Forward monitoring system and method

Abstract

To provide a forward monitoring system capable of recognizing a track stably even if a point group corresponding to a track acquired by a sensor is incomplete.SOLUTION: A forward monitoring system includes: a laying face point group removing part for removing a laying face point group from sensor output; a first narrowing-down part for extracting a first track candidate object in which a computer compares a cross sectional shape of an object, which is information included in a residue of the removal, with a track cross section stored in a memory; a second narrowing-down part for detecting a second track candidate object by comparing a width of the first track candidate object with a width of the track cross section; and a track curve complementary part for recognizing a track by complementing a track object including the second track candidate object. The sensor output includes: a point group corresponding to the track; and a point group corresponding to a laying face of the track. The laying face point group removing part extracts information about the object from the sensor output. The track curve complementary part recognizes the extracted object as a track and supports determination of the presence / absence of an obstacle in the vicinity thereof.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a forward monitoring system and method. [Background technology]

[0002] In technology for monitoring the area ahead of a traveling vehicle, it is necessary to estimate the vehicle's travel area in order to detect and respond to objects that may hinder the vehicle's travel. The vehicle's travel area can be defined as a solid object along the track. For such travel area estimation, sensor-based rail detection methods have been proposed in Patent Documents 1 to 3.

[0003] Patent Document 1 discloses a method for estimating rail positions by matching camera image data as sensor output with pre-prepared template image data of rails or reference distribution data of rail brightness values.

[0004] Patent Document 2 discloses a method of using distance measurement data obtained from a distance measurement sensor capable of measuring distances over a predetermined range in the track width direction, processing the distance measurement data to detect edges of the track in the track width direction, and estimating the track based on the detected edge positions.

[0005] Patent Document 3 discloses a method for determining rail positions by matching a point cloud extracted by determining a search range based on route information and track gauge acquired by a position estimation means with a point cloud of a rail template, and also discloses a method for determining rail positions by correctly matching by selecting an appropriate template from multiple types of templates. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 2020-006788 [Patent Document 2] Japanese Patent Publication No. 2020-155047 [Patent Document 3] Japanese Patent Application Publication No. 2016-212031 Summary of the Invention [Problem to be solved by the invention]

[0007] The rail detection using camera images in Patent Document 1 inevitably suffers from a drop in performance at night. Also, the distance measurement sensor formed by the laser radar in Patent Document 2 has difficulty accurately identifying the track when three or more objects are detected, including things like ground coils that are at the same height above the ground as the track.

[0008] In the method of Patent Document 3, even if a 3D laser radar is used, rail detection may fail if the self-position is indefinite or inaccurate. Furthermore, because the surface of a rail has a high specular reflectivity, the laser irradiated onto the rail top surface is specularly reflected and does not return to the sensor, resulting in missing points from the rail top surface, making accurate matching difficult. Furthermore, even if a template with an incomplete shape is selected for matching, if it matches an object other than a rail, it will be difficult to determine the rail position.

[0009] The present invention has been made in consideration of the above-mentioned problems, and its purpose is to provide a forward monitoring system that can stably recognize a trajectory even when the point cloud corresponding to the trajectory acquired by the sensor is incomplete. [Means for solving the problem]

[0010] The present invention, which solves the above problem, is a forward monitoring system equipped with a sensor that captures a trajectory that guides a moving body and a computer that processes the sensor output, and includes a laying surface point cloud removal unit that removes a laying surface point cloud from the sensor output, a first narrowing down unit that extracts a first trajectory candidate object by comparing the cross-sectional shape of the object, which is the information contained in the remaining data after the removal, with trajectory cross sections stored in the computer's memory, a second narrowing down unit that detects a second trajectory candidate object by comparing the width of the first trajectory candidate object with the width of the trajectory cross section, and a trajectory curve completion unit that recognizes a trajectory by complementing trajectory objects including the second trajectory candidate object, wherein the sensor output includes a point cloud corresponding to the trajectory and a point cloud corresponding to the laying surface of the trajectory, and the laying surface point cloud removal unit extracts object information from the sensor output, and the trajectory curve completion unit recognizes objects protruding from the laying surface as a trajectory based on the extracted object information, and outputs information to assist in determining whether or not there is a traveling obstacle near the recognized trajectory. [Effects of the Invention]

[0011] According to the present invention, a forward monitoring system can be provided that can stably recognize a track even when the point cloud corresponding to the track acquired by a sensor is incomplete. This can output highly accurate rail detection information to reduce erroneous or non-identified obstacle detection, thereby supporting improved safety in vehicle operation. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a perspective view showing a typical example of obstacle detection by a forward monitoring system according to a first embodiment of the present invention (hereinafter also referred to as "this system"). [Figure 2] 3 is a flowchart showing the processing procedure of the system according to the first embodiment. [Figure 3] FIG. 2 is an explanatory diagram of a coordinate system applied to the present system. [Figure 4] FIG. 2 is a plan view for explaining orbit center estimation in the system according to the first embodiment. [Figure 5] FIG. 2 is a front cross-sectional view for explaining an object extraction method in the system of the first embodiment. [Figure 6] FIG. 10 is a front cross-sectional view for explaining the narrowing process based on the rail cross-sectional dimensions in the system of the first embodiment. [Figure 7] FIG. 10 is a front cross-sectional view for explaining the narrowing process based on the rail-to-rail dimension in the system of the first embodiment. [Figure 8] 10 is a flowchart showing the processing procedure etc. of the present system (embodiments other than the first embodiment are also referred to as "the present system") according to the second embodiment of the present invention. [Figure 9] FIG. 10 is a plan view for explaining the narrowing-down process using a predicted trajectory line in the system according to the second embodiment. [Figure 10] FIG. 10 is a plan view for explaining the narrowing-down process using a predicted trajectory line corresponding to a curve and a branch in the system according to the second embodiment. [Figure 11] FIG. 11 is a front cross-sectional view for explaining the narrowing process based on the rail cross-sectional dimensions in the system of the third embodiment. [Figure 12] FIG. 11 is a plan view for explaining the narrowing-down process using a predicted trajectory line in the system according to the fourth embodiment. [Figure 13] FIG. 10 is a front cross-sectional view for explaining an object extraction method in the system of the fifth embodiment. [Figure 14] FIG. 20 is a front cross-sectional view for explaining an object extraction method using the system of the sixth embodiment. [Figure 15] FIG. 13 is a plan view for explaining a branch determination method in the system according to the seventh embodiment. [Figure 16] FIG. 13 is a plan view for explaining a method of installing sensors in the system of the eighth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments 1 to 8 of the present invention will be described with reference to the drawings. Note that the present invention is not limited to these embodiments. Below, stable rail (track) detection by this system will be described using a track basically composed of two rails as a representative example. This system can also be applied to new transportation systems and monorails with track configurations different from this representative example. Note that this system is applied not only to trains that operate between railway stations, but also to vehicles that move within station premises and vehicle depots, and all of these application targets are referred to as vehicles here. [Example]

[0014] Fig. 1 is a perspective view showing a typical example of obstacle detection by this system 10. This system 10 supports the safe operation of a moving body (hereinafter also referred to as a "vehicle") 20 that travels while being guided along a track 40 as shown in the obstacle detection example of Fig. 1. This system 10 is composed of a forward monitoring sensor 101, an object detection unit 103, a rail detection unit 104, and an obstacle determination unit 105, which are mounted on the vehicle 20 as on-board devices, but components other than the sensor 101 may also be ground equipment (not shown).

[0015] Each functional unit of this system 10 is a processing function formed by a computer executing a program stored in memory, and some are not shown because they are not visually distinguishable. Therefore, the term "unit" is used for convenience of explanation, and alternative terms such as "processing" can be interpreted as having substantially the same meaning. The computer can be a single-chip microcomputer, but a personal computer or the like can also be used. In addition, by utilizing communication functions, the computer can also serve as part of the computers installed in the traffic control unit or other ground equipment.

[0016] The system 10 uses a computer to process data acquired by a forward monitoring sensor 101 to determine an obstacle. That is, the system 10 and a forward monitoring method using the same (hereinafter also referred to as "the method") employ a laser radar as the forward monitoring sensor 101, extracts and compiles a cloud of points protruding above the ground from the cloud of points acquired by the system, and estimates that an object that matches the external shape characteristics (hereinafter also referred to as "dimensional characteristics") of a rail 40 is a rail 40.

[0017] The object detection unit 103 estimates the position of an object ahead of the vehicle, and the rail detection unit 104 estimates the position of the rail 40 ahead of the vehicle. Next, the obstacle determination unit 105 estimates the vehicle travel area 50 based on the rail position estimation result, and determines whether the object position estimation result is within the vehicle travel area 50.

