The Performance of Distributed Acoustic Sensing for Tracking the Movement of Road Vehicles

In this paper we introduce Distributed Acoustic Sensing (DAS) and discuss its potential application in Intelligent Transportation Systems (ITS). We present the fundamentals of the technology and data processing methods required for converting the raw data into information that can be used to track the position of vehicles along a road in real time. We present and discuss the results of experiments that explore the dependency of defined DAS signal quality metrics as a function of vehicle speed, size and distance from the fiber, as well as how the cables are installed. The experiments indicate that decisions concerning cable installation could influence data signal to noise by up to ~10 dB. Finally, we demonstrate that DAS systems that operate by measuring phase shifts of laser light backscattered along the fiber, are capable of tracking smaller and slower moving traffic than DAS systems that rely only on measuring the intensity of the backscatter.

The Performance of Distributed Acoustic Sensing for Tracking the Movement of Road Vehicles Peter J. Thomas , Yngve Heggelund , Inge Klepsvik , Jeremy Cook, Erling Kolltveit, and Torgeir Vaa Abstract-In this paper we introduce Distributed Acoustic Sensing (DAS) and discuss its potential application in Intelligent Transportation Systems (ITS).We present the fundamentals of the technology and data processing methods required for converting the raw data into information that can be used to track the position of vehicles along a road in real time.We present and discuss the results of experiments that explore the dependency of defined DAS signal quality metrics as a function of vehicle speed, size and distance from the fiber, as well as how the cables are installed.The experiments indicate that decisions concerning cable installation could influence data signal to noise by up to ∼10 dB.Finally, we demonstrate that DAS systems that operate by measuring phase shifts of laser light backscattered along the fiber, are capable of tracking smaller and slower moving traffic than DAS systems that rely only on measuring the intensity of the backscatter.
Index Terms-Distributed sensing, fiberoptics, intelligent transportation systems, traffic monitoring.

I. INTRODUCTION
T HE subject of sensor systems for road traffic flow mon- itoring is gaining increasing attention in recent years.
The key function of technologies comprising an intelligent transportation system (ITS) applied to a given sector is to allow for real-time monitoring of the general congestion level, or level of service [1].Systems can also provide the input data for algorithms providing extra functionality, such as vehicle tracking and the detection of events that can influence traffic flow, such as breakdowns and stoppages, and vehicles travelling in the wrong direction or with excessive speed.Such information, when available in real-time, is useful for road authorities for maintaining safe and effective operation, for example by facilitating decisions related to control measures such as the opening/closure of lanes and roads, and adjusting speed limits.Information from ITS systems can be made available to road-users through mobile phone applications, Radio Data System -Traffic Message Channel (RD-TMC) or Variable Traffic Signs (VTS), allowing oncoming traffic to adjust route or regulate speed according to the conditions.Measurement data from ITS systems is also invaluable for making decisions related to road planning.
There are many types of sensor technologies that have been developed for ITS applications.Video based approaches [2], [3], [4], have received particular attention due to recent developments in high resolution, low cost camera technology and high-speed computer graphics hardware that allow for real time image analysis techniques.While video systems are relatively simple to install and can be effective for traffic flow monitoring and event detection, video based systems are often influenced by external factors such as weather variations, and the presence or accumulation of rain, snow and dirt on the lens.In addition, algorithms for tracking and event detection can be sensitive to camera position, occlusion by other traffic, dynamic background features, small changes in the traffic position, and changes in the illumination conditions.
Inductive loops [5], [6] comprise electrical cables buried in the road, as metallic vehicles pass this induces a measurable current modulation in the cables.Inductive loop configurations can be used to measure vehicle speed, classify vehicles according to their length, and are often used in combination with axel counters.Inductive loops are the mainstay of many ITSs today, thanks to their high durability and physical robustness over long time periods.A significant disadvantage with inductive loops are that installation and repair processes are expensive and cause damage to the road surface.
Microwave based radar technologies [7] are playing an increasingly significant role in modern ITS systems.Typically radar beam cones intersect traffic at a nominal angle of 90 degrees and be used to report the presence and velocity of vehicles as they pass.Radar data can also be used to classify different vehicles and can be used to cover multiple lanes of traffic [8].The popularity of radar systems is driven by their non-intrusive installation in common with video technology, and measurements that are invariant to changes in weather and lighting conditions.A drawback with radar systems is that they must be recalibrated at regular intervals, bringing significant operational and maintenance costs.
An inherent problem associated with these sensor technologies is that each sensor node covers only a point or limited section of road.Therefore in order for a longer section of road to be covered, large numbers of sensors are required, organized as an discrete array of point sensors.This results in high installation and maintenance costs related to complicated communication and power networks serving every individual physical point sensor in the shape of an individual camera, radar or inductive loop.Therefore, there has been a lot of effort towards the development of easy to install, low complexity sensors that can be easily integrated into low bandwidth wireless networks, such as passive infra-red sensors [9].Passive acoustic sensors also fit in this category [10], these listen to the sound generated by passing cars and are capable of providing vehicle speed when used on their own or as part of an arrayed system [11], [12].Such sensors are typically mounted on existing infrastructure such as road signs and lamp posts.While wireless sensor network approaches can offer better scalability than more traditional approaches, the requirement for separate sensors, communication and batteries drives up installation and maintenance costs, and particularly where large special coverage or high spatial resolution is required.
Relative to conventional acoustic sensors, a Distributed Acoustic Sensor (DAS) offers superior scalability and an ability to make highly spatially resolved measurements.DAS systems use a fiber optic cable installed alongside the road that, in common with conventional passive sensors, listen for vibrations induced in the road by passing vehicles.DAS effectively turns the optical fiber into an array of independently addressable acoustic sensors.A key advantage of DAS systems is that the low-cost passive optical fiber performs the sensing, data and power transmission functions that would typically require distinct hardware components for a solution based on point sensors.Furthermore, since the hardware necessary for probing the fiber can be installed at a single, environmentcontrolled location away from the roadside, maintenance and repair costs are minimized.These characteristics mean that DAS systems are very cost efficient when scaled up to longer stretches of road.To date there have been relatively few quantitative studies of DAS applied to road traffic flow [13], [14].The focus of previous studies has been to devise data processing strategies for detecting traffic as well as quantifying vehicle velocity and classifying vehicle type.Other studies have focused on exploiting the background noise generated by traffic for seismic based structural monitoring of the subsurface [15], [16].In the industrial sphere, road administrations have begun pilot studies aimed at investigating the potential for DAS in ITS.North America's first DAS pilot installation covers a 7 km stretch of highway in North Dakota [17], and has been used to quantify traffic speed and volume.Similar pilot installations are appearing at other locations across North America [18] and Europe [19].This work reports on some of the work being carried by the Norwegian Pubic Road Administration as part of the Borealis project [20], where one of the aims was to make an assessment of the potential of DAS for monitoring a remote road stretch with challenging winter conditions where traffic stoppages are common.A second aim of the Borealis project was to assess how choices made during the installation of the system influence the DAS signal quality.To the best of our knowledge there have been no experimental studies that quantify how the quality of DAS signals for measuring traffic are influenced by factors such as the characteristic of the passing traffic, the road conditions, how the cable is installed, and the choice of interrogation technology.These elements form the main contributions of the current work and will help road infrastructure mangers in the design of new DAS installations.
In the first part of this paper we give an introduction to DAS technology, and how the data is processed in the context of road traffic monitoring.Following this we present the details and results from experimental DAS investigations carried out at three separated locations on the Norwegian highway.We demonstrate that depending on the vehicle velocity and size DAS signal-to-noise (SNR) levels vary within the ranges 4.0±1.1 dB and 1.9±0.05dB respectively.DAS SNR was shown to vary by 10 dB depending on the cable installation parameters.Ducted cable degraded the SNR by 1.6±0.4dB relative to unducted installations, although the long-term durability of ducted cables was shown to be greater.An additional improvement of 6.0±0.4 dB was found by installing the cables within the road structure versus in a typical telecom trench alongside the road.A further improvement of SNR by 2.3±0.6 dB could be achieved by optimizing the vertical placement of the cables.The SNR corresponding to a vehicle travelling in the lane closest to the fiber was measured to be 8 dB greater than when the vehicle was traveling along an adjacent lane.In addition to being easier to handle and install, we also demonstrate a modest SNR benefit of 0.8±0.4dB when using specially designed acoustic sensor cable, relative to ordinary communication cable.We also measured DAS SNR levels to be 3.0±0.7 dB greater in summer relative to winter, due to the differing road conditions.Furthermore, the signal was influenced by 2.6±0.2dB, depending on whether the vehicle was travelling up or downhill.Finally, we demonstrate that DAS systems that operate by measuring phase shifts of laser light backscattered along the fiber are capable of tracking smaller and slower moving traffic than systems that rely only on measuring intensity variations in the backscatter.The significance of all the above factors on potential real-world performance of DAS for traffic motoring applications is discussed.NORCE were responsible for developing the DAS interrogators used within the project as gathering and processing the data.The Norwegian Pubic Road Administration were responsible for building the test sites and facilitating data collection.Both NORCE and the Norwegian Pubic Road Administration contributed to writing the paper.

