Detection Rate of Congestion Patterns Comparing Multiple Traffic Sensor Technologies

This paper investigates the detection rate of various freeway congestion patterns and compares them across different traffic sensor technologies. Congestion events can be categorized into multiple types, ranging from short traffic disruptions (referred to as Jam Wave) to Stop and Go patterns and severe congestion scenarios like Wide Jam. We analyze multiple traffic data sets, including speed data from loop detectors, travel time measurements from Bluetooth sensors, and floating car data (FCD) collected from probe vehicles. Each combination of congestion pattern and detection technology is thoroughly examined and evaluated in terms of its capability and suitability for identifying specific traffic congestion patterns. For our experimental site, we selected the freeway A9 in Germany, which spans a length of $\mathrm {157~km}$ . Our findings reveal that Bluetooth sensors, which record travel times between two locations, are barely suited for detecting short traffic incidents such as Jam Waves due to their downstream detection direction, contrasting with the upstream congestion propagation. Segment-based speed calculations prove more effective in identifying significant congestion events. FCD tend to recognize Stop and Go patterns more frequently than loop detectors but often underestimate severe congestion due to their sensitivity to penetration rates and data availability.

INDEX TERMS Congestion patterns, speed reconstruction, traffic state estimation.

I. MOTIVATION
C ONGESTION on freeways is constantly increasing [1].
Many and different traffic sensors have been installed and widely tested for surveillance purposes, traffic planning, online traffic control, and congestion detection.On the one hand, there are stationary sensors such as loop detectors, which detect local speeds and volumes.On the other hand, there are section-based travel time measurements such as WiFi or Bluetooth (BT) sensors, which collect timestamps of moving devices when they pass certain detectors.Using the differences of locations and timestamps, travel times can be computed and, from this, also an average speed.A third emerging method to observe traffic is probe vehicles, which gather trajectories of their traces, so-called floating-car data (FCD).
The review of this article was arranged by Associate Editor Cecilia Caterina Pasquale.
Each detection technology has advantages and disadvantages.Loops, for example, cause high installation and maintenance costs but result in spatially and temporally exact data.Therefore, they are restricted to certain locations only.BT sensors are rather cheap but yield an inaccurate speed distribution because they are section-based (inaccurate location of a bottleneck) and need a second detection at a downstream sensor to determine congestion upstream (inaccurate time of a bottleneck).FCD depend on their penetration rate, which is convoluted by many influences, such as the fleet composition, or congestion warnings, and their deployment costs, which recommend alternative routes.Also, the data quality is influenced by the aggregation rate or the map matching process.If the penetration rate is high, FCD are a valuable input to traffic control and congestion detection.
This paper explores the circumstances, where, when, and under which conditions each of the specific detection technologies performs best.Traffic patterns can be categorized into multiple types and for each combination of traffic pattern and detection technology, we evaluate the effects of the combination.Moreover, we give recommendations where to install (in case of fixed sensors) or when to apply (in case of FCD) the detection technologies.Here, we restrict ourselves to freeways and to congestion detection.To this end, we focus on four congestion patterns and their detection rates depending on the detection technology.
The paper is structured in the following way.First, the state of the art of congestion detection and classification is described.Then, we introduce data from each sensor technology and describe the experimental site.Section IV discusses the results of the detection rates.The conclusion summarizes the work and main findings.

II. STATE OF THE ART
This section gives a literature review on detecting and classifying congestion.Further, we describe an algorithm that identifies congestion patterns in a given space-time domain.

