An Indoor Global Localization Technique for Mobile Robots in Long Straight Environments

Scan matching methods have been widely applied in the fields of autonomous localization and mapping. However, in structured environments where feature differences are less significant, such as long straight corridors, conventional positioning algorithms often suffer from characteristic mismatching, resulting in lower accuracy. As such, a new global localization algorithm, based on environmental difference evaluation and correlation scan matching fusion, is proposed in this study. In this process, the surrounding space was evaluated using a priori understanding of the environment based on a linear fit. Corresponding evaluation and positioning results from correlation scan matching were then modified using dynamic selection and a posture updating strategy. The performance of the proposed technique was compared with other conventional methods using open datasets exhibiting long straight features and a series of tests conducted in a physical corridor. Results showed that the proposed algorithm could effectively improve localization accuracy in narrow environments. The translation and rotation absolute pose errors were reduced by an average of 27.29% and 25.82%, respectively, compared with a correlation matching approach that does not consider the surrounding geometry. These results suggest the proposed technique offers higher adaptability and positioning accuracy in narrow environments.


I. INTRODUCTION
The simultaneous localization and mapping (SLAM) building problem involves placing a mobile robot in an unknown starting location in an unfamiliar environment. The robot must then incrementally construct a map of this environment while simultaneously using this map to calculate its absolute location [1]- [3]. Mobile robots can output their current position quickly and accurately using an existing map and the surrounding environment, which are critical for the subsequent realization of path planning, obstacle avoidance, and other functions. It is also more practical to track the pose of a robot against a known map, which could be constructed from the workspace of other mobile robots.
In recent years, the scanning matching global localization method, based on 2D laser input data, has played an important role in the positioning of mobile robots. Scan The associate editor coordinating the review of this manuscript and approving it for publication was Hui Xie . matching seeks a set of translation and rotation parameters to maximize the overlap of two scanned point clouds after alignment [4]- [6].Two or more consecutive scanned laser point clouds can then be unified in the same coordinate system, by solving for the coordinate transformation relationship. The current scanned point cloud can also be registered with an established map to recover changes in robot body posture. This process is illustrated in Figure 1.
Existing scan matching techniques can be divided into three categories: point-based, mathematical feature-based, and likelihood field estimation-based algorithms. The iterative closest point (ICP) model and its variations represent point-based scan matching algorithms, which are also more sophisticated approaches. In this process, the basic model is provided with a scanning point cloud and a reference point cloud for current laser data. These are used to solve the rigid body transformation matrix and minimize the Euclidean distance for corresponding points [7], [8]. The most representative scanning matching based on mathematical features is the normal distribution transform (NDT), which maps discrete 2D points from a single scan into a segmented continuous and differentiable probability density, defined on a twodimensional plane [9], [10]. The scan matching method based on the likelihood field estimation constructs a likelihood field using existing grids. It then specifies the search space, calculates a score for each pose in the search space, and determines the variance of a match using the pose score. Correlation scan matching (CSM) is an application of the likelihood field raster technique [11]- [13]. The primary advantage of this approach is that current scan data and multi-scan historical data can be used to improve matching accuracy. Global optimization can then be ensured through grid division. However, in the case of small environmental differences, such as long straight corridors, depth information acquired by a laser rangefinder does not change significantly with time, which complicates attitude estimation [14], [15].
Scan matching algorithms typically do not consider the effects of long straight environments on localization accuracy. This is particularly true for conventional CSM algorithms that require the use of adjacent likelihood regions and boundary features for registration. However, long-term corridor features do not include enough information to estimate complete poses, which results in fewer matches and larger positioning errors. Environments with linear and planar structures, such as offices and factories, often contain large numbers of long and straight spaces. These natural features should be considered to improve positioning accuracy for industrial applications [16], [17].
This study presents novel a localization technique suitable for long and straight indoor environments. The approach fully considers the characteristics of such spaces and evaluates the corresponding geometry with a line-fitting algorithm. A priori understanding of the current robot environment was acquired using line fitting. The localization results exhibit varying dependence on the dynamic selection of a robot posture updating strategy. The contributions of this work focus primarily on the following three points: 1) A prior understanding of the robot's current environment is critical for global positioning in long straight corridors. This study proposes a new methodology for evaluating the characteristics of long straight environments, based on linear fitting, using only laser sensors. 2) In long corridors with no obvious textural features, global positioning based on scan matching lacks forward and backward constraints, preventing a robot from perceiving its own X direction coordinates and causing position drift. To solve this problem, correlation scan matching algorithms and line fitting were combined in this study, to develop a more accurate global localization result for narrow corridor environments.
3) The method proposed in this article was tested in both virtual and physical environments. Tests performed using a public data set showed this method is more robust than ICP or combined RANSAC and more accurate for judgments in long straight environments. Test results in physical environments showed improved reliability in static corridors. The remainder of this article is organized as follows. A description of related work is provided in Section 2. Theoretical definitions, derivations, and the environmental detection algorithm workflow for 2D laser scanning matching are included in Section 3. A validation experiment is then conducted, and the results are analyzed in Section 4. Section 5 concludes this article.