A. Data Acquisition: DAS Interrogation
Distributed fiber optics sensors consist of a passive optical fiber and an interrogator unit located at one end of the fiber.Such systems measure physical parameters at every longitudinal position along the fiber optic cable.In addition to distributed acoustic measurement systems, distributed fiber optic sensors have been developed that measure temperature [21], strain [22] and even chemical parameters such as water detection [23], [24], [25] along the sensing fiber.Distributed fiber optic sensor interrogators operate by sending pulses of laser light along a fiber.As the light pulses propagate, DAS utilizes the Rayleigh backscatter, that is the most intense scatter and has the same frequency as the laser pulse.Rayleigh backscatter occurs due to inhomogeneities in the refractive index of the fiber core that are generated naturally during the fiber production processes.These inhomogeneities act as partially reflecting mirrors, with positions and reflection strengths that modulate as a function of acoustic signals impinging on the fiber.These modulations lead to changes in the intensity and phase of the backscattered light [26].Time domain analysis of these signals forms the basis of DAS systems.Some DAS systems focus on recovering acoustic signals by only using the intensity variations in the Rayleigh backscatter, while others operate by measuring the phase-shifts in the scatter over a predefined gauge length [27].DAS systems typically collect data at rates on the order of GB/s, and how these data are processed to provide actionable information for end users forms the basis of a great deal of research.DAS systems have found applications in a wide range of settings, particularly within the oil and gas industry [28], [29] and for geophysical monitoring [30].The basis of a DAS based ITS (see Fig. 1) involves the identification of the characteristic modulations in the Rayleigh backscatter induced by acoustic signals generated by passing traffic, and tracking their movement along the fiber as a function of distance/time.
The main limitation governing the refresh rate of DAS systems and therefore the acoustic frequencies that can be resolved is the requirement that all backscatter from each laser pulse must have left the fiber before introducing the next pulse.For a 10 km fiber, this limits the maximum refresh rate to 10 kHz, far higher than required for resolving acoustic signals passing traffic generate.Natural attenuation of the light by the fiber limits the range, modern fibers typically have losses as low as 0.2 dB/km.This means for light scattered at a point in the fiber 50 km from the interrogator, the intensity when detected at the interrogator will be approximately a factor of 10 less compared to the light scattered close to the interrogator.
For the majority of measurements presented in this paper, an Intensity (phase insensitive) DAS interrogator was used that sampled the backscatter with a frequency of 250 MHz, corresponding to a spacing between samples of 0.4 m.The interrogator used laser pulses with a duration of 45 ns and a pulse repetition of 20 kHz.