A. CONGESTION DETECTION STRATEGIES
Incident detection algorithms in online traffic control have been researched for decades.An early system is the COMPASS system [2], installed in the Metropolitan Toronto area.It combines two algorithms, the All Purpose Incident Detection and the Double Exponential Smoothing algorithms, which are able to make use of short-term forecasting techniques to detect irregularities of traffic variables such as volume, occupancy, or speed.The authors of [3] discuss an advanced incident detection algorithm with filters and extracts non-recurrent traffic features from recurrent congestion.In [4], freeway bottlenecks are identified.The authors use a speed difference threshold as an indicator of bottleneck identification.In [5], the authors predict traffic congestion and travel times based on a clustering of historical congestion maps and identify consensual days based on 10 months of data on a French freeway.The authors of [6] examine spatio-temporal congestion patterns identification on urban freeways based on FCD.In [7], a multisource data framework for traffic state estimation is described.In [8], the authors estimate link travel times based on sparse GPS data.The authors of [9] make use of data from connected vehicles to give a more accurate traffic flow state estimation.
In [10], the authors compare FCD and loop data on congestion recognition using predefined congestion types [11].They state that on an alpine freeway with a high share of heavy-duty vehicles and low sampling rates of FCD trajectories, FCD do not allow for identifying short-term congestion patterns.The article [12] investigates the estimation of traffic incidents from a hybrid observer method.Traffic incidents are detected by using an improved automatic incident detection (AID) technique based on lane-changing speed.
In [13], the author suggests an AID algorithm based on a comparison of BT travel times.It first filters out too low and too high travel speed values and, second, executes a timedependent comparison to neighbor values on the same road stretch.This algorithm determines the start and end time of congestion or incident and hence provides the times when a warning should be shown to passing drivers on variable message signs.Their algorithm works fast and reliably in terms of incident detection.The authors of [14] describe a novel AID method for freeways, based on the usage of data provided by BT sensors and an unsupervised anomaly detection approach.They validate their proposal using real data from a highway in Tel Aviv, Israel.Their method is based on characterizing anomalous traffic conditions by exploiting the fact that anomalies tend to be isolated.
The mentioned papers deal with the detection of congestion but do not distinguish several categories of traffic patterns.Also, they do not compare more than two detection technologies on a stretch with more than just a few kilometers.In our research, we want to tackle the question, which congestion detection strategy is best for which kind of traffic sensor technology.

B. CONGESTION CLASSIFICATION
Emerging freeway congestion on the roads follow certain patterns and can therefore be classified into only few types.A basic and well-known traffic state classification is the threephase traffic theory developed by Kerner et al. [15], [16], [17], [18], [19], [20].This theory distinguishes three phases of traffic: free-flow, synchronized flow, and wide moving jam.Hence, two of these phases indicate congestion.The wide moving jam phase is a congested stretch that propagates upstream with a constant mean speed at its downstream congestion front as one coherent structure through all other traffic states or bottlenecks.In synchronized flow, the downstream front is usually fixed to a certain location as for example bottlenecks.
Helbing et al. [21], [22], [23] formulated alternative traffic flow theories and defined five congestion phases: homogeneous congested traffic, oscillatory congested traffic, triggered stop-and-go waves, moving localized cluster, and pinned (standing) localized cluster.The authors derive conditions for congested traffic states, starting from the instability diagram of a traffic flow model.They analyze the occurrence, appearance, spreading in space and time, and the related increase in travel time.The terminology of traffic phases is discussed and an empirical evidence of the existence of a phase diagram of traffic states is given.Apart from different discussions and investigations, it is pointed out that combinations of on-and off-ramps create different patterns than a single isolated on-ramp.
In [11], the authors introduced four congestion types, which were also critically reviewed in [24], [25], [26], [27]: Jam Wave, Stop and Go, Wide Jam, and Mega Jam.The single congestion wave is a thin stripe within the spacetime diagram implying a temporarily low speed.Stop and Go describes several narrow stripes representing congestion waves separated by uncongested sections.Wide Jam is a broad area with predominant congestion speed.An extensive area with the domination of congestion speed values is Mega Jam.It represents a wide-spread traffic breakdown and mainly happens during severe traffic or major incidents.Fig. 1 shows examples as space-time representation of the speed distribution.The x-axis corresponds to the time, the y-axis to the space, while the driving direction of vehicles would be from bottom to top.The color represents the speed.This classification in four categories is also used here because it was originally initiated for this kind of freeway (e.g., without general speed limits) and it fits best the local knowledge of the freeway operators.
Concluding, few approaches for congestion classification exist.To the best of our knowledge, they have not been applied to multiple traffic detection technologies to account for their similarities and differences.