[0018] A forward monitoring sensor (hereinafter also referred to as "sensor" or "laser radar") 101 is installed on the vehicle 20 and acquires point cloud information of the forward direction. An object detection unit 103 estimates the position of an object in front of the vehicle 20 based on the sensor output of the sensor 101. The object detection unit 103 may detect an object using data from at least one of the sensor 101 and a second forward monitoring sensor 102 having a different property.

[0019] The second forward monitoring sensor 102 has the same purpose as the sensor 101, but to provide redundancy, it is a sensor 102 with different properties, such as a camera or millimeter-wave radar. The rail detection unit 104 has a function of receiving point cloud information from the sensor 101 and detecting the rails 40 ahead of the vehicle.

[0020] The obstacle determination unit 105 has a function of determining whether or not a detected object is an obstacle based on the results of the object detection unit 103 and the rail detection unit 104. More specifically, if the object is located within the vehicle travel area 50, like object 301 in Fig. 1, it will interfere with vehicle travel, and so the obstacle determination unit 105 determines that it is an obstacle. Conversely, if the object is located outside the vehicle travel area 50, like object 302 in Fig. 1, it will not interfere with vehicle travel, and so the obstacle determination unit 105 determines that it is not an obstacle.

[0021] The vehicle travel area 50 is a two-dimensional or three-dimensional area in which the vehicle 20 will travel in the future. An object 301 present in the vehicle travel area 50 is deemed to pose a risk of collision with the vehicle 20 and is determined to be an obstacle. One example of a method for calculating the vehicle travel area 50 is to use the rail detection results to determine the midline between two rails 40 as a track axis 501, and determine the area from the track axis 501 to an area a distance 502 away as the vehicle travel area 50. The distance 502 may be calculated based on legal standards such as a construction gauge or a vehicle gauge, for example.

[0022] Next, we will explain the operation of the system 10. Figure 2 is a flowchart showing the processing procedures of the system 10 of Figure 1, and shows point cloud data D101, point cloud coordinate data acquisition processing S101, track center point estimation processing S102, intra-section point cloud extraction processing S103, object extraction processing S104 with ground point cloud removal processing, narrowing-down processing S105 (first narrowing-down) based on rail cross-sectional dimensions, narrowing-down processing S106 (second narrowing-down) based on rail spacing dimensions, determination S107 of whether or not predetermined point cloud processing has been completed, rail curve interpolation processing S108, and rail estimated curve D102.

[0023] Hereinafter, each of these processes will be indicated by only the symbol S. In Fig. 2, S101 to S108 in particular clearly indicate the operation of the rail detection unit 104. If S107 is Yes, the process proceeds to S108, and if No, the process returns to S102. First, in S101, the rail detection unit 104 acquires point cloud data D101 input from the sensor 101, and converts it into a vehicle-fixed three-dimensional xyz coordinate system Σ0 (hereinafter abbreviated as "coordinate system Σ0") defined in Fig. 3.

[0024] 3 is an explanatory diagram for the definition of the coordinate system Σ0 applied to the present system 10. The coordinate system Σ0 is a right-handed coordinate system having an x-axis in the direction along the rail 40, a y-axis in the direction of the sleepers, and a z-axis in the direction of the zenith. Furthermore, the origin O is defined such that the x-origin is the position of the front of the vehicle, the y-origin is the midpoint of the front of the vehicle in the y direction, and the z-origin is the position of the bottom surface in the z direction of the vehicle underframe (not shown).

[0025] In the subsequent steps S102 to S106, the system 10 defines an inspection section based on the estimated track center, detects the rails 40 using the point cloud within the inspection section, and estimates the track center point of the next inspection section using the rail detection results.By repeating this operation from near the vehicle to far away, the system 10 detects all of the rails 40 within that range.The inspection section is an area within a certain distance starting from the estimated track center point, and is defined at approximately equal intervals along the rails 40 or the x-axis.

[0026] Fig. 4 is a plan view for explaining the track center point estimation process S102 by the system 10 of the first embodiment. The method of calculating the estimated track center and the inspection section in S102 differs depending on whether the vehicle 20 is closest to the vehicle 20, as shown in Fig. 4 (G20), or not, as shown in Fig. 4 (G21).

[0027] First, in the case of the vehicle closest to the rail 40 (G20), the y origin in the coordinate system Σ0 is often between the two rails 40. Furthermore, the cause of fluctuations in the z-direction position of the rail 40 relative to the vehicle 20 is the vertical movement of the vehicle 20 caused by a suspension device such as an air spring. Therefore, the relative position in the z direction between the vehicle 20 and the rail 40 can be considered to be roughly constant.

[0028] Therefore, the system 10 defines a certain range starting from A0 by setting the y coordinate of the estimated track center to y = 0, the z coordinate to the z position of the rail top surface when the vehicle 20 is stationary, and the x position to x = 0. In defining coordinates in this way, the system 10 can acquire (hereinafter simply referred to as "acquire") the coordinates of a point cloud including the left and right rails 40.

[0029] The definition of the inspection section in the case of the area closest to the vehicle can be an area that includes A0 as the starting point, a position that is a distance Dy0 away in the positive and negative y direction, and a position that is a distance Dx0 away in the positive x direction, as shown in Figure 4 (G20).

[0030] Next, using Figure 4 (G21), we will explain how to calculate the estimated track center Ai at a position further back than the nearest position. As shown in Figure 4 (G21), the estimated track center Ai is the midpoint of the left and right rail detection results G211. This estimated track center Ai is closer to the vehicle 20 than the inspection section Di, and is the midpoint of the left and right rail detection results G211 for the inspection section Di-1, which is one inspection section before the inspection section Di.

[0031] The left and right rail detection result G211 may be a set of representative points of the left and right rail detection results in the inspection section Di-1, or may be a set of the farthest point of the left rail 40 and the farthest point of the right rail 40 in the point cloud corresponding to the representative points of the left and right rail detection results.

[0032] Regarding the method of calculating the inspection section Di, as shown in FIG. 4 (G21), a coordinate system Σi corresponding to the inspection section Di is defined, and the estimated trajectory center Ai is set as the starting point, and the area is defined as including a distance Dyi in the positive and negative directions of yi and a distance Dxi in the positive direction of xi from the starting point Ai.

[0033] The orientation of the coordinate system Σi may be parallel to that of the coordinate system Σ0. Alternatively, it may be rotated with respect to the coordinate system Σ0. When rotating, an origin Ai of the predicted track axis line yi = ci is created based on the rail detection results in the inspection sections D0, D1, ..., Di-1 (see Figure 16). The tangent direction to the curved rail detected at the origin Ai becomes the xi axis.

[0034] When the rail 40 is curved, the yi axis of the coordinate system rotated in this way is closer to being parallel to the sleeper direction of the rail 40 than the y axis of the coordinate system Σ0. Since the gauge is roughly equal to the distance between the rails in the sleeper direction, when narrowing down the rail candidates using the gauge in the subsequent processing, it is expected to be more accurate to use the yi direction distance of the rail candidates narrowed down in the previous processing rather than the y direction distance of multiple rail candidates. This completes the operation of S102.

[0035] Next, in S103, the point cloud present in the inspection section calculated in S102 is extracted as an in-section point cloud including the rail 40. Furthermore, in S104, the point cloud corresponding to the ground is removed from the in-section point cloud, and multiple objects Coi (FIG. 5) protruding from the ground are detected.

[0036] 5 is a front cross-sectional view illustrating an object extraction method in the system 10 according to the first embodiment, and is a diagram illustrating a method for removing the ground point cloud in S104 of FIG. 2 and a method for detecting multiple objects Coi protruding from the ground. As shown in FIG. 5, first, in S104, the point cloud within the section indicated by ● and ◯ is projected onto a two-dimensional plane consisting of the yi and zi axes of the Σi coordinate system. Next, in S104, a line or curve lgi connecting the ◯ corresponding to the ground is identified.

[0037] 5, the system 10 further removes from the point cloud within the section points those points whose zi distance to the line or curve lgi is equal to or less than a threshold value Δzg and those points whose zi coordinates are smaller than those of the line or curve lgi. This allows the system 10 to remove the points indicated by circles that correspond to the ground and extract the points Cpi that protrude above the ground, indicated by circles.

[0038] The system 10 may detect a curve lgi or the like by using a curve calculation method that removes the point cloud Cpi that protrudes from the ground as an outlier, focusing on the fact that the number of points corresponding to the ground tends to be greater than the number of points protruding from the ground. Alternatively, the system 10 may detect a curve lgi or the like by using a simple regression line that is an extension of a known number of points.