B. Processing of Raw Data
Irrespective of whether a given DAS system is designed to recover the intensity or phase of Rayleigh backscatter, the changes that occur when vehicles travel alongside an optical fiber are recovered by applying Fourier transforms covering discrete sections of data in space and time.In the current implementation we use sections of 1024 pulses with an overlap between sections of 512 pulses.For a pulse repetition rate of 20 kHz, this implies a time resolution of 0.05 s and a frequency resolution of ∼40 Hz.The result is a space-time matrix of the intensity/phase waveforms.For traffic data, we observe that most of the signal intensity is in the frequency range 100 to 600 Hz.Therefore, when evaluating the space-time matrix describing the overall signal level, we integrate over this frequency range, see  (a) shows a thin band of noise at low frequency at all times, this could be explained by sub-hertz seismic noise/thermal fluctuations [16], combined with the limited resolution bandwidth of the system.Fig 2 .(a)also indicates that a slightly higher signal to noise ratio could have been achieved by decreasing the lower frequency bound for the integration.This was not done in order to exclude the influence of potential distant low frequency noise sources.The signal to noise ratio achievable on further raising the upper bound was marginal or negative in most cases and was therefore not done.
The stochastic nature of the backscatter variation along the fiber is accounted for in each sample point by subtracting a centered rolling average over a window of processed shots, and then dividing by the local standard deviation (background noise).The resulting space-time matrix contains "z-score" values, where a positive z-score means that the backscatter-variation in the sample is higher than average.The movement of vehicles travelling alongside the fiber cable appear as lines of high values in the z-score matrix, see Fig. 3(a).A more detailed theoretical description for Distributed Acoustic sensing can be found elsewhere in the literature, for example in [26].

C. Recognition of Vehicles From the Processed Data
In order to determine the instantaneous location of road traffic using the DAS z-score matrix, algorithms are used that discriminate lines of high value from the background noise.
The vehicle recognition algorithm used in these investigations relies on the Hough-transform [32] as the basic building block.The Hough transform can generally be used to extract features (lines, general shapes) from images.Here, the algorithm is used to find line segments, which are then grouped together to form vehicle tracks.In a functioning DAS based ITS, these tracks could provide the basis for vehicle location and velocity algorithms as well as event detection algorithms, such as those aimed at detecting traffic flow stoppages.

D. Data Quality Metrics
In general, the quality of DAS datasets will influence the effectiveness of algorithms for tracking road traffic, and therefore how suitable the datasets are for specific functions.For example, high quality DAS data can be used to track all vehicles with a high degree of precision and reliability, and facilitate the detection of complex traffic behaviors such as queue formation.The usefulness of lower quality DAS data on the other hand may be limited to simpler problems such as car counting, average speed and level of service assessments.
In order to compare different DAS traffic datasets collected under different conditions, we have devised some quality metrics.The quality metrics were applied to time-space regions of the z-score matrix that contain a vehicle track identified by the Hough transform algorithm, see Fig. 3(b).
The first metric, the signal to noise ratio (SNR), describes the ratio between those parts of the z-score matrix associated with the detected track, and the background noise.It is common to express the SNR on a logarithmic scale as in (1); where s is the signal (the z-score associated with the track) and n is the background noise.E(•) denotes the expectation value, which for signal and noise is approximated by the average value over a track, and the average value over a track-free area respectively.A higher SNR score for a given set of conditions is an indicator of higher sensitivity, and a greater ability to detect weaker signals, like those from smaller and slower moving vehicles.
The SNR value associated with a given track is an average over the selected space-time window.This means that provided the track contains regions with significantly higher z-scores than the average, a track could maintain a high SNR value even if there are gaps in the track, i.e. the tracks lack continuity.Track continuity is an important variable for evaluating the real-world performance of DAS based ITS because it directly influences the demands on algorithms that use z-score tracks as their inputs.This includes algorithms for vehicle counting, measuring vehicle velocity, and detecting events such as traffic stops.Therefore in order to complement the SNR metric, we define the signal-continuity (CNT) metric.This is defined as the ratio of the sum of the lengths of the detected track segments to the total expected track length as in equation ( 2).
where l i is the length of detected track segment i, and L is the expected total track length.The expected total track length is calculated using a best fit curve fitting through the detected track segments.
Both the SNR and CNT metrics defined here are intended to be applied to datasets corresponding to a single vehicle travelling at constant velocity.

III. EXPERIMENTAL INVESTIGATIONS
The majority of the tests reported in this paper were carried out at two test sites along the two-lane E8 road in northern Norway; one site at Sidnodjeva (site A) and other at Gardeborgbakken (site B), see Fig. 4. Test site A was Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.approximately 750 m in length and had no incline.Test site B was 610 m in length with an approximate average 1:15 incline.Each test site was fitted with fiber cables that were all installed parallel to the road, and had a range of different specifications in relation to the types of cable used and how these were installed.
Standard cables used for communication were used as well as two cables designed for acoustic sensing applications.One of the special acoustic sensing cables was BRUsens AC4 from Solifos, with a 6.5 mm outer diameter non-metallic design with aramid yarn for strain relief and a Polyurethane outer jacket.The second acoustic sensing cable was a BRUsens AC3 with a 4.5 mm diameter design and where the optical fiber were contained within a stainless steel central metal loose tube.The metal tube was protected by stainless steel strength members and a polyacrylate outer sheath.The more physically robust AC3 cable was chosen for direct installation within the road structure, whereas AC4 was preferred when used together with protective cable ducts.Road structures are made up of asphalt, base course and sub base course layers, see Fig. 5(a).
Note that at the time of testing neither test site A or B had an asphalt layer.The base coarse layer was composed of stones with a maximum diameter of 22 mm and bound by bituminous materials.The sub-base course layer was constructed of stones in the size range from 22 mm to 90 mm.Fiber cables were installed in both the base course and sub-base course layers for site A, and only in the base course layer at site B and C. It is not desirable to have fiber cables in the asphalt layer since this upper layer is most susceptible to damage and is periodically removed and replaced as a part of maintenance procedures.Fiber cables were also installed at different lateral positions, see Fig. 5(b).The majority of the cables were installed in  the road structure at a short distance from the edge (∼ 1m), while others where either installed close to the centre of road, or in a nearby trench used for housing cables for telecommunication links.Some cables were installed in tubes of diameter ∼ 20 mm while others installed directly into the road structure.
In order to compare intensity and "phase sensitive" interrogator technology, separate investigations were carried out at test site C at Fløyfjell tunnel, on the E16/E39 road in Bergen, see Fig 4(b).Test site C is approximately 1 km in length, and comprises of two lanes with no gradient, both for northbound traffic.A single cable was installed in the base-course layer and was positioned towards the side of each lane forming a loop.
The different cable installation configurations are summarized in table I.In order to reduce measurement uncertainties at test sites A and B, the different configurations (except configurations 6 and 7) were installed both lanes, with each lane being split into two subsections, where the cables were terminated.The cables in each subsection were connected together at roadside junction boxes to give a single test fiber consisting of multiple cable and installation configurations.