II. RELATED WORKS
This section summarizes existing 2D laser scan matching methods and related line fitting models. Some of the most relevant and recent works are discussed, including point scan matching, mathematical feature scan matching, and likelihood field scan matching. A brief review of line fitting is also included.

A. SCAN MATCHING
Scan matching is a common technique for registering laser data acquired in different poses and determining relative pose transformations. The laser source is used to scan a flat object surface at a fixed angular resolution α i , returning several successive measurements {ρ i | · · · ρ i−1 ,ρ i , ρ i+1 · · · }. The values of laser scanning points in the coordinate system of the mobile robot can then be determined from the angular resolution and distance measured by the laser. The position of the mobile robot in the global coordinate system can be acquired through a coordinate transformation, using the relationship between the current position and the next position, represented by(X i , Y i , θ). Scan matching was typically conducted using ICP and its variations, NDT or CSM.
ICP is a popular optimal registration algorithm based on the least-squares method. In this process, corresponding relational point pairs are selected repeatedly for two sets of laser-point cloud data with similar local geometric features. An optimal rigid body transformation is then calculated until the convergence requirements for accurate registration are met [18], [19]. Zhu et al. proposed an indoor robotpositioning model based on ICP scan matching, which was used to modify a robot motion model and reduce the uncertainty of robot attitude in the prediction step [20]. Censi described an ICP variation using point-to-line metrics and a precise closed-form that minimized the corresponding metrics [21]. Pavlov et al. [22] proposed a new methodology for accelerating ICP, based on the Anderson acceleration technique. This algorithm is typically faster than the standard Picard iterative method. Although ICP can achieve high accuracy in scan matching, its strong dependence on the quality of initial data associations reduces its applicability in practical applications. In addition, its computational efficiency and robustness need to be improved further [23], [24].
In addition to ICP-based models, scan matching methods such as NDT use mathematical properties to characterize data and inter-frame pose changes. Biber and Strasser applied NDT to laser scan matching between frames, defining a score for the probability density after registration [9], Li et al. addressed the difficulty of extracting typical features in unstructured rescue environments, using NDT scan matching to estimate probability distributions for laser data. Their proposed NDT-EKF algorithm combined NDT scan matching with the EKF framework [25]. He et al. used NDT to match the current keyframe with a local map and construct a new map at a lower frequency [26].
Both ICP and NDT exhibit certain requirements for initial values. As such, correlation based on exhaustive matching performs well in a variety of settings. Olson described a matching algorithm based on the correlation of two lidar scans. This approach resolved the issue in a probabilistic framework and identified a rigid-body transformation that maximized the probability for observed data [12], Olson also developed a new multi-resolution scan matching approach that makes exhaustive matching practical, even for large positional uncertainties [11]. Hess et al. combined branch definition algorithms to increase search efficiency based on CSM scan matching [27]. Song et al. introduced a global localization strategy for 2D laser ranging robots, based on depth-first searches and grid diagrams [13].
Scanning matching, a common positioning technique, has been used in a variety of studies that rely on external sensors. For example, Yuxiang et al. developed an indoor positioning system, based on Wi-Fi signal strength, and implemented a fingerprint matching algorithm derived from the improved KNN algorithm and Gaussian process regression [28]. Codas et al. proposed a localization model based on a particle filter, by integrating RFID sensors and odometers [29]. This algorithm successfully localized a mobile robot with bounded error estimation, while dead reckoning approaches using only odometry caused the error to diverge. Relying solely on an odometer during global positioning will lead to the accumulation of positioning errors and inaccurate results. Martin et al. leveraged recent advances in deep learning and variational inference to correct dynamical and observation models for state-space systems [30]. This approach trains Gaussian processes on the residuals between the original model and the ground truth, and has been applied to publicly available datasets for robot navigation based on two-wheel encoders, a fiber optic gyro, and an inertial measurement unit.
Most location techniques using external sensors require changing the existing environment, which is often impractical. In addition, odometer-based dead-reckoning positioning models calculate new position values using only the value derived in the previous step. As a result, the error and uncertainty of position estimation are cumulative and increase with time. Requiring odometers to coordinate with other sensors, such as an RFID [29], reduces the integrity of the environment. In other words, before the positioning process can occur, the corresponding sensing equipment must be assembled in the environment where the robot is to be located.
The primary goal of this article is the realization of robot positioning in long straight corridors, based on scanning matching, without changing the environmental state. Both ICP and NDT models depend heavily on the initial pose, which can easily lead to local minima. The primary advantage of the CSM method is that current-frame laser data and multiframe historical data are used to improve the accuracy of probability matching, which can maintain global optimums in raster divisions. However, CSM is prone to mismatching when environmental differences are small [17]. These salient features should be considered for global localization in artificial environments composed of linear features, such as homes, offices, and factories [16].