[0039] Next, a method for calculating the plurality of objects Coi protruding from the ground surface will be described. The system 10 calculates the plurality of objects Coi protruding from the ground surface by grouping the point clouds Cpi protruding from the ground surface based on the distance between the points, and assembling them into a plurality of object units. For example, one method is to group points in the point cloud Cpi whose inter-point distance is equal to or less than a predetermined threshold D, by assuming that they constitute the same object. This completes the operation of S104.

[0040] Next, of the first narrowing-down process (S105, S106) of the present system 10, S105 shown in Fig. 6 will be described. Fig. 6 is a front cross-sectional view for explaining the narrowing-down process (S105) based on rail cross-sectional dimensions. In S105, an object having the dimensional characteristics of the rail cross section is extracted from the multiple objects Coi, and a first rail candidate object Co1i is obtained.

[0041] 2, 8 (S105), and 6, the system 10 performs a first narrowing-down process (S105) using a first incomplete template T1 (FIG. 6) that simulates the dimensional characteristics of a rail cross section. That is, the system 10 considers that, among the multiple objects Coi, an object that falls within the range of the first incomplete template T1 has the dimensional characteristics of a rail cross section, and extracts a first rail candidate object Co1i.

[0042] The first incomplete template T1 is a figure created based on the dimensions of the rail cross section, e.g., a rectangle with margins added to the maximum vertical and horizontal dimensions of the rail cross section. The midpoint of the bottom edge of the first incomplete template T1 is set to be positioned on a line or curve lgi corresponding to the ground. With this setting, the system 10 extracts a first rail candidate object Co1i with dimensions similar to the rail cross section. The system 10 then processes the sensor output using the first incomplete template T1. The first incomplete template T1 has the effect of removing noise from the point cloud that is floating above the ground.

[0043] Furthermore, the present system 10 can recognize rail point clouds that are missing enough to complete the rail cross-sectional contour, by considering them as being within the range of the first incomplete template T1. Therefore, according to the present system 10, even point clouds that are missing the top surface rail portion, which was an issue in Patent Document 3, can be stably recognized as rails 40 by leaving them as rail candidates. This completes the operation of S105. Next, by comparing the distance between the first rail candidate objects Co1i with the gauge, a narrowing-down process based on the rail spacing is performed (S106) to extract second rail candidate objects Co2i (FIG. 9).

[0044] FIG. 7 is a front cross-sectional view for explaining the narrowing-down process (S106 in FIGS. 2 and 8, which corresponds to the second narrowing-down process in the present invention) based on the gauge dimension of the rail 40 by the system 10 of the first embodiment. The system 1 compares whether the measured distances between a plurality of first rail candidate objects Co1i match a known gauge. First, since the gauge is the distance between the inside of the left rail 40 and the inside of the right rail 40, it is necessary to calculate the position of the first rail candidate object Co1i that is considered to be on the inside of the rail 40.

[0045] Therefore, as shown in Figure 7, the system 10 creates a track center line yi = Ci that passes through the estimated track center Ai and is parallel to the zi axis, and regards objects with large yi coordinates on the track center line yi = Ci as left rail objects, and objects with small yi coordinates as right rail objects. In other words, the system 10 regards first rail candidate objects Co1i whose yi coordinates are all larger than Ci as objects Co1i(l) that are estimated to be the left rail 40, and regards first rail candidate objects Co1i whose yi coordinates are all smaller than Ci as objects Co1i(r) that are estimated to be the right rail 40.

[0046] The system 10 excludes any object Co1i whose yi coordinate straddles Ci, determining that it cannot be the rail 40. Next, the system 10 can determine candidate points on the inside of the left and right rails 40 by taking the minimum value in the yi direction of the object Co1i(l) estimated to be the left rail 40 as the representative point of object Co1i(l)A, and taking the maximum value in the yi direction of the object Co1i(r) estimated to be the right rail 40 as the representative point of object Co1i(r).

[0047] Finally, the system 10 extracts pairs of all objects Co1i(l) and Co1i(r) whose representative point distances are within a distance range created based on the track gauge, and designates these pairs as second rail candidate objects Co2i (FIG. 9). The second rail candidate objects Co2i form a pair of right rail object Co12i(r) and left rail object Co12i(l).

[0048] Here, a method for extracting the right rail object Co12i(r) and the left rail object Co12i(l) will be described using Figure 7. In the example of Figure 7, there are two left rail objects, Co1i(l)A and Co1i(l)B, and one right rail object, Co1i(r).

[0049] The distance between the representative point Cr1i(l)A of the left-side object Co1i(l)A and the representative point Cr1i(r) of the right-side object Co1i(r) is within the recognition range that is equal to the track gauge. Also, the distance between the representative point Cr1i(l)B of the object Co1i(l)B, which is to the left of the middle, and the representative point Cr1i(r) of the object Co1i(r) is too close and falls outside the recognition range. In the case illustrated in Figure 7, the system 10 recognizes the object Co1i(l)A as the left-rail object Co12i(l) and the object Co1i(r) as the right-rail object Co12i(r).

[0050] By performing the second narrowing-down process based on the gauge, the system 10 can further remove, from the first rail candidate objects Co1i extracted by the first narrowing-down process, objects that have the same length and width dimensions as the rail cross section and height from the ground, but are not rails 40, thereby improving the success rate of rail detection and enabling more stable rail detection.

[0051] Next, in S107, it is determined whether processing has been completed for the point cloud to be processed. If processing has not been completed, the process returns to S102, and a second rail candidate object Co2i+1(r) is calculated for the next inspection section Di+1. If processing has been completed, the process proceeds to S108, where a rail detection result is calculated. In S108, all second rail candidate objects Co2i, Co120, Co121..., Co12n calculated for each inspection section D0, D1,..., Dn (see FIG. 16) are connected by complementation, and the rail detection result is output.

[0052] The second rail candidate object Co2i (FIG. 9) is complemented for each of the left and right rails 40. That is, the system 10 extracts the right rail curve by finding a complement curve using all right rail objects Co20(r), Co21(r), ..., Co2n(r) in each inspection section D0, D1, ..., Dn.

[0053] Similarly, the system 10 extracts the left rail curve by calculating an interpolated curve using all left rail objects Co20(l), Co21(l), ..., Co2n(l) in each inspection section. The right rail curve and the left rail curve are integrated and output as a rail estimation curve D102 (Fig. 2), which is the rail detection result. [Example]

[0054] In Example 2, a rail detection method with improved performance achieved by adding a narrowing-down section using track prediction lines to the configuration of Example 1 will be described with reference to Figures 8 to 10. The system 10 in Figure 8 improves noise resistance in rail detection performance and can also detect left and right rail pairs at junctions.

[0055] Fig. 8 is a flowchart showing the narrowing down process using the trajectory prediction line in the present system 10, particularly the operation of the rail detection unit 104. In the second embodiment (Fig. 8), a narrowing down process S109 using the trajectory prediction line (referred to as the third narrowing down process in the present invention) is added between S106 and S107 in the first embodiment (Fig. 2). In this S109, the second rail candidate object Co2i output by the narrowing down unit based on the rail spacing dimension is further narrowed down and a branch determination is made using the trajectory prediction line, and a third rail candidate object Co3i is output.

[0056] The predicted track line is a straight line or curve that represents the estimated position of the rail 40 in the inspection section Di, calculated using the rail detection results of the inspection sections D0, D1, . . . , Di-1 that precede the current inspection section Di.

[0057] 9 is a plan view for explaining the narrowing-down process S109 using a trajectory prediction line in the system 10 according to the second embodiment. Fig. 9 illustrates a method for creating a trajectory prediction line and the narrowing-down process S109 of the second rail candidate objects Co2i using the trajectory prediction line. First, the method for calculating the trajectory prediction line is as follows.

[0058] That is, the present system 10 complements the left rail objects Co30(l), Co31(l), ..., Co3i-1(l) included in the third rail candidate objects Co30, Co31, ..., Co3i-1 calculated in the sections D0, D1, ..., Di-1 before the current inspection section Di.

[0059] The system 10 also calculates an extrapolated line Sple(l) of the complementary line Spl(l) of the left rail 40. The system 10 also creates, as a predicted trajectory line, an extrapolated line Sple(r) of the complementary line Spl(r) of the right rail 40 calculated by complementing the right rail objects Co30(r), Co31(r), ..., Co3i-1(r).

[0060] In Example 2, the system 10 also uses third rail candidate objects in sections D0, D1, ..., Di-1 before and after the current inspection section Di to calculate the predicted trajectory line, but it may also use only the section Di-L, ..., Di-1 at point L (not shown) before the inspection section Di.