IV. RESULTS AND DISCUSSION
Unless specified, the data presented from roadside cables concerns passing test vehicles on the nearside lane.

A. Background Noise
As explained in section II-D., the background noise value is used for compensating the z-score matrix data for the stochastic nature of the DAS signals.In order to estimate the background noise, we used the standard deviation for data from time periods where there was no traffic.Note that all the recording during these experiments were made under the same fair weather conditions, and the background noise may increase under precipitation or wind.We found no strong evidence that the cable used, or its placement had a significant influence on the background noise.We also do not see any physical reasons why different cable configurations should have different noise levels under quiet conditions, since all cables contained similar optical fibers following the International Telecommunication G657 standard, therefore constraining the Rayleigh backscatter level.Self-induced interrogator noise should be equal for all datasets.As a result of these observations, a common noise level was applied to all the different cable configurations when computing the SNR.

B. DAS Signal Characteristics as a Function of the Passing Vehicle Properties
As the vehicle speed increases, the increased rotational speed of the tyres leads to greater noise generated by the tyre-road interaction.As a result, the vehicle speed has a significant impact on DAS signal.This can be clearly seen in Fig. 6 that demonstrates how the apparent quality of the vehicle space-time DAS track generated by a passing truck degrades as the velocity decreases from 50 km/h to 10 km/h.
The apparent quality of the tracks in terms of the SNR and CNT metrics defined in II D. are shown in Fig. 7.The data shows that the SNR decreases by an average of 4.0±1.1 dB as the vehicle velocity decreases from 50 km/h to 10 km/h.A decrease in track continuity (CNT) of 23±5.7 % was observed over the same velocity range.This is a significant difference, and high DAS performance at low velocities would  depend strongly on the quality of the tracking algorithms used.Both the CNT and SNR data in Fig. 7 show an approximately linear trend with velocity, and by extrapolation can be used to make reasonable predictions on how signal quality is likely to be at value above, below and in between those for which data was collected.
In addition to the vehicle velocity, another factor determining the acoustic noise energy generated by the tyre-road interaction is the vehicle mass.This is revealed on comparison of the DAS signal quality when vehicles of different size pass.The data in Fig. 8 shows a more clearly visible track for the truck than for the lighter vehicles.
In Fig. 9, the data shows that the DAS SNR was 1.4±0.6 dB greater for the truck compared to the van travelling at the same speed.This is partly due to the greater mass of the truck, and a greater surface contact area due to a three axel, 10-wheel the van.In contrast, the SNR was only 0.5±0.6 dB greater for the van than for the estate car, since the van was only 30 % heavier, with a similar contact area to the road.The CNT value of the signal for the truck, van and estate car at 50 km/h were measured to be 100 %, [98.3, 100] % and [97.5, 99.2] % respectively, where the square brackets denote the interquartile range.These close lying values indicate that real-world performance for DAS is potentially good for a wide range of vehicle types travelling at 50 km/h.Note that although only the driver was present in the vehicle during the testing, this discussion implies that the DAS signal strength could also be dependent on the number of passengers, through the influence on the vehicle mass.
Also, we observe that the track corresponding to the truck is wider than for the two smaller vehicles.These observations present fundamental bases for an algorithm to differentiate between vehicle types.
The effect of the lateral vehicle to fiber separation on the DAS data quality was investigated.The data in Fig. 10(a), corresponds to an average lateral separation of 1-1.5 m between the near-side wheels and the fiber.The appearance of this data is similar to in Fig. 10(b), where the separation was in the range 0-0.5 m.In contrast, the appearance of the track degrades significantly in Fig. 10(c) where the estimated lateral separation was 3.5-4 m.This observation is consistent with the fiber optic sensing association's guidelines [31].The SNR of the track data in Fig. 10 (a), (b) and(c) are 27.5 dB, 27.5 dB and 19.7 dB respectively, whereas the continuity was 100 % for all vehicle-fiber separations.In terms of real-world performance, this data suggests that the DAS data quality is robust to small natural lateral deviations (∼ <1 m) of passing traffic within any given lane.Furthermore, the 8 dB SNR drop observed when the vehicle was traveling in the adjacent lane suggests that continuous tracking of individual vehicles across multiple lanes of traffic with a single DAS cable would be challenging, particularly for smaller vehicles travelling at lower velocities.
Given that a sensing cable is installed in each lane, the reduction of the DAS signal with increasing lateral separation between vehicle and cable serves as an excellent filter, ensuring the data from a given cable is dominated by vehicles  travelling within the closest lane, and will not be significantly influenced by more distantly located noise sources encountered in typical urban environments such as trains, and traffic in adjacent lanes or intersections.It follows that a dual cable configuration would be an advantage if lane identification for passing traffic is desirable.Furthermore the data implies that in order to maximize SNR, cables should not be positioned such that they are separated from the road traffic of interest by interfering noise sources, such as paths for cyclist and pedestrians.
We re-analyzed some of the data for the truck travelling at different velocities, while adjusting the 100-600 Hz band pass condition introduced in IIB to cover the range 50-1000 Hz and did not see a significant change in the SNR or CNT values.This is consistent with our observation that there was no strong correlation between the vehicle velocity and frequency content of the data.
It is worth mentioning that the DAS data is in general dependent on the directionality of the incident seismic waves, which have p and s wave components, and the cable will be more sensitive to the component which causes the greatest disturbances parallel to the cable.For perpendicularly incident waves, the cables will be relatively incentive to the faster p wave component.A more detailed analysis of s and p wave arrival is often carried out in experiments involving stationary impact sources, where s and p wave contributions can be unambiguously identified from one another [30].