C. DETECTION OF CONGESTION PATTERNS
Analyzing emerging congestion just as a binary variable (congestion prevailing or not) is often too rough.Depending on the use cases of the data collection, e.g., for traffic control or traffic planning, a differentiation into more than one type is useful.The following approach introduced by [24], [28], [29], [30], [31] automatically identifies and classifies emerging congestion in a space-time domain.We briefly summarize their methods and results, in order to apply them to various traffic detection technologies in Section IV.First, we describe the identification of congestion clusters, then the assignment of a congestion type to each cluster.

1) CONGESTION CLUSTERS
Fig. 2(a) sketches a space-time representation of traffic.Both space and time are discretized into arbitrary intervals, which results in a grid with a constant speed per cell.Congestion is deemed to occur if the speed of such a cell got assigned to a value less than a critical congestion velocity v crit [32].Since for the identification of congestion only cells containing a speed value below the threshold v crit are relevant, the spatiotemporal speed is divided into cells with speed v ≥ v crit and cells fulfilling the condition v < v crit (marked in red in Fig. 2

(b)).
A congestion cluster describes a set of cells belonging to the same congested regime.Basically, it consists of two requirements, of which at least one must be fulfilled: (1) Cells within this cluster intersect in at least one corner with a congested cell, and (2) the temporal offset of two congested cells is not larger than a time t merge .This temporal offset is defined as the travel time of a virtual vehicle.It is computed via virtual trajectories, which simulate vehicles driving through the space-time domain (Fig. 2(c)).In each cell, the trajectory takes the predefined speed.If a virtual trajectory hits another cluster within time t merge , both are merged.The result is visualized in Fig. 2(d).
The shape of the final cluster k is defined as the convex hull of all congested cells belonging to the identified congestion cluster (Fig. 3).A convex hull makes the cluster's shape and thus the assignment of a cell to a congestion cluster unique.
Each so computed cluster k has a certain area A k in a space-time unit, e.g., km • min.Very small clusters, that do not satisfy a minimum size of A min , are eliminated and are out of scope for the classification.If these single cells were not merged into a more distinctive cluster, they are likely to be data inaccuracies, outliers, or correspond to single slow vehicles not implying congestion.

2) CONGESTION TYPES
After having identified all relevant congestion clusters, each of them gets assigned to one congestion type according to several conditions.To this end, virtual trajectories traverse the entire space-time area starting from the spatial beginning of the considered road stretch, each with a certain time offset t r .They continue with the prevailing speed in each discretized cell.In contrast to the previous section, the simulated vehicles do not stop after a travel time of t merge but continue to the end of the space-time domain to traverse the entire area.
The assignment of a congestion type is done in two steps: (1) determination of a congestion type for each virtual trajectory and (2) determination of a congestion type for each cluster.
Step (1): The speed profile of a virtual trajectory is analyzed.Let t 0 be the time at which a trajectory enters the first congested cell of a cluster and let similarly t 1 be the time at which this trajectory leaves the last congested cell of the same cluster.Briefly, this trajectory experiences congestion between t 0 and t 1 .Let further t JamWave be the maximum duration of a Jam Wave, t MegaJam the minimum duration of a Mega Jam, and n StopandGo the minimum number of speed drop downs.If the trajectory experiences a short incident (t 1 −t 0 ≤ t JamWave ), the congestion type is set to Jam Wave.If t 1 − t 0 is greater than t MegaJam , it is a Mega Jam.Otherwise, the number of speed drops below v crit between t 0 and t 1 is calculated.If there are fewer than n StopandGo speed breakdowns, the congestion type is a Wide Jam, otherwise Stop and Go traffic.
Step (2): Each cluster is embedded in an environment of free-flow conditions and traversed by several trajectories with a uniformly distributed starting time.Fig. 4(a) shows the virtual trajectories inside the cluster: those which intersect the congestion cluster in solid lines, all others dotted.Each trajectory experiences a different speed profile.The profile of the bold trajectory is sketched in Fig. 4(b).Each trajectory is assigned to one of the congestion types.The congestion type of a cluster is basically determined driven by a frequency-based assignment, refined by some additional requirements [29], depending on the number of distinct congestion types n 2types and n 3types , respectively.If the congestion type could not be determined uniquely, it is set to a generic type Mixed.
The presented methodology determines isolated congestion clusters, each having the shape of a convex hull of all affected cells, a certain area, and one of four congestion types.We refer to the original papers for a more detailed description [24], [28], [29], [30], [31], [32].The rest of this paper is based upon this methodology.To the best of our knowledge, none of the algorithms described in literature, deals with 3+ detection technologies on a stretch of 100+ km for a period of several months.