[0061] Next, the process of narrowing down the second rail candidate objects Co2i will be explained. In the inspection section Di, second rail candidate objects Co2i whose distance from the extrapolated lines Sple(l), Sple(r) is within the threshold Δyr are extracted as third rail candidate objects Co3i. At this time, as in Example 2, if the number of extracted objects is one left rail object Co3i(l) and one right rail object Co3i(r), it is determined that no branching has occurred.

[0062] Fig. 10 is a plan view for explaining the narrowing-down process using a trajectory prediction line corresponding to a curve and a branch in the system 10 of the second embodiment. On the other hand, if the number of extracted objects is two (G31) for the left rail objects Co3i(l) and two (G32) for the right rail objects Co3i(r) as shown in Fig. 10, the system 10 determines that a branch has occurred. Furthermore, the system 10 distinguishes between the left rail pair P3i(p=l) and the right rail pair P3i(p=r) for the third rail candidate objects Co3i.

[0063] In Example 2, in the inspection section Di, of the third rail candidate objects Co3i, the left rail object belonging to the left rail pair P3i(p=l) is called Co3i(l, p=l), and the right rail object is called Co3i(r, p=l). Also, in the inspection section Di, the left rail object belonging to the right rail pair P3i(p=r) is called Co3i(l, p=r), and the right rail object is called Co3i(r, p=r). Furthermore, in sections where no branching occurs, the notation p=r, p=l is omitted, and the left rail object is called Co3i(l), and the right rail object is called Co3i(r).

[0064] When a branch determination is made in inspection section Di, rail detection is performed in the next inspection section Di+1, which is divided into an inspection section Di+1(p=l) for estimating the left rail pair and an inspection section Di+1(p=r) for estimating the right rail pair. Specifically, in the inspection section Di+1(p=l) for the left rail pair, the track center point in S102 is determined based on the left rail candidate object Co3i(l,p=l) of the left rail pair and the right rail candidate object Co3i(r,p=l) of the left rail pair.

[0065] Furthermore, in S109, a predicted trajectory line is calculated using the inspection sections D0, D1, ..., Di and the third rail candidate object Co3i detected in the left rail pair inspection section Di (p = l). The trajectory center point of S102 in the right rail pair inspection section Di+1 (p = r) is calculated based on the left rail candidate object Co3i (l, p = r) of the right rail pair and the right rail candidate object Co3i (r, p = r) of the right rail pair (S102).

[0066] Furthermore, in S109, a predicted trajectory line is calculated using the third rail candidate objects Co3i detected in the inspection sections D0, D1, . . . , Di and the right rail pair inspection section Di (p=r).

[0067] In the second embodiment, the rail curve interpolation unit does not interpolate the second rail candidate objects Co20, Co21, ..., Co2i, Co2n. The rail curve interpolation unit interpolates the curves of the left rail 40 and the right rail 40 for the third rail candidate objects Co30, Co21, ..., Co2n calculated in each inspection section, and outputs the rail estimation curves as the rail detection results (S107).

[0068] In S109, if it is determined that a branch exists in the inspection section Di, curve interpolation processing is performed for each of the following first to fourth targets, and a rail estimation curve is output as a rail detection result.

[0069] The first objects are the left rail candidate objects Co30(l), Co31(l), ..., Co3b-1(l), Co3b(l, p=l), ..., Co3n(l, p=l) of the left rail pair.

[0070] The second target is the right rail candidate objects Co30(r), Co31(r), ..., Co3b-1(r), Co3b(r, p=l), ..., Co3n(r, p=l).

[0071] The third object is the left rail candidate objects Co30(l), Co31(l), ..., Co3b-1(l), Co3b(l,p=r), ..., Co3n(l,p=r) of the right rail pair.

[0072] The fourth object is the right rail candidate objects Co30(r), Co31(r), ..., Co3b-1(r), Co3b(r, p=r), ..., Co3n(r, p=r).

[0073] By adding a narrowing-down process using track prediction lines, it is possible to exclude object pairs other than the rail 40 that has a width similar to the track gauge and is included in the second rail candidate object Co2i, thereby further improving the success rate of rail detection and enabling more stable rail detection.

[0074] In addition, by adding the rail 40 branching determination process, rail detection can be performed even when a branch occurs in the rail 40, enabling stable rail detection. Furthermore, since detection is performed separately for the left rail pair and the right rail pair, when the branching direction is given by other means, it is also possible to detect only the rail 40 in the branching direction. [Example]

[0075] In Example 3, in S105 (the first narrowing down in the present invention) where the narrowing down process based on the rail cross-sectional dimensions of Example 1 or Example 2 is performed, the first rail candidate object Co1 is extracted using a second incomplete template that reproduces a part of the rail cross-sectional shape, rather than a first incomplete template that has a rectangular shape and is created based on the maximum length and width dimensions of the rail cross-section.

[0076] Fig. 11 is a front cross-sectional view for explaining the narrowing-down process based on rail cross-sectional dimensions in the system 10 of Example 3. That is, Fig. 11 shows the definition (G40) of the second incomplete template in Example 3, and a diagram (G41) of the second incomplete template and the point cloud in the inspection section Di projected onto the yizi plane.

[0077] The second incomplete template is created based on a partial shape of the rail cross section, assuming that there may be defects in the point cloud corresponding to the rail 40. For example, attention is paid to the property that a laser irradiated onto the side surface of the rail tends to return to the laser radar 101. Therefore, as shown in FIG. 11 (G40), point clouds Tp1 and Tp2 that reproduce only the side surface of the rail 40 constitute the second incomplete template. That is, in the system 10, the point cloud Tp1 is defined as the second incomplete template To2A, and the point cloud Tp2 is defined as the second incomplete template ToB.

[0078] 11 (G41), a rail object To2A is set based on the shape of the second incomplete template T2A, and a rail object To2B is set based on the shape of the second incomplete template To2B. In the third embodiment, the shapes of the rail objects To2A and To2B are set to rectangular shapes.

[0079] Next, we will explain how to extract first rail candidate objects Co1i using second incomplete templates T2A and T2B in the inspection section Di. First, the second incomplete template T2A is slid on the yizi plane, and the similarity to the point cloud constituting the multiple objects Coi at each yizi position is calculated, and multiple positions of the second incomplete template T2A where the similarity is equal to or greater than a predetermined threshold are extracted. Furthermore, multiple rail objects To2A corresponding to the multiple extracted second incomplete templates T2A are designated as first rail candidate objects Co1i.

[0080] The second incomplete template T2B is then slid on the yizi plane, the similarity between each yizi position and the point cloud constituting the multiple objects Coi is calculated, and multiple positions of the second incomplete template T2B where the similarity is equal to or greater than a predetermined threshold are extracted. Furthermore, multiple rail objects To2B corresponding to the multiple extracted second incomplete templates T2B are added to the first rail candidate objects Co1i.

[0081] The above is the operation of the narrowing-down process S105 based on rail cross-sectional dimensions in Example 3. According to Example 3, even if there is a defect in the point cloud corresponding to the rail 40, it is possible to extract multiple rail candidate positions without template matching failure, thereby enabling stable rail detection.

[0082] Furthermore, the system 10 of the third embodiment prevents rail detection failures that result in matching only to objects other than the rail 40 when there are multiple objects similar to the rail shape. The system 10 of the third embodiment first uses the second incomplete template to extract multiple rail candidates including a point cloud corresponding to the rail 40, thereby avoiding the phenomenon of matching only to other objects. The system 10 of the third embodiment can further narrow down the rail point cloud using rail shape characteristics such as the gauge in subsequent processing, thereby extracting the rail point cloud with higher accuracy. [Example]

[0083] In the fourth embodiment, attention is focused on the fact that the rail shape changes smoothly in the inspection section. In the fourth embodiment, which applies this, a trajectory prediction line and a branch determination method different from those in the second embodiment are described. FIG. 12 is a plan view for explaining the narrowing-down process using the trajectory prediction line in the system 10 of the fourth embodiment (the fourth narrowing-down process in the present invention). That is, (G50) and (G51) in FIG. 12 are diagrams showing a method in the system 10 for extracting third rail candidate objects Co3i using the trajectory prediction line within the inspection section Di. In FIG. 12, (G50) shows a case where no branch occurs within the inspection section Di, and (G51) shows a case where a branch occurs in the inspection section Di.

[0084] The system 10 of the fourth embodiment first calculates, in the yizi coordinate system, coordinates that are ΔD away from the estimated trajectory center point Ai of the inspection section Di in the positive and negative directions of yi, and defines the point ΔD away in the positive direction of yi as Sbi(l) and the point ΔD away in the negative direction of yi as Sbi(r). Next, the system 10 calculates parameters that best match the point group Cp2i that constitutes the second rail candidate object Co2i for a line lb(l) that passes through point Sbi(l) and a line lb(r) that passes through point Sbi(r) and is parallel to line lb(l).