C. DAS Signal Characteristics as a Function of the Cable Installation and Road Conditions
The amount of acoustic energy transferred to the optical fiber from the environment will depend on the mechanical construction of the fiber cable and how it is installed in the road structure.This in turn could have an impact on the DAS data quality and real-world performance.We observed that strong signals are observed both when the fiber cable was placed in a cable duct in the road structure, and when the cable was installed directly into the road structure, see Fig. 11.However, the direct installation data appears to give a more stable signal intensity along the cable, indicating a more uniform acoustic coupling with the environment.Fig. 12. show that these observations are also reflected in the statistics covering a larger dataset, where the SNR for data correspond to fibers installed with and without a cable duct was in the intervals [27.4, 27.7] dB and [28.3, 29.4] dB respectively, representing a difference of 1.6±0.4dB.
This observation is consistent with the fiber optic sensing association's guidelines [33], and with the observations concerning the acoustic response of ducted and unducted fiber cables in [34] The continuity was 100 % for all of the datasets corresponding to direct installation of the cable in the road  structure, and for all but one of the datasets where a cable duct was used.This last observation shows that under the given conditions, high continuity was maintained when the cable was installed in a cable duct, even though some partial dips of the signal intensity were observed.While the data indicates good real-world performance potential for both cases, using cables installed directly in the road structure should provide more robust DAS data for vehicle tracking algorithms, particularly where slow moving or small traffic vehicles are concerned, where the acoustic signal is weaker.
Superior DAS SNRs were observed for sensor cables installed in the base course ([28.3, 29.4] dB), relative to the sub-base course ([26.0,27.7] dB), representing a difference of 2.3±0.6 dB, see Fig. 13.A possible explanation is the reduced ability of surface acoustic waves to penetrate beyond the base course layer into the sub-base course layer.The signal continuity under the given conditions are high for cables in both the base and sub-base course cases, indicating similar real-world DAS performance.However, the SNR values lead us to conclude that sensor cables in the base course layer provide a more robust data input for vehicle tracking algorithms.
Fiber optic cables are often found alongside roadsides as part of communication systems, and many of these cables have free capacity that could potentially be exploited by DAS based traffic monitoring systems.The free capacity can either be in the form of when the DAS laser is used in "dark" fibers that are not being used for communication purposes, or is tuned to an unoccupied frequency band.Exploiting this free capacity would be extremely attractive since the installation of cable represents a significant portion of the total cost of installing DAS systems.The previous datasets were acquired using premium cables, specially adapted for DAS applications through their construction.Furthermore the majority of the previous datasets were acquired using cables either installed within the road structure.Communication cables on the other hand are often well sound-insulated and installed in trenches outside the main road construction.The representative data in Fig. 14 isolates the influence of non-ideal cable design from non-ideal installation on the data.The data shows that choosing a standard telecommunications cable, when installed in the road structure, results in relatively little degradation relative to a specially adapted acoustic fiber cable.This is partially consistent with the results in [35], where a loose tube cable construction not dissimilar from the communication cable used in this investigation, gave a better acoustic response than acoustically optimized cables identical to the ones used here.
However, there is a significant signal degradation when cables were installed in a trench approximately 2.5-3 m away from the roadside, and is consistent with [33].These observations are also reflected in the SNR figures calculated from a larger dataset, see Fig. 15, with the SNR for road-installed special acoustic sensing fiber being greatest, [28.1, 28.7] dB, with a modest degradation to [26.9, 27.9] dB for road-installed communication fiber cables, and a significantly lower SNR of [21.6, 22.5] dB for the trench installed fibers.The difference between the highest and lowest values, for special fiber installed in the road structure, and trench installed fiber, was 6.0±0.4 dB.The weaker signals from the trenched fiber, was anticipated since the acoustic energy generated by the tyre road interaction radiates in all directions within the road structure, and the energy transferred to the fiber reduces with the acoustic transmission distance, that was greatest for the trenched fibers.Furthermore, the fact that the cable trench lies outside the main road construction may contribute further to the signal degradation due to boundary reflections and higher porosity of the mass surrounding the trench, compared to within the road construction.The strong dependence of the DAS data quality on vehicle-cable separation discussed in section IV-B suggests that this was the dominating reason why the trenched fiber gave poorer results.
Despite the relatively low SNR for trenched fiber, the CNT value remained high, with only one of the dataset represented in Fig. 15 demonstrating non-perfect continuity.This indicates that while real-world DAS performance is expected to be poorer for trenched installed fibers, the data could still be sufficient for characterizing the traffic flow, and therefore provide the basis for a highly cost efficient ITS.
At test site B, that had a 1:15 (6.7 %) gradient, we observed that DAS SNRs were higher for vehicles travelling uphill compared to downhill (2.6±0.2 dB), see Fig. 16.This could be due to the difference in force vectors exerted by the tyres on the road surface for up-and downhill travel, that we estimate to be around 14 %.
An additional contribution to the signal difference could be from increased vehicle motor and transmission noise when travelling uphill.It is unlikely that this feature would negatively impact the real-world system performance except for more extreme gradients found on minor roads, where the application of DAS is less relevant.
The amount of acoustic energy generated by the tyre-road interaction is also determined by the road conditions with harder, rougher road surfaces generally expected to generate stronger acoustic signals.It is therefore anticipated that the strength of DAS signals to passing traffic increases with rougher road surfaces.While there were no observable spatial variations in the road surface quality at site A or B, there were seasonal changes in the surface conditions.During the summer the road surface was clear, but during the winter large areas of the road surface was partially coated in compacted snow, creating a softer and quieter road surface.This was reflected in the data shown in Fig. 17 where the average SNR over three different cable configurations decreased by 3.0±0.7 dB from summer to winter.
Fig 17 also shows a larger variation in the SNR across datasets acquired in winter relative to summer, most likely because of the patchiness of the snow covering, and the fact that the vehicles take slightly deviating paths through the test section during winter due to the lower visibility of road markings.Regarding the track continuity during winter time, while this was not consistently 100 % as it was in the summer, the signal during the winter generally stayed above 90 %.The larger SNR observed in the data from the roadside cable in Fig. 17 relative to the road centre cable is most likely vehicle proximity related.In order to get a complete picture of the role temperature plays in determining the DAS signal quality,  it may be relevant to consider through modelling, the effect of acoustic impedance mismatching between ground layers with different temperature.Such an analysis is beyond the scope of the current paper.
Note from the site B data shown in Fig. 17 that the roadside cables installed directly in the road correspond to a higher SNR than for the roadside cable in the duct, this is consistent with the data in Fig. 12, corresponding to the same cable configurations at test site A, and the same test conditions.For configuration 3 the SNRs were very consistent between Site A [27.4, 27.7] dB, and site B [29.1, 29.5] dB.For configuration 1, the SNR deviated somewhat between Site A [28.3, 29.4] dB and site B [31.9, 32.4] dB.The reason for the slightly higher SNR at site B is most likely due to the fact that the vehicle was travelling uphill, increasing the vehicle force vector perpendicular to the road surface, and contributing extra motor and transmission noise.The difference could also be due to the fact that the cables at site A were installed during construction of the road, whereas the cables at site B were retrofitted on an older road stretch by a standard microtrenching procedure.This difference in the installation method may have influenced the transmission efficiency of the acoustic energy to the fiber.Related to this, comparing the data acquired (for similar sized vehicles at 50 km/h) from the asphalted road test site C, with the data test sites A and B, where no asphalt was present indicates that the presence of an asphalt layer results in a SNR reduction of around 3 dB.This is consistent with a reduction in the acoustic energy generated by the interaction of the tyres with the soft asphalt layer.It is unlikely that this reduction would negatively impact the real-world performance of the system.However, the sensitivity in general of DAS to road surface conditions may provide useful functionality for maintenance planning related to degradation and alerting to safety issues.
Regarding the long-term ageing of the cables, OTDR measurements at site B revealed that the average loss of fibers installed in ducts increased from 0.20±0.02dB/km to 0.30±0.01dB/km over a 3.5 year time period.For fibers cables installed directly in the road, attenuation increased from 0.19±0.02dB/km to 0.63±0.08dB/km over a four-year time period.This indicates that despite the heavier construction of fiber cables installed directly in the road, these cables appear to degrade more rapidly, mostly likely due to the increased exposure to frost heaving.These data suggest that the use of cable ducts could be important for ensuring high performance of DAS systems over decadal time scales.Note that while the measured drop in fiber transmission would be difficult to Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
compensate for in an otherwise optimized DAS system, it is unlikely to have would have significantly affected the DAS system performance, even for the fiber cable installed without cable ducts.
A general observation for all the datasets is that the Hough transform tracking algorithm performed robustly for a large range of SNR values, giving a high degree of continuity and demonstrates the suitability of the algorithm for the purpose.The robustness of the continuity of the SNR and continuity metrics to adjustments in the parameters of the Hough transform algorithm was tested.This seemed to affect the results of all the fiber cable installations in similar proportion and did not alter the overall conclusions.We are therefore confident that these metrics forms a valid basis for ranking the cable installation methods.
We observed that for DAS datasets where several vehicles were present, the data processing procedure outlined in section II-D.tends to diminish the signals from small vehicles that follow closely after large vehicles.The main reason for this is the division by the local standard deviation, whose size will be affected by the signal from the large vehicles and therefore may become too large and unrepresentative.This adverse effect can be remedied by using a more robust method that account for the variation in background noise due to other traffic.However, this type is much more computationally expensive and has therefore not been used for the results in this paper, where only one vehicle was present in each dataset.