B. LINE EXTRACTION
A line extraction model, based on 2D laser data, was found to be the most useful in describing the long and straight characteristics of corridor environments. Several different line extraction algorithms have been proposed in recent years. Only basic versions of these algorithms are discussed here, chosen for their performance and popularity in mobile robotics and feature extraction.
Borges proposed a segmentation method for line extraction in 2D range images [31], This model has been used for mobile robot navigation systems and dynamic map building. Zhang et al. used the split-and-merge technique for map building and described localization techniques for a mobile robot equipped with a 2D laser rangefinder [32]. Schaefer at el. used a novel probabilistic method to extract polylines from raw 2D laser data [16], The goal of this approach was to determine a set of polylines that maximized the likelihood of a given scan. Gao et al. proposed a line segment extraction algorithm, using laser data, based on seeded region growing [33]. RANSAC has been used to roughly estimate 3D transformations by finding an optimum consensus 209646 VOLUME 8, 2020 model [34], a common linear fitting method. This algorithm has also been used for robust fitting of models in the presence of data outliers. The primary advantage of RANSAC is its robustness as a generic segmentation method that can be used with several types of features once the feature model is determined [35]. Samuel et al. proposed a weighted linefitting technique that weights the influence of each point on a fit, according to its uncertainty, which can be derived from sensor noise models [36].A key feature and contribution of this method is the concrete formula for the covariance of the line segment fitting while allowing for individual weighting of each measured point. Therefore, we use the weighted line extraction method to describe the long straight corridor environment.

C. PROBLEM STATEMENT
Although a variety of scan matching and line extraction techniques have been proposed and implemented, it is still necessary to improve global localization accuracy in narrow corridor environments. When a mobile robot is in this type of environment, with no other references, the input of two laser beams returns the same scan point. As a result, the matching algorithm returns the same score at all locations in the mapped area. The region where the mobile robot is located can then not be correctly matched, regardless of whether a scan line is added or not. As shown in Figure 2 The single geometric structure of long corridors results in a loss of forward and backward constraints in current LiDAR SLAM techniques when point cloud matching is done at the front end. This prevents the robot from perceiving its own X direction coordinates in the corridor, causing position drift. Misjudging the correlation of any laser point in the two beams, while leaving the correlation of other points unchanged, will produce a new correlation configuration that differs from the ideal value. As in visual SLAM, incorrect data associations can lead to catastrophic failures in the localization processes [37]. An incorrect global pose relationship RT will then reduce the accuracy of a pose. As such, this study attempts to improve global localization accuracy for mobile robots in narrow environments. Line extraction is adopted to evaluate these straight environments and the dependence of scan matching results on the dynamic selection of a robot posture updating strategy is investigated.