[0085] Here, the system 10 calculates the sum of the distances between lb(l) and the point group Cp2i plus the sum of the distances between lb(r) and the point group Po2i as the objective function I. When no bifurcation occurs in the objective function I for the slope a of the line lb(l) as shown in FIG. 12 (G50), there is only one type of line lb(l) and line lb(r) that best matches the point group Cp2i, and therefore the objective function I is considered to have a single minimum value.

[0086] On the other hand, when a bifurcation occurs as in (G51), there are two lines lb(l) and lb(r) that closely match the point group Cp2i, and therefore the objective function I is considered to have two minimum values. Therefore, the system 10 finds the minimum value of the objective function I, and if there is one minimum value that is below a predetermined value, it determines that no bifurcation has occurred and calculates the slope aopt of the line lb(l) at which the minimum value occurs.

[0087] On the other hand, if there are two minimum values ​​below a predetermined value, it is determined that a branch has occurred, and the larger value of the slope a of the line lb(l) that results in the minimum value is defined as the slope aopt(p=l) of the line corresponding to the left rail pair, and the smaller value is defined as the slope aopt(p=r) of the line corresponding to the left rail pair.The method for extracting third rail candidate objects will be explained below.

[0088] First, when no branching occurs as in Figure 12 (G50), the point cloud within a predetermined distance from the straight line lb (l, a = aopt) having aopt is defined as the point cloud Cp3i(l) corresponding to the left rail object Co3i(l) of the third rail candidate object Co3i. Also, the point cloud within a predetermined distance from the straight line lb(r) having aopt is defined as the point cloud Cp3i(r) corresponding to the right rail object Co3i(r) of the third rail candidate object Co3i.

[0089] Next, the system 10 determines the left rail object Co3i(l) from the point cloud Cp3i(l), and also determines the right rail object Co3i(r) from the point cloud Cp3i(r). As a result, the system 10 detects the third rail object Co3i. As a more specific method for calculating Co3i(l) and Co3i(r), the system 10 may define the object group resulting from grouping the point cloud Cp3i(l) by distance as Co3i(l), and the object group resulting from grouping the point cloud Cp3i(r) by distance as Co3i(r). Alternatively, the point cloud Cp3i(l) itself may be defined as the object Co3i(l), and the point cloud Cp3i(r) itself may be defined as the object Co3i(r).

[0090] Furthermore, when a branch occurs as in (G51), for the left rail pair, the point cloud within a predetermined distance from the straight line lb(l,p=l,a=aopt) having aopt(p=l) is defined as the point cloud Cp3i(l,p=l) corresponding to the left rail object Co3i(l,p=l) of the third rail candidate object Co3i. Also, the point cloud within a predetermined distance from the straight line lb(r,p=l,a=aopt) having aopt(p=l) is defined as the point cloud Cp3i(r,p=l) corresponding to the right rail object Co3i(r,p=l) of the third rail candidate object Co3i.

[0091] Next, the system 10 finds the left rail object Co3i(l,p=l) from the point group Cp3i(l,p=l), and similarly finds the left rail object Co3i(r,p=l) from the point group Cp3i(r,p=l), thereby detecting the third rail candidate object Co3i(p=l) of the left rail pair.

[0092] For the right rail pair, the straight line lb(l,p=r) having aopt(p=r) and the point cloud within a predetermined distance are defined as the point cloud Cp3i(l,p=r) corresponding to the left rail object Co3i(l,p=r) in the third rail candidate object Co3i, and the straight line lb(r,p=r) having aopt(p=r) and the point cloud within a predetermined distance are defined as the point cloud Cp3i(r,p=r) corresponding to the right rail object Co3i(r,p=r) in the third rail candidate object Co3i.

[0093] Next, the system 10 detects the third rail candidate object Co3i(p=r) of the right rail pair by determining the right rail object Co3i(l,p=r) from the point group Cp3i(l,p=r), and similarly determining the right rail object Co3i(r,p=r) from the point group Cp3i(r,p=r).

[0094] The system 10 of the fourth embodiment focuses on the smooth change in rail shape. Therefore, the system 10 can prevent erroneous determination that a branch exists even when second rail candidate objects Co2i, which are not the rail 40 but are arranged at intervals similar to the width of the rail 40, are close to the rail 40. As a result, the system 10 can perform more stable rail detection.

[0095] Furthermore, in the present system 10 of the fourth embodiment, the straight lines lb(l) and lb(r) are used, but the present system 10 may also use curves. For example, for a quadratic curve cb(l) that passes through the point Sbi(l) and a quadratic curve cb(r) that passes through the point Sbi(r) and is parallel to the curve cb(l), the present system 10 may employ a method of calculating parameters that best match the point group Cp2i that constitutes the second rail candidate object Co2i. [Example]

[0096] In the fifth embodiment, a rail detection method will be described for cases where only the top surface of the rail is exposed and the rest is buried, such as at a tramway track surface or at a railroad crossing, as shown in Fig. 13. That is, the system 10 of the fifth embodiment adds processing in the ground point cloud removal S104, thereby enabling object detection even when the rail 40 does not protrude from the track surface.

[0097] Fig. 13 is a front cross-sectional view for explaining an object extraction method in the system 10 of Example 5. That is, Fig. 13 is a diagram showing a ground point cloud removal method in Example 5. The road surface of a road that intersects with tracks (rails) 40 at a railroad crossing or the like, or a road that also serves as a tram track 40 (hereinafter collectively referred to as "railroad crossing, etc.") roughly coincides with the top surface of the rails 40, so if ground detection is used to extract an object above the ground, nothing can be extracted.

[0098] To address this issue, we focus on the fact that the ground height at railroad crossings, etc. changes suddenly compared to before and after entering the crossing, and that when a railroad crossing, etc. is viewed in a yizi cross section, there is a point cloud on part of the side of the rail, and detect objects consisting of a point cloud including the rail 40.

[0099] First, Δzg is calculated by subtracting the ground height zi-1 calculated by ground estimation in the inspection section Di-1 from the ground height zi calculated by ground estimation in the inspection section Di. Δzg takes a value close to 0 unless the ground height changes suddenly, and Δzg >> 0 when the vehicle enters a railroad crossing or other similar area for the first time in the inspection section Di. Therefore, if Δzg exceeds the threshold, it is detected that the inspection section Di has entered a railroad crossing or other similar area, and the subsequent processing is carried out.

[0100] In the inspection section Di, the system 10 removes from the point cloud within the section the point cloud whose zi-direction distance from the line or curve lgi is equal to or less than a threshold value Δzg and the point cloud whose zi coordinate is greater than that of the line or curve lgi, thereby removing the point cloud corresponding to the ground and extracting the point cloud Cpi that protrudes downward from the ground. The method for calculating the multiple objects Coi using the point cloud Cpi is the same as the method described in Example 1. According to the system 10 of Example 5, rail detection can be performed even at a railroad crossing or the like where an intersecting road or the like is at the same height as the top surface of the rail 40. [Example]

[0101] In the sixth embodiment, a rail detection method different from that in the fifth embodiment will be described for the case where the rail 40 exists at a railroad crossing or within the track area of ​​a tram. In the sixth embodiment, attention is paid to the fact that grooves exist only at positions where the wheels pass at a railroad crossing or the like, and multiple objects Coi including the rail 40 are detected by detecting the grooves. First, the method for determining the railroad crossing or the like in the inspection section Di is the same as in the fifth embodiment.

[0102] Next, if the inspection section Di is determined to be a railroad crossing or the like, in Example 6, the point cloud is projected onto the yizi plane and then grouped, and if the y-direction distance between groups is within a specified range, it is determined to be a groove through which wheels pass.

[0103] Fig. 14 is a front cross-sectional view showing a method for detecting multiple objects Coi including a rail 40 by detecting grooves in Example 6. In the case of Fig. 14, the inter-group distance between groups CoiA and CoiB is w1, and the inter-group distance between groups CoiB and CoiC is w2. If these w1 and w2 (hereinafter, the same applies to only the symbols) are within a range of values ​​set with reference to the wheel width, it is determined that the groove is one through which the wheel will pass.

[0104] After determining that w1 and w2 are grooves through which the wheels pass, the system 10 determines whether the groove is along the left or right rail 40 using the yi coordinate Ci of the estimated track center Ai. If the yi coordinate of the groove is greater than Ci, the system 10 determines that it corresponds to the left rail 40, sets a group to the left of the groove, and makes it one of the multiple objects Coi. Similarly, if the yi coordinate of the groove is smaller than Ci, the system 10 determines that it corresponds to the right rail 40, sets a group to the right of the groove, and makes it one of the multiple objects Coi.