D. Monitoring of Multi-Vehicle Traffic Flow
The data discussed in IV C show that the degree to which noise generated by any given vehicle contributes to the background decreases rapidly with distance from the vehicle.This implies that it should be relatively straightforward to use DAS for identifying single vehicles in multi vehicle traffic data, without background noise from the other vehicles interfering.In order to illustrate this, we recorded data at test site c during day-time traffic.In this case we used an intensitybased interrogator, see Figure 18.The average traffic speed was around 75 km/h, and the CNT measure is seen to be generally good.The road consists of two lanes in the same direction inside a tunnel, and the tracks that appear and/or disappear over the recorded time interval are due to lane changes, confirmed both by video camera observations and data from a DAS cable installed close the adjacent lane (not shown).The CNT for the track indicated with yellow in the lower part of Fig. 18 is 79 %.A detailed analysis of traffic line patterns as a function of the state of traffic flow is beyond the scope of the current paper.

E. Performance of Phase Sensitive DAS Technology
In the previous sections, the focus was on determining the influence of cable installation on the DAS system performance.While the data quality was high for a wide variety of situations data quality was significantly reduced for lower vehicle velocities and size, potentially limiting usefulness for certain traffic monitoring applications.In addition to the cable installation, data quality will also be influenced by the choice  of interrogator technology.All the measurements presented in previous sections were collected using an "intensity" type DAS interrogator.A "phase" type interrogator, while technologically more complex and expensive, has some desirable properties such as higher sensitivity and more linear response to acoustic disturbances [36].Fig. 18 shows DAS data from test site C, simultaneously collected using an intensity type and a phase sensitive interrogator, each connected to separate fibers within the same cable.The data shows that the phase sensitive system gives a stronger and more continuous response to a passing car.The difference is most apparent where the car is moving slowest, i.e.where the acoustic signal is smallest.At 10 km/h the phase sensitive system gave a continuity of 49.8 % relative to the 18.4 % from the intensity system.This is a highly significant result, given that slow vehicle velocity is perhaps the factor with greatest negative impact on real-world DAS performance, section IV-B.
Another important variable for phase sensitive systems is the gauge length, the length of fiber over which assessments of phase shifts are made.Gauge lengths are typically on the order of a few meters and was 16 m for the data shown in Fig. 19.While a detailed assessment of the effect of gauge length on the DAS data for traffic applications is beyond the scope of this paper, in general longer gauge lengths use longer optical pulses meaning a stronger signal may be expected, particularly for large vehicles.On the other hand, spatial averaging effects caused by longer gauge lengths would degrade the spatial resolution, meaning that shorter/smaller vehicles may become more difficult to observe.