III. METHODOLOGY
In narrow environments, it is necessary to determine geometric characteristics using line extraction. Scans can then be used to match the output global poses. An overview of the proposed approach is shown in Figure 3. The framework can be divided into three steps: scanning matching, environment evaluation, and pose fusion. Scanning matching included the most popular correlation scanning matching algorithms. Environmental assessment included linear fitting and difference assessment methods. Pose fusion occurred after environmental assessment. The global localization problem based on scan matching can then be described as follows. The same map features are observed at different times to acquire rotation and translation changes for the posture of the robot between adjacent frames.

A. CORRELATIVE SCAN MATCHING
Unlike the NDT and ICP algorithms that depend heavily on initial pose, correlation scan matching is robust to large initialization errors. The target of the CSM is defined by starting a search from the previous pose frame and finding an optimal pose transformation, such that the position of the scan point in the map corresponds to the grid with the largest possible degree of occupation. This approach utilizes a probabilistic framework to search for a rigid transformation that maximizes the probability of having the observed data.
Global localization based on scan matching can be defined as the prediction of trajectories under the conditions of a known map. The probability model can be described as p(x t |x t−1 , m t−1 , z t , u t−1 ) and used to estimate the robot position at time t, given the robot position map m t−1 , a control u t−1 at the previous instance, and the observation z t at the current instance. In this context, Bayes rule can be expressed as: Since the sample weight factor in p(z t |x 1:t−1 , m 1:t−1 )will be normalized to 1, and is independent of x t it will be canceled in the process of algorithm standardization. As such, no calculation is required.
(2) VOLUME 8, 2020 The term p(x t |x t−1 , u t ) can be acquired using IMU and an odometer. The observation model p(z t |x t , m 1:t−1 ) then becomes the primary focus of algorithm development. Each lidar scanning observation z t is independent of the others. As such, the observation model can be described as: The proposed global localization is based on a calculation of maximum likelihood estimation for p(z|x t , m) Therefore, equation (3) can be used to express the problem as [38]: wherex t is the final output pose, while m grid represents known map. Accurately defining this term is the primary goal of maximum likelihood estimation in the proposed algorithm. The matching and positioning processes for known maps involves outputting a transformation relationship for the coordinate system of the mobile robot, when the laser in the current frame is aligned with the point cloud in the map. The optimal rigidity transformation is assumed to bex t = ( x, y, ϕ). The position of the radar point after a rigidity update can then be expressed as: The process of identifying an optimal rigid transformation can be understood as searching for all possible pose orientations in a search interval, then scoring the results. The pose with the highest score represents the current optimal transformation result, as shown in Figure 4. The process FIGURE 4. The process of exhaustive matching, the angular directions, the X axis, and the Y axis of the triple cycle, matching the optimal pose. When the best registration is found, the translation and rotation will be output.
for high-resolution matching scores is consistent with that of low-resolution scores. Finally, the optimal conversion is output for the score with the highest match.
The conventional multi-resolution CSM model is described in Algorithm 1 [11], [12], [27]. The required input includes a grid map (to be located) and current laser data, as well as a fuzzy value representing the initial pose. The output is a transformation relationship developed from the optimal matching score. The input laser data are initially rasterized and blurred using a Gaussian kernel. Each raster is then assigned a score based on the mapping results. Exhaustive matching is then conducted and the score is calculated for θ, x, and y.
The third line of Algorithm 1 is the Gaussian smear process. If a grid coordinate is selected by a laser scan, the surrounding grids will also be assigned according to a Gaussian distribution (the correlation step). Lines 4-9 and 10-15 are two exhaustive-search matching processes with differing grid 209648 VOLUME 8, 2020 precisions. The final output is the optimal pose transformation in the laser coordinate system and the corresponding matching score.