III. DATA
We apply the previously described congestion identification and detection methodology to multiple traffic sensor technologies along a freeway stretch in Germany.This section describes all available data sets including their preprocessing and the experimental site.

A. TRAFFIC DETECTION TECHNOLOGIES
We compare the congestion detection rate of three sensor technologies, loop detectors (LD), Bluetooth (BT) sensors, and Floating Car Data (FCD), as these are among conventionally available data collectors on roads [33].Additionally, a data fusion approach combines the resulting speed distributions and is also compared as a fourth speed data set.This section describes each measurement technique in general and gives an overview of the applied data sets.and maintained properly.Often, they are used in a speedtrap configuration where one loop is followed by another at a known spacing.However, their usage remains limited to just a few corridors due to the high installation and maintenance costs.Their resulting raw speed distribution is sketched in Fig. 5. Let X × T be a space-time domain.Every detector l 1 , l 2 , l 3 observes passing vehicles continuously.Their speed values are indicated applying the color bar on the right.Spatially between loops and temporally between single measurements, no speed information is available.
The raw data set used here has a very low error rate comparing to the ground truth derived from probe vehicles [28].It is interpolated and smoothed using the well-known approach Adaptive Smoothing Method (ASM) by Treiber et al. [34], [35], [36].Wherever the spacing of two neighbored loops is ≤10 km, these areas are spatially and temporally interpolated to obtain the full speed distribution.Briefly summarized, the ASM smooths raw data of a sparse input source in two traffic-characteristic directions: v c denoting the wave speed in congested traffic conditions, and v f denoting the wave speed in free-flow conditions.In a discrete time-space domain, the resulting complete speed matrices V c (t, x) and V f (t, x) are combined cell-wise: The weight w(t, x) is adaptive and favors low speeds: with V thr a threshold where weight w(t, x) equals to 0.5 and V a parameter to control the steepness of the weight function.The so received data set provides average speed values in cells with lengths of 500 m and 1 min.

2) BLUETOOTH DATA
WiFi or BT sensors register each device that passes the sensor by reading its unique media access control (MAC) address.If re-identified at a further sensor downstream, the matching of both hardware addresses allows the derivation of a travel time and therefore the average speed of the device in the respective segment [37], [38], [39], [40], [41], [42].Fig. 6 illustrates the space-time domain of resulting speeds.
Each line corresponds to one vehicle or device re-identified at two or more locations l 1 , l 2 , l 3 .The slope and the color of the line corresponds to the average segment speed of the vehicle.BT installation is cost-efficient, as both driving directions can be covered by one sensor solely.Moreover, origin-destination relations can be derived while tracking a device over a long distance.However, the sensors collect information from a sub-sample of the vehicle collective only (depending on the penetration rate).Since they are conceivably placed several kilometers apart from each other, the average speed can be less granular.BT sensors detect in a certain range in their surrounding, which involves a fuzzy location of the detected device.Furthermore, traffic flow cannot be detected.BT sensors determine travel times at the end of a segment.This results in delayed detection of increasing travel times during congestion, which is a major drawback for accurate real-time information [43].
Note that the driving direction cannot be determined based solely on a single BT measurement.By matching a corresponding entry in the database for an individual hardware address hash value, the direction of travel is known.The raw data set is preprocessed and interpolated using the Low-Resolution Travel Time Smoothing Method (LTSM) [44].This method considers travel times in predefined cells.All paths crossing a certain cell determine the average speed of this cell, weighted by their spatio-temporal share inside the cell.This way, data errors are tried to be eliminated.The result is a speed distribution with discretized cells, again of sizes 500 m and 1 min.
3) FLOATING-CAR DATA FCD, also called probe data, are gathered from vehicles or smartphones that determine their current position via GNSS and report it on a regular basis to a central server.Time and space differences allow for reconstructing the probe's speed profile on its route.Fig. 7 illustrates the space-time domain resulting from FCD.Each trace corresponds to an FCD trajectory.Its color is smooth due to the frequent location updates every few seconds, depending on the fleet.
However, the amount of FCD observations depends on the penetration rate, since FCD represent only a sub-sample of the whole fleet.This limitation becomes less and less relevant with sample rates of some percent of the total traffic volume, in particular on freeways during peak hours.Another aspect is the time delay.FCD-based traffic state estimation is quite convoluted with many influences such as update frequencies from vehicles to the back-end server, the fleet size of floating cars, the current traffic flow, and the provider treatment (Fig. 7).The data set for this study is provided by BMW [45], [46].Data are interpolated to a space-time distribution using the Phase-Based Smoothing Method (PSM) [47].The PSM is an approach that is based on concepts of the ASM.It was developed to reconstruct space-time traffic speeds with higher accuracy given only FCD.It utilizes findings summarized by the Three-Phase traffic theory [16], [19] in order to distinguish between localized and moving congestion.The method outperforms the ASM in a study reported by [47].We refer to the original paper for a detailed method derivation and evaluation.The output result is again a smoothed speed distribution of discretized cells with sizes 500 m and 1 min.