V. CONCLUSION
In this paper we introduce Distributed Acoustic Sensing and discuss its potential application for ITS system in the context of existing methods.We present the basic technology and data processing methods required for converting the raw data into information that can be used to track the position of vehicles through a stretch of road.We also presented and discussed the results of experiments designed to explore the dependency of defined DAS signal quality metric on variables relating to the passing traffic, and the conditions surrounding the fiber installation.We demonstrate a range in DAS SNR levels spanning, 4.0±1.1 dB and 1.9±0.05dB depending on the vehicle velocity and size respectively.Low vehicle velocity presents the most challenging conditions for DAS technology in real world situations.The SNR varied by 1.9±0.05dB depending on the vehicle type, indicating that DAS technology could be suitable for classifying the type of passing vehicles.
Concerning the cable installation, we demonstrate DAS data quality varies across a range covering several dBs, with best data quality resulting from cables installed directly within the road structure as opposed to being installed in cable ducts in the road structure or in roadside trenches.We demonstrate the DAS SNRs vary within the range 2.3±0.6 dB depending on the vertical placement of the cables, with best results when cables are placed in the base course layer.In addition to being easier to handle and install, we also demonstrate a modest SNR benefit of 0.8±0.4dB when using a specially designed acoustic sensor cable, relative to ordinary communication cable.In summary, while the impact of individual cable installation choices do not make a large effect on SNR, different combinations of all fiber installation parameters could easily make a 10 dB difference in the data SNR, therefore impacting on the real world performance.The cable installation is therefore a critical element in the system design.By studying the cable transmission over time, we also demonstrated that cables installed directly in the road age more rapidly than those protected using a tube.
With respect to the road conditions, we showed that DAS SNR levels to be 3.0±0.7 dB greater in summer relative to winter, due to the differing road conditions.we also demonstrated a 2.6±0.2dB SNR difference depending on whether the vehicle was travelling up or down a gradient.This sensitivity means that DAS could find applications in monitoring the condition of the road.
Finally, we demonstrated that phase sensitive DAS systems are capable of tracking smaller and slower moving traffic than intensity based DAS systems.This is particularly important since slow vehicle velocity is one of the major factors impacting DAS performance in real-world applications.
In addition to the data SNR, another factor impacting the real world performance of DAS systems is dependent on the performance of algorithms to identify and track vehicles in the data.In this paper we defined a continuity metric to quantify this performance.We demonstrate the performance of the tracking algorithm used here was capable of maintaining reliable tracking over a wide range of signal qualities.The maintenance of high tracking continuity for a wide range of signals, in combination with low computational overhead is an important condition for real-time monitoring.In the current study, the continuity of DAS tracks in most cases was maintained close to over 90 % even when the SNRs dropped by around 7-8 dB relative to the datasets with highest SNR.High continuities were maintained even for small vehicles at moderate speeds.The reduction in vehicle velocity was by far the most important factor influencing the continuity metric.Continuity dropped to [67.8, 86.7] % for the truck travelling at 10 km/h travel speeds, and [6.3, 13.4] % for the smallest vehicle type.
While the tracking algorithm used this work would be challenging to implement in real-time on relatively modest computer hardware, the authors have developed a prototype real-time tracking algorithm based on Kalman filtering [32].This algorithm was tested at test site B, although the results of this test are beyond the scope of the current paper, and will form part of a future publication.
DAS in combination with the method and techniques described in this paper, could potentially form the basis of an effective ITS tool across a wide range of applications from vehicle counting and average speed measurement to detecting complex travel behaviors such as car breakdowns and the formation of queues.Other possible functionalities are vehicle classification and road condition monitoring.In order for such a system to be realized having sufficient reliability and accuracy, the most pressing future challenge is for the development and verification of high-performance real-time methods for processing the DAS data as well as real-time event recognition algorithms.As DAS technology continues to improve, with high sensitivity interrogators [37], cable architectures [38], including weak Fiber Bragg grating arrays [31], and advanced data processing/machine learning [39] forming active fields of research, the performance of DAS systems for ITS applications is certain to improve dramatically in the coming years.The results and conclusions from this paper are also relevant when planning DAS systems designed for other transport monitoring applications, such as Railway [40] and airport monitoring [41].
This work was mainly focused on single vehicle traffic, where the mass, position and velocity at any given time were accurately known.While this technique allowed us to make some accurate assessments of DAS signal quality for a range of different situations, including vehicle type, position velocity, and cable installation parameters.However the method is less well suited to assessing how DAS data can be used for detecting complex multivehicle traffic flow patterns.In this situation the experiments would require large numbers of vehicles in combination with other sensor systems such as inductive loops and video based approaches.Future experiments could also include vehicle loading factor as a variable.The relationship between progressive degradation of the road surface with DAS signal quality needs to be investigate further, and could potentially find use cases in road maintenance planning.

Fig. 1 .
Fig. 1.(a) Sketch of a DAS based ITS system, consisting of an interrogation unit that sends laser pulses into an optical fiber, and interprets the Rayleigh backscatter.(b) The red curve shows the characteristic features of raw DAS data corresponding to a single laser pulse.The trace consists of a series of reflection peaks due to inhomogeneities in the fiber refractive index.The intensity (I) and phase of the backscatter of a subsequent trace (black dotted line) is modified by an acoustic signal at a specific distance.Note the intensity of the backscatter along the fiber decreases due to attenuation.

Fig. 2 .
Fig. 2. (a) The frequency content of DAS signals at one location along a fiber.The data represents a mostly quiet noise situation, except for when a truck passes after 17 seconds (b) The acoustic energy integrated between 100 Hz and 600 Hz for the same time period.

Fig. 3 .
Fig. 3. (a) z-score matrix for DAS data from a roadside cable as a vehicle travels alongside at 50 km/h.The fiber cable consists of two sections connected in series and are separated by a loop, causing an apparent step in the line of high z-score values.The step is calibrated out in subsequent processing.(b) Tracks detected by the Hough algorithm.The data is sectioned manually into space-time windows (dotted-line) in order to evaluate and compare the data quality between different sections.