[0105] In the case of FIG. 14, the yi coordinate of the groove corresponding to w1 is larger than Ci, so it corresponds to the left rail 40, and a group is set to the left of the groove, and it is one of the multiple objects Coi. The yi coordinate of the groove corresponding to w2 is smaller than Ci, so it corresponds to the right rail 40, and a group is set to the right of the groove, and it is one of the multiple objects Coi. The dimensions of the group are set in advance based on the vertical and horizontal dimensions of the rail cross section. According to the system 10 of Example 6, even if it is not possible to directly acquire a point cloud of the rail side at a railroad crossing, etc., it is possible to detect objects including the rail 40 as long as the groove is detected, so rail detection is possible. [Example]

[0106] In the seventh embodiment, a branch opening direction estimation process is added to estimate the opening direction of a branch when a branch has occurred. The branch opening direction estimation process is added immediately before rail curve complementation S108, and also transmits information on the opening direction to rail curve complementation S108. After the left rail pair and right rail pair are complemented in rail curve complementation S108, the rail detection results and information on the opening direction are output as an estimated rail curve.

[0107] FIG. 15 is a plan view for explaining a branch determination method in the system of Example 7. That is, FIG. 15 is a diagram showing a method for estimating a branch opening direction in Example 7. Example 7 uses information on third rail candidate objects estimated by the method described in Example 2 or Example 4. Specifically, it uses a left rail pair total point group Cp3(p=l) in the inspection section D0, D1, ..., Dn(p=l) and a right rail pair total point group Cp3(p=r) in the inspection section D0, D1, ..., Dn(p=r).

[0108] First, find the closest position xmin(p=l) in the direction of travel of the left rail pair total point group Cp3(p=l) and the closest position xmin(p=r) in the direction of travel of the right rail pair total point group Cp3(p=r). If there is a branching tongue rail 40 ahead of the vehicle, the right rail pair or left rail pair will be located after the tongue rail 40, so xmin(p=l) and xmin(p=r) will no longer be the same.

[0109] Therefore, if xmin(p=l) and xmin(p=r) are apart by a predetermined distance or more, the position of xmin(p=l) or xmin(p=r) that is farther from the vehicle can be regarded as the tip position of the tongue rail 40. Next, the width of the left and right rails 40 of the left rail pair and the width of the left and right rails 40 of the right rail pair at the tip position of the tongue rail 40 are calculated.

[0110] One method for calculating the width of the left and right rails 40 is to calculate the point on the left rail 40 and the point on the right rail 40 that are closest in x-coordinate to the tip position of the tongue rail 40 from the rail pair total point group Cp3, and use the absolute value of the difference between the point on the left rail 40 and the point on the right rail 40 as the width of the left and right rails 40.

[0111] Furthermore, the value Dtr obtained by subtracting the width of the right rail pair from the width of the left rail pair (40) is used as the traffic direction determination value. If the traffic direction determination value is positive, the left rail pair is considered to be open, and if it is negative, the right rail pair is considered to be open.

[0112] The above is the method for determining the open direction of a branch in the seventh embodiment. In addition, if necessary, only the rail pair in the open direction is sent to rail curve complement S108. With the above method, when a branch occurs, it is possible to determine whether the left or right side is open, and in the case of a system that needs to detect only open branches, more stable rail detection is possible. [Example]

[0113] In the eighth embodiment, an installation setting will be described that enables the laser radar 101 to capture as many point clouds of the rail 40 as possible. Fig. 16 is a plan view for explaining the installation method of the sensor 101 in the system 10 of the eighth embodiment. That is, Fig. 16 is a diagram showing the detectable range of rail objects relative to the installation angle of the laser radar 101, and G61 is the installation method recommended in the eighth embodiment.

[0114] 16 (G60), the point cloud corresponding to the ground surface acquired by the laser radar 101 has a layer structure L0, and the angle between the layer structure L0 and the curved rail R in the inspection section D2 is close to 0 degrees. When set in this way, the system 10 cannot acquire a point cloud corresponding to the rail 40, and is unable to detect the rail in the inspection section D2.

[0115] On the other hand, if the laser radar 101 is installed so that the point cloud corresponding to the ground has a layer structure L1 as shown in Figure 16 (G61), the angle between the layer structure L1 and the curved rail R in the inspection section D2 is maintained away from 0 degrees. The system 10 configured in this way can acquire a point cloud corresponding to the rail 40, enabling rail detection in the inspection section D2.

[0116] Therefore, in the system 10 of Example 8, the mounting position and mounting angle of the laser radar 101 are set so that the angle between the rail curve expected during operation and the layer structure irradiated onto the ground is equal to or greater than a predetermined value. The system 10 of Example 8 can maintain rail detection far ahead of the vehicle 20 even when the vehicle 20 approaches a curve, thereby enabling more stable rail detection. Note that the layer structures L0 and L1 refer to the trajectory of the laser irradiation sweep operation of the laser radar 101.

[0117] [supplement] Rail detection methods can be broadly divided into two types. In the first method, rails ahead of the vehicle are detected from the position estimation results by comparing the vehicle's self-position estimation results with a map that clearly shows the latitude and longitude of the rail locations. In this method, the accuracy of rail detection depends on the accuracy of the self-position estimation. Therefore, there is a problem in that rail detection is difficult in areas where the self-position is uncertain. The present system 10 employs the second method, which can solve this problem.

[0118] In the second method, a sensor 101 mounted on a vehicle acquires data in front of the vehicle, including the rails 40, extracts the rails 40 in front of the vehicle from the sensor output, and recognizes their location. This rail detection method using the sensor 101 can detect the rails 40 without relying on the accuracy of self-position estimation.

[0119] The main configuration, operation, and effects of the system 10 will be summarized below. [1] The system 10, shown in Figures 1, 3, and 8, is a forward monitoring system equipped with a sensor 101 that captures a trajectory 40 that guides a moving object 20, and a computer that processes the sensor output to improve the accuracy of the monitoring information. The system 10 includes a laying surface point cloud removal unit, a first narrowing down unit, a second narrowing down unit, and a trajectory curve completion unit.

[0120] The sensor output includes a point cloud corresponding to the track 40 and a point cloud corresponding to the surface of the track 40. The surface point cloud removal unit extracts object information by removing the surface point cloud from the sensor output. As a result, the surface point cloud removal unit extracts objects protruding from the surface of the track 40.

[0121] The first narrowing down unit extracts a first trajectory candidate object by comparing the trajectory cross section stored in the computer's memory with the cross-sectional shape of the object. The second narrowing down unit detects a second trajectory candidate object by comparing the width of the first trajectory candidate object with the width of the trajectory cross section. The trajectory curve complementing unit complements the trajectory objects including the second trajectory candidate object and recognizes the trajectory 40 by connecting the points to estimate the outline.

[0122] The system 10 provides the position information of the track 40 thus recognized to the operator or monitoring entity (hereinafter also referred to as the "operator, etc."). The operator, etc., who receives this information, is assisted in determining whether or not there is an obstacle 301 to the moving object 20. According to the system 10, even if the point cloud corresponding to the track 40 acquired by the sensor 101 is incomplete, the system 10 can stably recognize the location of the track 40. In other words, the system 10 has the effect of supporting the monitoring of the operator, etc., by recognizing the location of the track 40 with high accuracy.

[0123] [2] The track 40 in the present system 10 described in [1] above may be configured with a plurality of rails 40, for example, two pairs of rails. The rails 40 are not limited to applications for long-distance transportation, and examples of mobile objects 20 that travel short distances of several hundred meters include cranes such as gantry cranes. The present system 10 may have the configurations shown in Figures 6, 7, 9, and 11.

[0124] The laying surface is the ground, for example, a surface including the curve lgi in Figure 6. The laying surface point cloud removal unit is a ground point cloud removal unit. The first track candidate objects are multiple objects including the first rail candidate object Co1i, but are not limited to a square (rectangular) shape as shown in Figure 6. The first narrowing down unit extracts the first multiple rail candidate objects by comparing the cross-sectional dimensions of the multiple objects with the rail cross-sectional dimensions stored in the computer's memory. The second narrowing down unit detects the second rail candidate object Co2i by comparing the dimensions between the first multiple rail candidate objects with the dimensions between the multiple rails, for example, the track gauge.

[0125] The track curve complementing unit detects the rail 40 by complementing a plurality of objects including the second rail candidate object Co2i. In this form, even if the point cloud corresponding to the rail 40 acquired by the sensor 101 is incomplete, the system 10 can stably recognize the rail 40 because the actual rail 40 laid on the route on which the moving object 20 travels is stored in the computer memory and the cross-sectional dimensions are compared with the sensor output.