4) FUSED DATA
In addition to the three separate speed distributions, a fused data set is considered.The fusion is conducted as a simple superimposition of all data sets.It takes the average mean per cell over all data sources, followed by an ASM-interpolation in order to fill data voids, where none of the data sources yielded any speed values.The result is again a speed distribution in a uniform grid with step sizes of 500 m and 1 min.

B. EXPERIMENTAL SITE
As a test site, the German autobahn A9 between Munich and Nuremberg is chosen with a stretch length of 157 km.Munich is located at km 529, Nuremberg at km 372.As an example, the detection rate is discussed based on data collected on May 29, 2019 in Northbound driving direction (Munich towards Nuremberg, decreasing kilometers) during severe traffic congestion.Each data source was obtained as a cell-wise speed representation with a grid size of 500 m and 1 min.All in all, time-discrete data of 70 loop detectors and 35 BT sensors are available.Raw FCD have a data availability rate of 24% (share of cells containing speed data).The direction of travel in the contour plots is illustrated as 'from bottom to top'.Therefore, the plots are depicted with a reversed vertical axis.Both BT and FCD are available along the entire road stretch.Loop detectors are available in the Munich and the Nuremberg area, and few in the middle around the city of Ingolstadt (km 450-470) (Fig. 8).

IV. DETECTION RATE OF CONGESTION PATTERNS
This section discusses the detection rate of each congestion type depending on the data source.It investigates how each data set performs in identifying each congestion type.If loop detectors are more than 10 km apart from each other, the space in between is not interpolated.These locations without data can be accepted as part of the input speed matrix.To apply the algorithm, all cells representing an unavailable location over the entire period as well as unavailable times (e.g., time zone offsets) are set to a freeflow speed v freeflow .During the assignment of a congestion type to an identified congestion cluster, all other areas except the cluster are set to free-flow conditions (Fig. 3).Similarly, these locations or times of undefined speed values can be filled with free-flow defaults.Note that assigning a freeflow value is not valid for single undefined speed value cells surrounded by defined speed value cells.This interpolation of particular cells is done using the smoothing algorithms described earlier.
The identification of congestion clusters and the assignment of a congestion type to each pattern utilize the parameters given in Tab. 1.In order to minimize the number of assignments to Mixed -the undefined congestion type n 2types and n 3types are set relatively low in an empirical way [29].As assessment method, number and size of each recognized congestion type is considered per data source.The size of a congestion cluster is defined as the area of its convex hull.