B. LINE EXTRACTION FOR ENVIRONMENTAL EVALUATION
Man-made environments such as houses, offices, or factory floors are typically composed of linear structures. As such, line extraction was used to evaluate flat surfaces in narrow spaces, distinguishing their geometries from similar structures for scan-matching positioning in complex environments.
In this study, the straightness of an object was described by the slope of a fitted line and the length of the object was represented by the length of the line segment. Weighted line fitting is a common 2D line extraction algorithm based on expectation maximization [31]. This approach weights the influence of each point on the overall fit, according to its uncertainty derived from sensor noise models. The robot moves through multiple poses where are represented by (X i , Y i , θ). A frame of laser scan contains fixed angular resolution α i and several successive measurements ρ i . the measured distance to the environment's boundary in the direction denoted by β i . Generally, ρ i be comprised of the true range and additive noise term. At each pose, the robot gathers a range scan and the scan point coordinates are described in the laser coordinate frame: where ε ρ and ε β are considered the distance noise and angle noise, respectively, which are assumed to be a zero-mean Gaussian. The noise model for the sensor uses a maximum likelihood approach to formulate a general strategy for estimating the line of best fit from a set of non-uniformly weighted range measurements. This process is shown in Algorithm 2 below. Linear parameters were generated randomly from the laser data and weights for the remaining points were calculated using the linear model, until the maximum number of iteration steps or convergence was reached. The complexity of this algorithm is low and, as shown in Figure 5, it was replicated in TurtleBot2 with SICK LMS111.
However, fitting a line segment to the 2D laser data is not the primary goal. The collected data were used not only for scan-matching positioning but also for subscription during the fitting process. Our objective was to evaluate the environment encountered by the mobile robot using these linear parameters.
The slopes k 1 and k 2 for segments L 1 and L 2 were calculated by line fitting, as shown in Figure 6. The values of k 1 and k 2 determined how the environment was classified (i.e., long straight), as follows: The term was used to quantify varying degrees of environmental properties and distinguish narrow structures from VOLUME 8, 2020  the general background. When was close to 1 and ψ was close to π/2, differences in sample points were obvious and the environment was not considered straight. When was close to 0 and ψ was close to π, the corresponding structure was assumed to be long and straight. As such, the mobile robot is instructed to process the location results from scanning matching.
The results of line fitting may correspond to continuous short lines, in which case the environmental characteristics are relatively evident and superior positioning information can be acquired with CSM. In addition to the angle threshold, we must also consider the distance threshold. The distance s between a fixed point on the fitted lines (P i ) and either of the endpoints (P i−1 or P i+1 ) can be expressed as: The angular threshold can then be defined as: A value of D that is smaller than indicates the fitting line to be a continuous short line, representing a rapidly changing environment.

C. GLOBAL LOCALIZATION ALGORITHM CONSIDERING LONG AND STRAIGHT ENVIRONMENT
In the case of narrow environments, an improved global location can be acquired through the evaluation of environmental differences, using line fitting and CSM-based scanning matching.

Algorithm 3: Global Localization Algorithm for the Long Straight Environment
Input: maps, readings ς, d, t Output: In the previous section, angular and distance parameters were acquired and threshold values were set according to empirical results. For example, when the degree of environmental difference is small, environmental similarity is high. The global pose value acquired after CSM matching is then introduced into the odometer and internal sensor to produce a joint global pose in the laser frame for P i , This process can be described as: where p represents the influence of environmental similarity on positioning results. This is an empirical threshold, considered to be variable depending on the situation. The dependence of scan matching results on the dynamic selection of a robot posture updating strategy was investigated using environment information from the previous position. This global localization method considering environmental diversity is represented in Algorithm 3. This algorithm first develops a priori understanding of the environment through line extraction and then provides the corresponding global pose through CSM.

IV. VERIFICATION AND EXPERIMENTS
The proposed algorithm was evaluated through a series of experiments. The assessment metrics for global positioning errors are first introduced and the algorithm is then tested using a public data set. Finally, the model was tested in a physical environment using a mobile robot in a long straight corridor. The results from each experiment are then presented and discussed.