[0126] [3] In the present system 10 described in [2] above, the first narrowing-down unit may have a first incomplete template created based on the rail cross-sectional dimensions as shown in FIG. 6. This first incomplete template may be a rectangle (basic) or any shape created based on the maximum length and width dimensions of the rail cross-section. The present system 10 may extract, as the first plurality of rail candidate objects, multiple objects whose cross-sectional dimensions fall within the range of the first incomplete template. This form of the present system 10 narrows down the candidates using the first incomplete template, so that the rail 40 can be detected with a certain degree of certainty.

[0127] [4] In the present system 10 described in [2] above, the first narrowing-down unit may have a second incomplete template created based on the rail cross-sectional shape as shown in Fig. 11 of Example 3. The present system 10 may extract, as a pair of rails 40, a first plurality of rail candidate objects, if the matching rate between the point cloud formed by the plurality of objects and the second incomplete template is equal to or greater than a predetermined value.

[0128] In this embodiment, the system 10 can detect the rails 40 with improved accuracy by further narrowing down the number of templates using the second incomplete template in the first narrowing-down section. The second incomplete template shown in Fig. 11 improves detection accuracy by capturing the characteristic shape of the rail side surface by taking advantage of the characteristic of the rail 40 in which the rusted side surfaces other than the top surface are more likely to reflect laser radar.

[0129] [5] In the present system 10 described in [2] above, the second narrowing-down unit may determine a plurality of representative positions of the first plurality of rail candidate objects from the track axis prediction line yi=ci shown in Figures 7 and 8. The distance between the plurality of representative positions is determined based on the dimensions (gauge) between the plurality of rails (S106 in Figure 8), and the first rail candidate object Co1i, which is a predetermined distance, may be determined as the second rail candidate object Co2i. In this form of the present system 10, the second narrowing-down unit may estimate, for example, the second rail candidate object Co2i that matches the gauge stored in the computer memory as the rail 40 with a high probability.

[0130] [6] In the present system 10 described in [2] above, the moving object 20 may be specialized for a railway vehicle 20, as shown in FIG. 1. This form of the present system 10 further includes a third narrowing-down unit, which will be described with reference to FIGS. 2 and 8. The third narrowing-down unit may extract second rail candidate objects Co2i within a predetermined range as third rail candidate objects Co3i for a first rail candidate line created using rail detection results at positions closer to the vehicle 20 than a predetermined section. The present system 10 applied to a railway vehicle 20 can detect rails 40 with sufficient accuracy, taking practicality into consideration.

[0131] [7] In the present system 10 described in [6] above, when there are multiple previous section rail objects in the section immediately preceding the predetermined section, the third narrowing-down unit performs the processing described in Figures 9 and 10. That is, the third narrowing-down unit separates the previous section rail candidate objects into a left rail pair including a rail object on the left and a right rail pair including a rail object on the right.

[0132] This embodiment of the system 10 extracts second rail candidate objects Cp2i within a predetermined range based on the rail detection results for the previous section that is closer to the vehicle 20 than the predetermined section, and the left rail candidate line created using the left rail pair. Similarly, the system 10 extracts second rail candidate objects Cp2i within a predetermined range based on the rail detection results for the previous section that is closer to the vehicle 20 than the predetermined section, and the right rail candidate line created using the right rail pair. In this way, the system 10, which is specialized for a railway vehicle 20 on which two rails 40 are laid parallel to each other, can detect the two rails 40 with sufficient accuracy.

[0133] [8] The system 10 described in [2] above may further include a fourth narrowing down unit. This fourth narrowing down unit extracts second rail candidate objects Cp2i within a predetermined distance from the second rail candidate line as fourth rail candidate objects, as shown in G50 in Fig. 12. The fourth rail candidate objects are extracted as the smallest objective function created based on the distances between multiple second rail candidate lines and the second rail candidate objects Cp2i.

[0134] The multiple second rail candidate lines are created based on the track axis estimated positions obtained from the track axis predicted line yi=ci. Here, instead of the second rail candidate object Cp2i, the second rail candidate point cloud that constitutes it may be used. In this form, the system 10 can detect the rails 40 with high accuracy even in curved sections by measuring the gauge of the two laid rails 40 using the fourth narrowing unit.

[0135] [9] In the present system 10 described in [8] above, the fourth narrowing-down unit may exclude the fourth rail candidate object from the second rail candidate objects Cp2i to determine the other rail candidate objects, as shown in G51 in Fig. 12. In this form, the present system 10 may extract other rail candidate lines for which an objective function created based on the distances between a plurality of second rail candidate lines and other rail candidate objects is equal to or less than a predetermined value.

[0136] Here, instead of the other-rail candidate objects, a point cloud of other-rail candidate objects that constitute the other-rail candidate objects may be used. The present system 10 may further have a function of adding the extracted other-rail candidate line and other-rail candidate objects at a predetermined distance to the fourth rail candidate objects. In this form, the present system 10 can detect, if there is a tongue rail that constitutes a point machine in addition to the two laid rails 40, by using the fourth narrowing-down unit, since it is located at a short distance from the estimated track axis position.

[0137]

[10] The system 10 in [2] above may extract objects protruding below the ground level detected by the ground point cloud removal unit, relative to the ground level in the range immediately preceding the predetermined range, as shown in Figure 13 of Example 5 and Figure 14 of Example 6. Furthermore, when the ground level in the predetermined range rises discontinuously, the system 10 may detect such raised ground as a road intersecting the track or a road that also serves as the track surface. In other words, this form of the system 10 can detect rails not only on the track surface of a tramway, but also in cases where only the top surface of the rail is exposed and the rest of the rail is buried, such as at a railroad crossing.

[0138]

[11] In the above [7], the system 10 determines the presence of a point based on the following criteria, as shown in Fig. 15. That is, the system 10 determines the presence of a point when the relative positions of the nearest position xmin(p=l) in the direction of travel of the left rail pair total point group Cp3(p=l) corresponding to the left rail pair and the nearest position xmin(p=r) in the direction of travel of the right rail pair total point group Cp3(p=r) corresponding to the right rail pair are separated by a predetermined distance or more.

[0139] The system 10 also calculates the difference between the width of the left rail pair minus the width of the right rail pair in the section farther from the vehicle 20 between the nearest position xmin(p=l) in the direction of travel of the left rail pair point cloud and the nearest position xmin(p=r) in the direction of travel of the right rail pair point cloud. Depending on the result, the system 10 determines that the left rail 40 is open if the difference is positive, and determines that the right rail 40 is open if the difference is negative. This type of system 10 can determine the open direction when there is a point on the rail 40.

[0140]

[12] When the system 10 described in

[11] above determines that the left rail 40 is open, it detects only the left rail 40 in the subsequent section. When the system 10 determines that the right rail 40 is open, it detects only the right rail 40 in the subsequent section. This type of system 10 detects rails only on the side of the point determined to be an open section, resulting in less waste and better monitoring efficiency.

[0141]

[13] As shown in Figure 16, the layer structures L0 and L1 and their sweep states, in the present system 10 described in [1] above, the sensor 101 may organize the acquired point cloud in layers. The mounting position and mounting angle of the sensor 101 may be such that the expected track direction and layer direction are within a predetermined angle range that allows rail detection to be maintained. This type of present system 10 can maintain high-precision rail detection at positions away from the moving object 20 even when the moving object 20 approaches a curve, enabling more stable rail detection.

[0142]

[14] In the present system 10 described in [2] to

[13] above, as shown in Fig. 1, the sensor 101 is installed on the vehicle 20 traveling along the rails 40, and may further include a rail detection unit 104 and an obstacle determination unit 105. The rail detection unit 104 detects the rails 40 based on information acquired by the sensor 101.

[0143] The obstacle determination unit 105 detects obstacles 301 within a predetermined range that is set based on the rails 40 detected by the rail detection unit 104. The system 10 configured as described above determines whether or not there are obstacles 301 to the vehicle 20, and provides the information to the operator, etc., who are then supported in driving safely.

[0144] This method can be summarized as follows.

[15] As illustrated in Figures 2 and 8, this method is a forward monitoring method that uses information obtained by processing the sensor output that captures the trajectory 40 that guides the moving body by a computer, and the computer performs the following processing by executing a program stored in memory and forming each functional unit that processes information.

[0145] First, the laying surface of the track 40 is detected from the point cloud (D101) captured by the sensor 101 (S101-S103). Next, the laying surface point cloud removal unit extracts objects protruding from the detected laying surface (S104). Next, the first narrowing down unit extracts first track candidate objects by comparing the cross-sectional shape of the extracted objects with the track cross section stored in the computer (S105). The second narrowing down unit compares the width of the detected first track candidate object with the width of the track cross section to detect second track candidate objects (S106).