A. LOOP DETECTOR DATA SET
Applied to loop detectors' speed measurements, the results of the congestion type assignment show seven congested areas on this day, once Jam Wave, once Stop and Go, twice Wide Jam, once Mega Jam, and twice Mixed.The Mixed cluster at km 380 in the late evening is traversed by four trajectories if the temporal offset t r is set to 5 min.Thereof, two Jam Wave and two Wide Jam trajectories are identified.Therefore, the Mixed cluster remains unclassified.The other Mixed cluster at 16:00 at km 450 is crossed by three trajectories, Stop and Go, Wide Jam, and Jam Wave types, respectively.
The areas without detectors (between km 480 and km 470, km 450 and km 410, km 410 and km 400) are set to free-flow conditions such that congestion detection is not false positive (Fig. 9(a)).The tails of the jam occurring downstream of km 450 (in the unequipped region) appear as small single jam events instead of one merged cluster due to the detection area.

B. BLUETOOTH SENSOR DATA SET
Interpolated travel times derived from BT measurements detect more congestion clusters than loop detectors because single slow vehicles, which were not eliminated beforehand have a greater influence.Therefore and because the sensors are available from a longer road stretch than the loops, the results with the same parameterization show 21 congestion clusters, of which Jam Wave is assigned to 2, Stop and Go to 13, Wide Jam to 2, and Mega Jam to 2 clusters.Two clusters cannot be assigned to a unique congestion type and are set to the undefined congestion type Mixed.Both of them are touched by one Jam Wave and one Stop and Go trajectory each.
This prompts a need for a sensitivity analysis and some parameter adjustments, specifically the increase of A min and the decrease of t merge .With an average spacing of approximately 5 km between the BT sensors, it is suggested to set A BT min = 30 km•min.This value corresponds to (spatially short) congestion of at least 6 min or spatially more extended congestion that are both reasonable for a congestion definition measured by BT. t merge separates congestion occurrences.Especially during night hours when -supposedly -slow vehicles are moving around, the differentiation should be set sharp.Therefore, t BT merge = 1 min is recommended.Both parameter changes lead to more precise convex hulls of the clusters and reduce high (uncongested) speeds inside the congested clusters.
After these adjustments, the results for actual congestion (not derived from single vehicles) return 10 clusters, of which 1 Jam Wave, 5 Stop and Go, 2 Wide Jam, and 2 Mega Jam clusters are assigned (Fig. 9(b)).These analyses yield an underestimation of the congestion type Jam Wave.However, larger jam events can be detected reliably.One main finding is that Jam Wave is a pattern that is least detectable through travel time measurements with large sensor spacings due to the upstream propagation of congestion waves in contrast to downstream travel time measurements.

C. FLOATING CAR DATA SET
Fig. 9(c) depicts the reconstructed traffic speeds based on FCD.Speed measurements during the first two hours of that day are not available due to time zone shifts and are therefore set to free-flow conditions.The result shows 3 Jam Wave, 5 Stop and Go, 1 Wide Jam, 2 Mega Jam, and 2 Mixed clusters.The Mixed cluster at 20:00 at km 400 is traversed by two Jam Wave and two Wide Jam trajectories.The larger cluster at 9:30 around km 372 is passed by 3 Jam Wave, 3 Stop and Go, and 2 Wide Jam trajectories.In both cases, n 2types and n 3types are not fulfilled, respectively.
In contrast to the loop data set, FCD detect Stop and Go more often, specifically between the Ingolstadt and the Nuremberg area.This is mainly caused by different detection areas.Locations, which are considered by both detection technologies, are equally well detected in terms of their sizes but vary in the assignment of an appropriate congestion type.The BT data set, which does not have limitations in the common detection region, detects at the same locations and similar congestion types as FCD but varies in number and size of the detected congestion.BT detections are bound to their segment-based measurements, which means that congestion ranging over more than one segment is more often interpreted as single separate congestion events, e.g., Wide Jam and Stop and Go in the BT data set between km 460 and km 440 from 13:00 to 16:00.