A. EVALUATION OF ERROR MEASUREMENT
In some instances, the accurate and inaccurate positioning results could be distinguished visually, through a mapping effect. However, as algorithm performance improved and the amount of data increased, direct observational evaluation became increasingly poor. The evaluation methodology used in this experiment can be described as follows. Root mean square error (RMSE) and relative pose error (RPE) are common evaluation metrics in SLAM applications [39]- [42].
RMSE measures the mean deviation between an estimated value and the correct value. In 2D laser-based SLAM problems, two different RMSE measurements are needed to represent errors in both translation and rotation. Although RMSE is not particularly useful for rotation errors, translation errors in narrow environments are far more prevalent. As such, RMSE was used as the primary evaluation metric for the algorithm proposed in this study. The errors in 2D translations can be defined as: The rotational component of RMSE is then given by: where x x,y i = (x i , y i ) T represents the estimated robot position andx x,y i indicates the actual robot position.

B. VALIDATION ON DATASETS
Two different standard datasets, listed in Table 1, were used to evaluate the proposed algorithm. The map created with each dataset is shown in Figure 7. It is evident from the figure that these sets include data from a large number of long and straight corridors. It is also clear that global localization based on scan matching is prone to inaccurate positioning, due to the high similarity of relevant characteristics.
As such, five regions constructed from two datasets (ACES-Building and MIT-Killian-Court) were selected from the map and compared with the parameters of lines fitted in each region. This type of line extraction was used to develop an initial description of the environment.
However, each region may contain multiple corresponding pose numbers. For this reason, the length and angle corresponding to a pose in the middle of the region were selected as representatives of the region (see Tables 2 and 3).
The angle in these tables was measured in the laser coordinate system, where small angles indicate segments are approximately parallel to the Y-axis, while angles approaching 90 • indicate the segments are nearly parallel to the X-axis. The length refers to the magnitude of the fitted line segment. As seen in Figure 7, these line segments were shorter, with an angle between 0 • and 90 • , in non-straight environments. Conversely, fitted line lengths were longer and the angles were closer to 0 • or 90 • in long straight environments. This suggests that line fitting plays a descriptive role in environment classification.
The ACES-Building and MIT-Killian-Court datasets were collected in real environments, making it difficult to acquire accurate absolute motion information. As such, the optimized global pose approximation (with loop optimization) was used as the real value [43]. Three different pose trajectories are given first in Fig.8 and Fig.9. The mean square error of this global pose, with and without consideration of the straight environment, was then calculated.
It can be seen from the pose trajectory map that the pose without considering the long and straight characteristics of the environment exhibit high noise. In addition, the pose trajectory considering environmental factors coincides nicely with the assumed true value, which also indicates that environmental factors have a positive impact on positioning results.
Both ACES-Building and MIT-Killian-Court dataset contain a large number of long and straight environment features, and MIT-Killian-Court dataset scenario is much larger than ACES-Building dataset scenario. Although it is not in perfect agreement with the results of laser tracking, the pose trajectory produced by considering the characteristics of the long straight corridor is closer to the true value than the trajectory produced without environmental considerations. This was particularly true for large scenes, as is evident in Fig. 9. Figures 10 and 11 respectively show the translation and rotation absolute pose errors produced by considering the  influence of narrow environment characteristics for the ACES-Building dataset. The abscissa represents the number of positions, while the ordinate is an indication of the degree of error.
Loop optimization requires the loop to be checked before correcting the pose. However, no loop was detected prior to an X-axis value of 880. In other words, the approximate true value of the pose was the same as the value without optimization. As such, the decision of whether to consider the influence of long straight environments in global pose estimation is negligible.
However, it does not affect the evaluation of the experiment, after the first loop point (x = 880), it is easy to acquire a global pose that considers environmental influences, which offer more accurate positioning results. It can be seen from the figure that the method considering the characteristics of environment length and straightness achieved superior global positioning performance as measured by both translation and rotation absolute pose error.
The MIT-Killian-Court dataset is composed of infinite corridors, which are especially suitable for testing the proposed model. the differences between rotation and translation errors for models that do and do not consider the influence of narrow environments are easy to distinguish. As with the ACES-Building dataset, images before position number 1800 were omitted. As shown in Figs. 12 and 13, translation and rotation errors were particularly obvious in the stage of the experimental process.
In summary, the technique proposed in this article improved positioning accuracy in long straight environments. Translation errors that considered the influence of these environments were smaller than those that did not consider the environment. This approach also led to the suppression of rotation errors. A statistical analysis of the translation and rotation errors for the two datasets was determined using: where P RSME i and the Q RSME i represent represent the error values for models considering and not considering the influence of environmental characteristics, respectively.
As shown in Table 4, the average translation and rotation errors decreased by 27.29% and 25.82% respectively. These results indicate that considering the characteristics of long and straight environments leads to improved reliability. As discussed above, ICP and RANSAC are two commonly used methods for scan matching and line fitting, respectively. In this study, these methods were combined and used for long straight corridor positioning. The combined algorithm was tested using the ACES-Building and MIT-Killian-Court datasets, as shown in Figure 14. Figures 14a and 14b show output results for each dataset, produced by combining CSM matching and weighted line fitting. Figures 14c and 14d show output results from ICP and RANSAC. The blue and green lines in the figure indicate pose trajectory outputs. The ICP algorithm alone is not suitable for large-scale scenarios, the results for these two datasets are particularly fragmented. In addition, the RANSAC algorithm is a non-deterministic model that produces reasonable results with certain probabilities, which are directly proportional to the number of iterations. And the RANSAC can only provide estimates in variable 2D environments. As such, judgment results in narrow corridors that are often unsatisfactory.
It is evident from the thumbnail images that the fusion algorithm based on CSM is more robust than that of ICP. The experimental data shown in Table 5 includes the average output time for a single pose, which refers to the ratio between the output time of all poses and the total number of poses. The average matching rate is the ratio between the sum of the matching rate of a single pose and the total number of poses. In small datasets, such as ACES-Building, the average output time for single poses is relatively small, suggesting the fusion algorithm based on ICP offers lower complexity. However, the result is not ideal as the match rate is too low, resulting in lower accuracy.
A different effect was observed in larger datasets, such as MIT-Killian-Court. In the experiment, we found the two fusion methods required additional iterations to output a pose when the dataset was in the final stage. However, the fusion method based on CSM offered a higher matching rate, especially for large turning angles.