[0146] The trajectory curve complementing unit complements (S108) the trajectory objects, including the detected second trajectory candidate object, thereby more accurately recognizing the trajectory 40 (D102). Position information of the trajectory 40 based on this recognition is provided to the operator, etc. As a result, the operator, etc. is assisted in easily detecting whether or not there is a traveling obstacle 301 in the vicinity ahead of the moving object. According to this method, even if the point cloud corresponding to the trajectory 40 acquired by the sensor 101 is incomplete, the trajectory 40 can be stably recognized. This method outputs information that recognizes the trajectory 40 with high accuracy, and can support the operator, etc. in safe driving. [Explanation of symbols]

[0147] 10...Forward monitoring system, 20...Vehicle (moving object), 40...Rail 40, 50...Vehicle running area, 101...Forward monitoring sensor, 103...Object detection unit, 104...Rail detection unit, 105...Obstacle determination unit, 301...Object in front of the vehicle (e.g., obstacle such as a person), Co1i...First rail candidate object, Cp2i...Second rail candidate object, Co3i...Third rail candidate object, Ai...Estimated track center, Cp3(p=l)...Left rail pair total point cloud, Cp3(p=r)...Right rail pair total point cloud, D0 to D2...Inspection section, Point cloud data D101...Point cloud data D101, D102...Estimated rail curve, To2A, To2B...Second incomplete template, Tp1, Tp2...Point cloud (reproducing only the side of the rail 40), O...Origin, xmin(p=l)...Closest position in the direction of travel of all point clouds of the left rail pair, xmin(p=r)...Closest position in the direction of travel of all point clouds of the right rail pair, w1, w2...Distance between groups, yi...Track center line, yi=ci...Predicted track axis line, Σ0...Vehicle-fixed three-dimensional xyz coordinate system (coordinate system Σ0), Δzg...Threshold

Claims

1. A forward monitoring system comprising a sensor for capturing a trajectory for guiding a moving body and a computer for processing the sensor output, a laying surface point cloud removal unit that removes a laying surface point cloud from the sensor output; a first narrowing-down unit that extracts a first trajectory candidate object by comparing the cross-sectional shape of the object, which is information contained in the remaining object after the removal, with trajectory cross sections stored in a memory of the computer; a second narrowing unit that detects a second trajectory candidate object by comparing a width of the first trajectory candidate object with a width of the trajectory cross section; a trajectory curve complementing unit that complements trajectory objects including the second trajectory candidate object to recognize the trajectory; and the sensor output includes a point cloud corresponding to the track and a point cloud corresponding to a surface on which the track is laid; the installation surface point cloud removal unit extracts object information from the sensor output; The trajectory curve complementing unit recognizes an object protruding from the installation surface as a trajectory based on information about the extracted object, outputting information to assist in determining whether or not there is a traveling obstacle in the vicinity of the recognized trajectory; Forward monitoring system.

2. The track is composed of a plurality of rails, The laying surface is the ground, the laying surface point cloud removal unit is a ground point cloud removal unit, the first track candidate objects are a plurality of objects including a first plurality of rail candidate objects; the first narrowing unit extracts the first plurality of rail candidate objects by comparing cross-sectional dimensions of the plurality of objects with cross-sectional dimensions of a rail; the second narrowing unit detects a second rail candidate object by comparing a dimension between the first plurality of rail candidate objects with a dimension between the plurality of rails; the track curve complementing unit is a rail curve complementing unit that detects a rail by complementing a plurality of objects including the second rail candidate object; The forward monitoring system of claim 1 .

3. The first narrowing portion is A first incomplete template is created based on a rail cross-sectional dimension, The first incomplete template is a rectangle or an arbitrary shape created based on the maximum length and width dimensions of the rail cross section, extracting the plurality of objects whose cross-sectional dimensions are within the range of the first incomplete template as first plurality of rail candidate objects; The forward monitoring system according to claim 2 .

4. The first narrowing portion is a second incomplete template created based on the rail cross-sectional shape; extracting, as first plurality of rail candidate objects, a plurality of objects for which a matching rate between the point cloud constituting the plurality of objects and the second incomplete template is equal to or greater than a predetermined value; The forward monitoring system according to claim 2 .

5. The second narrowing portion is determining a plurality of representative positions of the first plurality of rail candidate objects from the track axis prediction line; A first rail candidate object is defined as a second rail candidate object when the distance between the plurality of representative positions is a predetermined distance calculated based on the dimensions between the plurality of rails. The forward monitoring system according to claim 2 .

6. the moving body is a railway vehicle, The system further includes a third narrowing-down unit that extracts the second rail candidate objects within a predetermined range as third rail candidate objects from a first rail candidate line created using rail detection results for a previous section that is closer to the vehicle than the predetermined section. The forward monitoring system according to claim 2 .

7. The third narrowing portion is A plurality of preceding section rail objects present in the section immediately preceding the predetermined section a left rail pair including a left rail object; a right rail pair including a right rail object; Separate into extracting the second rail candidate object from a predetermined range on a left rail candidate line created using the rail detection result of the previous section that is closer to the vehicle than the predetermined section and the left rail pair; extracting the second rail candidate object from a predetermined range on a right rail candidate line created using the rail detection result of the previous section that is closer to the vehicle than the predetermined section and the right rail pair; The forward monitoring system according to claim 6.

8. a plurality of second rail candidate lines created based on track axis estimated positions obtained from the track axis predicted line; the second rail candidate object or a second rail candidate point group constituting the second rail candidate object; extracting a second rail candidate line for which an objective function created based on the distance is minimum; extracting the second rail candidate object from within a predetermined distance range from the second rail candidate line as a fourth rail candidate object; Further having a fourth narrowing section, The forward monitoring system according to claim 2 .

9. The fourth narrowing portion is The fourth rail candidate object is excluded from the second rail candidate objects to be set as other rail candidate objects; the plurality of second rail candidate lines; the other-rail candidate object or a group of other-rail candidate object points constituting the other-rail candidate object; extracting candidate rail lines for which the objective function created based on the distance is equal to or less than a predetermined value; an other-rail candidate object within a predetermined distance from the other-rail candidate line is recognized as the fourth rail candidate object; The forward monitoring system of claim 8.

10. For the ground height in the range immediately before the specified range, If the ground level in a predetermined range rises discontinuously, it is detected as a road intersecting the track or a road that also serves as the paved surface; extracting objects protruding below the ground surface detected by the ground point cloud removal unit; The forward monitoring system according to claim 2 .

11. It is possible to determine the opening direction when there is a point machine in the middle of the rail, the nearest position in the direction of travel of the left rail pair point cloud corresponding to the left rail pair; the closest position in the direction of travel of the right rail pair point cloud corresponding to the right rail pair; If the distance is greater than or equal to a predetermined value, The nearest position in the direction of travel of the left rail pair point cloud, the closest position in the direction of travel of the right rail pair point cloud; Calculate the difference between the width of the left rail pair farther from the vehicle and the width of the right rail pair, If the difference is a positive value, the left rail is determined to be open. If the difference is a negative value, it is determined that the right rail is open. The forward monitoring system according to claim 7.

12. If it is determined that the left rail is open, only the left rail is detected for the section thereafter, If it is determined that the right rail is open, only the right rail is detected for the following section. The forward looking system of claim 11.

13. The sensor organizes the acquired point cloud in layers, The mounting position and mounting angle of the sensor are as follows: The angle of the layer direction relative to the expected trajectory direction is within a predetermined range. The forward monitoring system of claim 1 .

14. a sensor installed on a vehicle traveling along a rail; a rail detection unit that detects rails based on information acquired by the sensor; setting a predetermined range based on the rail detected by the rail detection unit; an obstacle determination unit that detects obstacles within the predetermined range; having A forward monitoring system according to any one of claims 2 to 13.

15. A forward monitoring method based on information obtained by processing, by a computer, sensor outputs that capture a trajectory for guiding a moving object, comprising: The computer executes a program stored in a memory and performs information processing on each of the functional units formed by the program, thereby a laying surface point cloud removal unit extracts an object protruding from the laying surface detected from the sensor output including the point cloud corresponding to the trajectory; a first narrowing-down unit extracts a first trajectory candidate object by comparing a cross-sectional shape of the object with the trajectory cross section stored in the memory; a second narrowing unit detecting a second trajectory candidate object by comparing a width of the first trajectory candidate object with a width of the trajectory cross section; a trajectory curve complementing unit that complements trajectory objects including the detected second trajectory candidate object to detect the trajectory; A forward monitoring method that assists in determining whether or not there is a traveling obstacle in the vicinity of the detected track.

Images