D. FUSED DATA SET
The data set fused from loops, BT, and FCD computes 10 congestion clusters (Fig. 9(d)).It shows one Jam Wave, 6 Stop and Go, 1 Wide Jam, and 2 Mega Jam clusters.
Congestion occurring at km 380 from 16:00 to 19:00 is interpreted as Stop and Go by all three data sets, whereas the fusion approach detects a Wide Jam.The segment-based travel times average the speeds prior to the congestion (yellow segments), which narrows the fused congestion.In contrast, the short congestion event occurring at the same location at 22:00 has small sizes in loop detector (LD), FCD, and the fused data set, whereas BT data detect significantly larger congestion.Congestion between km 440 and km 430 between 15:00 and 16:00 cannot clearly be detected by BT data, which create a yellow and a red phase but is well detected by FCD.The fusion creates a common area between the congested (red) BT data and the entire FCD cluster but this way, overfits this congestion into three separate parts.

V. RESULTS AND FINDINGS
Fig. 10 conjointly shows all recognized congestion clusters and their corresponding convex hulls.In general, clusters are detected in similar space-time regimes.The recognized Mega Jam at km 480 at 20:00 (region 3 ) is classified as a Wide Jam by LD because the downstream detection equipment ends at km 480.Analogously, the subsequent equipped area starts around km 462 and therefore, congestion can only be detected in a narrow stripe.Therefore, LD do not detect the whole congestion area.
Summarizing, all considered detection technologies are able to identify congested areas with a different level of detail.The basic requirement is a sufficiently small detector spacing.Reconstructed BT speeds are rarely suited to detect Jam Wave due to their downstream travel time propagation in contrast to upstream congestion propagation.ASM-smoothed loop data detect congestion reliably and are able to assign an appropriate type reasonably.FCD prefer congestion type Stop and Go over the other types, especially in comparison with BT, which is the detection technology with most overlapping locations.

A. QUALITY ASSESSMENT
The quality assessment of the recognition strategy utilizes the area of the convex hull defined by a cluster.Let a cluster k have two properties: a congestion type ct k and an area A k .The size of a cluster's area corresponds to the sum of spatio-temporal areas of all affected cells in the cluster.The size of all identified clusters is summarized in Table 2. Additionally, the related region according to Fig. 10   In regions where data from all detection technologies are available, one Mega Jam is identified (region 1 ).Its size varies between 3,900 (both LD and BT), 4,600 (FCD), and 4,000 (fusion) km•min.Hence, it is comparable and of similar size.BT prefers Wide Jam over Stop and Go, especially visible in the area where loop detectors are not installed.In region 6 , the BT Wide Jam, conjoint with the small Stop and Go cluster directly connected, sums up to an area of size 1528.25 + 141 = 1669.25,the corresponding FCD Stop and Go cluster amounts to 1770.3 km•min.The two jam waves (region 8 ), recognized as one Stop and Go pattern in FCD at 17:00 at km 400, are not detected by LD (out of their region) and cannot clearly be captured by BT.

B. DISCUSSION
This section discusses the suitability of the three detection technologies loop detectors, BT sensors, and FCD.The recognition of the four congestion types Jam Wave, Stop and Go, Wide Jam, and Mega Jam are compared on a freeway stretch.The evaluation shows that each detection technology has a different level of recognition.
In general, loop detector data are able to detect all kinds of congestion.Their temporally high accuracy and their small spacings in this study lead to an exact traffic state.If loops are installed in a short range of only few kilometers, it challenges the detection of a Mega Jam due to its greater extensions.Only if a Mega Jam lies ideally in this spatiotemporal area, it can be detected reliably.
BT measurements are not suited to detect a single or several jam waves (Stop and Go).The propagation direction of jam waves is upstream with a constant speed of approximately 15 km/h, in contrast to the travel time measurements which are downstream awaiting the second detection at the end of the segment.Especially, when the speed falls below v crit for a short period in a segment, BT speeds result in two options: Either congestion is not detected at all because the segment speed averages to a value greater than v crit .Or a congestion event supposed to be Jam Wave decreases the segment speed to a value below v crit which then increases the probability of detecting Wide Jam because of the large spacing of several kilometers.The same holds for Stop and Go traffic.Multiple jam waves after another are not reliably determined by travel time measurements.Wide Jam and Mega Jam are appropriately suited to be detected by BT.If congestion extends over segment boundaries, it is likely that BT data detect several individual congestion events, again due to the slope of the trajectories in driving direction and the congestion propagation in the opposite direction.
FCD are able to detect all congestion types well as long as the penetration rate of probe vehicles is high enough.During night hours, little or no speed information can be gathered.If the penetration rate is low, few vehicles define the speed information.In this case, false velocities, e. g. due to a wrong map-matching, cannot be averaged.If the vehicle fleet is large enough, a reliable speed reconstruction along with the determination of congestion types is possible.A drawback here is that the data set comes from a car manufacturer, which provides their customers with in-car traffic information.Prior to congestion, drivers receive an incident warning and are more likely to leave the freeway or use an alternative route.Therefore, the fleet size decreases upstream of congestion and increases again downstream of it, which reduces the number of vehicles driving through a congested area.Since in the near future, probe vehicle data will play a major role in terms of traffic data collection, the effects on the data quality should be tackled.
To conclude, data sources recognize congestion events in a similar way.Detected clusters have a comparable size, however, some sources prefer more and smaller clusters whereas others prefer fewer and larger clusters.