C. TESTING IN REAL CORRIDORS
To evaluate the performance of our proposed algorithm, a series of physical experiments were conducted. The VOLUME 8, 2020  effectiveness of the proposed algorithm in real environments was assessed using a mobile robot in a long straight corridor.
A real scenario using a long straight corridor was selected to test the proposed algorithm. The hardware system consisted of a laptop with an i7 processor and a TurtleBot2 mobile robot, equipped with SICK LMS111 lidar. This configuration is shown in Figure 15. A corridor with a length of 70 meters meters is shown in Figure 16. The hallway only contains doors, walls, and some fire-fighting equipment, which was suitable for testing the proposed algorithm.
The proposed global positioning algorithm was implemented using a TurtleBot2. The laser sensor was mounted on the mobile robot and the dataset from the corridor was acquired using manual operation. Figure 16 shows the pose trajectory and constructed map for the mobile robot in the actual environment.
Consistency was ensured by first collecting scanning information and then using global localization algorithms, both with and without long straight parameter adjustments, to process the data. The resulting translation errors are shown in Figures 17 and 18.
Test results in the actual scene indicate that considering the influence of long straight environments results in lower translation and rotation errors. It can also be seen that the algorithm plays a positive role in the global localization of the narrow corridor.

V. CONCLUSION
This study proposed a novel global localization method for mobile robots in long and straight indoor environments, in which the global pose was adjusted for the first time by combining a weighted line fitting model with correlation scan matching. The prior environment was determined using a line extraction method. The linear fitting model was used to determine environmental characteristics and, in combination with correlation scan matching, output global posture weights.
Scan data in the narrow corridor were combined with internal sensor positions during correlation scan matching to update weight values. A validation experiment, conducted using public datasets and a physical environment, produced relatively low errors, thereby verifying the effectiveness of the proposed model.