Fig. 4 .
Fig. 4. (a) Location tests sites A and B on the E8 highway that were used during the investigations.(b) Location of test site C on the E16/E39 highway.

Fig. 5 .
Fig. 5. (a) Representation of the road structure at the test sites.Cables were installed in the base course and sub-base course layers.(b) Representation of the lateral position of the installed cables, the figure is approximately to scale, with the width of the lane being 3.5 m.
During data collection the test sites were closed for normal traffic flow.Tests consisted of individual vehicles driving through the entire test site at an approximately constant speed, and holding an approximately constant lateral position.Constant speeds of 10, 30 and 50 km/h were chosen for the majority of the tests.The emphasis on lower velocities was due to their relevance for DAS based traffic stop detection and queue formation algorithms.In addition, low velocities are most challenging for DAS systems due to the reduced traffic generated acoustic noise.There were three vehicles used in the test at sites A and B; a Volvo 520 tipper truck, a VW transporter van, and a Skoda Octavia estate car.The vehicles were chosen because of their contrasting sizes and weight.The curb weights of the truck, van and estate car were approximately 9.0 tons, 2.0 ton and 1.4 tons respectively.At test site C, two test vehicles were used; an electric powered Kia Soul small passenger car with a curb weight of 1.7 tons, and a Mitsubishi Pajero 4 × 4 with a curb weight of 2.0 tons.For all experiments, only the vehicle driver was present in the vehicle, with no additional passengers.The measurements at the E8 tests sites were taken over two separate periods, in February 2018 and August 2018.All DAS measurements acquired using the same "intensity" type interrogator.Measurements at test site C were carried out during November 2018 and June 2019.The long-term ageing properties of the optical fibers installed at test site B were assessed through optical time domain reflectometry (OTDR) measurements of the loss, first during February 2018, and later during January 2022.

Fig. 6 .
Fig. 6.Space-time visualization of DAS data as a Truck passes through part of test site B. The approximate average speed of the truck was (a) 50 km/h (b) 30 km/h and (c) 10 km/h.The data was collected using cable configuration 1.The color scale is the same for all figures.

Fig. 7 .
Fig. 7. DAS SNR (blue) and Continuity (red) corresponding to multiple truck passes, as a function of speed,.The data was obtained using cable configuration 1.

Fig. 8 .
Fig. 8. Space-time visualization of DAS data corresponding to (a) a truck, (b) a van, and (c) a car driving through part of test site A at 50 km/h.All of the data was collected using cable configuration 3. The color scale is the same for all figures.

Fig. 9 .
Fig. 9. DAS SNR (blue) and Continuity (red) corresponding to multiple passes of different vehicle types.All vehicles were driving at 50 km/h.The measurements correspond to cable configuration 3.

Fig. 10 .
Fig. 10.Space-time visualization of DAS data corresponding to when the truck was passing through test site A at 50 km/h with a near-side lateral distance to the fiber cable of (a) 0-0.5 m, (b)1-1.5 m and (c) 3.5-4.0m.Cable configuration 5 was used.The color scale is the same for all figures.

Fig. 11 .
Fig. 11.Space-time visualizations of DAS data when a truck travel through part of test site A with an approximate average speed of 50 km/h.The two plots both correspond to cables installed (a) without a cable duct (configuration 2) or (b) with a cable duct.Both cables were installed in the sub-base course layer of the road (configuration 4).The color scale is the same for both figures.

Fig. 12 .
Fig. 12. DAS SNR (blue) and Continuity (red) corresponding to multiple passes of a truck at 50 km/h, for cables with or without a cable duct.The data on the left hand side corresponds to cable configuration 3, the data on the right hand side configuration 1.

Fig. 13 .
Fig. 13.DAS SNR (blue) and track continuity (red) data corresponding to multiple passes of a truck driving at 50 km/h through test site A. The data on the left and right corresponds\ to cables directly installed in the base (configuration 1) and sub base course (configuration 2) layers respectively.

Fig. 14 .
Fig. 14.Space-time visualization of DAS data as corresponding to when a truck drives through part of test site A, with an approximate average speed of 50 km/h, for different installation configurations.(a) Configuration 6; communication cable in a roadside trench (b) Configuration 5; communication cable installed in the sub-base course layer (c) Configuration 4: Acoustic sensor cable installed in the sub-base course layer.The color scale is the same for all figures.

Fig. 15 .
Fig. 15.DAS SNR (blue) and track continuity (red) for data corresponding to multiple passes of a truck driving at 50 km/h through test site A. The data corresponds to communication cables installed in a roadside trench (left, configuration 6) and installed in the sub-base course (middle, configuration 5).Also shown in the rightmost position are the figures when the special acoustic sensing cable is used in the sub-base course (configuration 4).

Fig. 16 .
Fig.16.DAS SNR (blue) and continuity (red) as a function of the driving direction for multiple passes of the truck through test site B at 50 km/h in both uphill and downhill direction.The data was collected using cable configuration 1.

Fig. 17 .
Fig. 17.DAS SNR (blue) and CNT (red) values corresponding to multiple passes of a truck uphill through Site B at an average approximate speed of 50 km/h.The measurements were collected during the (a) summer and (b) winter.The data was collected with cables installed in conduits at the road centre (left, configuration 7), and road side (centre, configuration 3), and roadside cables installed directly in the road structure (right, configuration 1).

Fig. 18 .
Fig. 18.DAS data recorded during a normal traffic situation at test site C.The left-hand side shows the space-time visualization of the traffic in one of the two lanes.The right-hand side shows the tracks detected by the Hough algorithm.The track shown in yellow has a CNT of 79 %.

Fig. 19 .
Fig. 19.Space-time visualizations of DAS data collected data as a small passenger car drives through test site c (a) Upper: Intensity-based interrogation technology, lower: Phase sensitive, 16 m gauge length.The speed was 10 km/h left of the first yellow dashed line, 20 km/h between the two yellow dashed lines, and 40 km/h right of the last yellow dashed line.The color scale is the same for both figures.

TABLE I CABLE
AND INSTALLATION CONFIGURATIONS USED DURING THE INVESTIGATIONS