VI. SUMMARY AND CONCLUSION
This paper deals with the detection rate of congestion patterns comparing multiple traffic sensor technologies such as speed measurements from loop detectors (LD), travel time measurements from Bluetooth (BT) sensors, and floating-car data (FCD) from probe vehicles.A state-of-the-art algorithm is applied that detects and categorizes congestion events in a given space-time domain.This method assigns one of the four congestion patterns Jam Wave, Stop and Go, Wide Jam, and Mega Jam to each identified congestion cluster.The novelty of our paper is to apply and evaluate this methodology to 3+ detection technologies on a stretch of 100+ km.
The evaluation of the detection rate of each detection technology and each congestion pattern shows that BT sensors are barely suited to detect short incidents such as Jam Wave or Stop and Go traffic.This is mainly due to different directions: congestion propagation is upstream whereas BT travel times are measured downstream.Further, BT measurements are segment-based, which results in congestion mostly restricted to the segments.If congestion extends to more than one segment, it is likely that several individual congestion clusters are identified.FCD and loop detectors (LD) classify identical congestion types if the detection area is not restricted.LD are not available throughout the entire stretch of the experimental site and therefore sometimes detect parts of a congestion event only.FCD congestion clusters tend to be larger than LD clusters.These findings will be used for recommendations where and when which detection technology should be installed or applied.While a hot spot analysis per data source per congestion type [29] yields locations and times, which are congestion-prone, these congestion patterns should be detected by the suitable detection technology.BT sensors should be installed, where larger congestion events such as Wide Jam or Mega Jam often occur, whereas FCD should be used for the detection of short incidents such as Jam Wave and Stop and Go.Loops, which are not as cost-effective as other detection technologies, should not be used while equipping a new road stretch.Moreover, they should be replaced by other stationary sensors such as radars or video detection systems in traffic planning.If well calibrated, these kinds of sensors can detect each congestion pattern well if available over a larger detection area with small sensor spacing (few hundred meters only) [30].
Depending on the use case of the data collections, the exact locations and times are more or less relevant.For instance, different control strategies (ramp metering etc.) require a higher level of detail.Therefore, loop detectors are recommended to be placed directly upstream of a ramp entrance.In contrast, for determining travel times in noncongestion-prone areas, a lower level of detail might be sufficient and Bluetooth sensors could be used.
Our next steps include a matching between reasons for the emergence of congestion and its resulting congestion type.Knowing the cause of an incident will improve the prediction of congestion types at certain locations or times.Further, a virtual reality device calibrated using data from this study can simulate different congestion patterns and their effects.

FIGURE 4 .
FIGURE 4. Assignment of a congestion type per cluster.
1) LOOP DETECTOR DATAInduction loops measure local velocities at a fixed point of the road.They detect speeds accurately when installed

FIGURE 10 .
FIGURE 10.Convex hulls of congested areas per data source.

TABLE 1 . Parameter values to identify separate congestion clusters.
is indicated.