Obstacle detection system

Abstract

To reduce time required for obstacle detection processing by image processing while selecting an appropriate ROI for improving obstacle detection performance in an obstacle detection system mounted on a track transport system.SOLUTION: In order to solve the above problem, an obstacle detection system includes a track detection unit that detects track information from data of an external sensor for recognizing an environment around a train, an ROI processing unit that cuts out a region of interest to be subjected to an object detection process from the track information, and an object detection unit that detects an object, wherein the obstacle detection system includes a track information recording unit that records the track information detected by the track detection unit, and the ROI processing unit changes the track information to be used for cutting out the region of interest on the basis of any one of information on a size of the region of interest and a track shape in front of the train.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an obstacle detection system mounted on a rail transportation system that travels on a fixed track. [Background technology]

[0002] In rail transport systems, where transport vehicles travel on a track, detecting obstacles on the track is important for improving the safety and operability of rail transport systems, since obstacles on the track cannot be avoided by steering. In recent years, research has been conducted on obstacle detection support systems that use external sensors such as millimeter-wave radar, laser radar, and cameras to detect obstacles on the track and support the driver by notifying the driver of the detection results. One example of detecting obstacles on the track using external sensors is obstacle detection using cameras and image processing. As a technology for recognizing and detecting objects using image processing, there is general object recognition technology, whose accuracy has rapidly improved in recent years thanks to deep learning. However, general object recognition technology tends to have a lower detection accuracy for objects with a small number of pixels in the image.

[0003] Generic object recognition technology has difficulty detecting objects with a small number of pixels in an image. Due to limitations in the computational cost of real-time processing, the input image size for generic object recognition technology is limited to a certain size, such as 608 x 608 or 416 x 416. Therefore, even if a high-resolution image is captured, it will be reduced in size when input into generic object recognition technology.

[0004] When the image is reduced, it becomes difficult to detect objects with a small number of pixels, such as obstacles at a distance. In order to detect objects with a small number of pixels, one possible solution is to set a specific area within the image as a ROI (Region of Interest) and process only the ROI. In other words, the extracted ROI is enlarged or reduced to the input image size, such as 608 x 608 or 416 x 416. However, since the area outside the ROI is not processed, it is necessary to select the coordinates and size of the ROI appropriately in order to achieve the purpose of obstacle detection.

[0005] Patent Document 1 discloses the following technology with the aim of providing a system for detecting objects in an image at high speed and with high accuracy: "In a system 10 for detecting obstacles within a structure gauge, an auxiliary storage device 113 includes an image acquisition unit 101 that acquires images captured from a moving object moving on a trajectory, a trajectory detection unit 103 that detects the trajectory in the image, an ROI processing unit 106 that determines the width of the trajectory at a reference ordinate value selected in the image, determines the size of a region of interest in the image based on the trajectory width and a coefficient, and determines the position coordinates of the region of interest in the image based on the selected ordinate value and the detected trajectory, and an obstacle detection unit 107 that executes object detection processing in the region of interest at the determined position coordinates." [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Publication No. 2023-34917 Summary of the Invention [Problem to be solved by the invention]

[0007] The obstacle detection system detects obstacles on the track and outputs a brake operation request to avoid collisions. Therefore, the shorter the processing time required for obstacle detection, the more likely it is that the train will run at a speed that avoids collisions with obstacles. As a result, the travel time between stations can be shortened, achieving both safety and convenience. However, the technology described in Patent Document 1 selects an ROI based on track information, so obstacle detection by image processing cannot begin until track detection processing, which identifies the track position on the image, is complete. It is generally known that track detection processing by image processing imposes a high computational load. This poses a problem of extending the time required for obstacle detection processing by image processing.

[0008] In order to address the above-mentioned problems, the present invention aims to reduce the time required for obstacle detection processing using image processing in an obstacle detection system installed in a rail transportation system, while selecting an appropriate ROI to improve obstacle detection performance. [Means for solving the problem]

[0009] In order to solve the above problems, one representative obstacle detection system of the present invention is an obstacle detection system that includes a track detection unit that detects track information from data from external sensors that recognize the environment around the train, an ROI processing unit that extracts a region of interest to be processed for object detection from the track information, and an object detection unit that detects objects, and is characterized by including a track information recording unit that records the track information detected by the track detection unit, and the ROI processing unit that changes the track information used to extract the region of interest based on information on either the size of the region of interest or the shape of the track ahead of the train. [Effects of the Invention]

[0010] According to the present invention, in an obstacle detection system installed in a rail transportation system, it is possible to select an appropriate ROI to improve obstacle detection performance while reducing the time required for obstacle detection processing using image processing.

[0011] Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a diagram illustrating a configuration of an obstacle detection system according to a first embodiment of the present invention. [Figure 2] FIG. 2 is a diagram schematically illustrating an example of an input image. [Figure 3] 1 is a flowchart showing a process executed by an ROI processing unit in the first embodiment of the present invention. [Figure 4] FIG. 10 is a diagram showing the relationship between an input image from an external sensor and an ROI cut-out image. [Figure 5] FIG. 10 is a diagram showing how the track appears for each calculation cycle when a train is running on a curved section. [Figure 6] 10 is a flowchart showing a process executed by an ROI processing unit in Example 2 of the present invention. [Figure 7] FIG. 10 is a diagram schematically illustrating a method for determining straight sections and curved sections in a second embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, an embodiment will be described with reference to the drawings. [Example]

[0014] The configuration of an obstacle detection system according to a first embodiment of the present invention will be described with reference to FIGS.

[0015] 1 is a diagram showing the configuration of an obstacle detection system. The obstacle detection system is composed of an external sensor 110 mounted on a train 100 and capable of sensing the external environment, and an obstacle detection device 120 that can detect obstacles ahead of the train 100. The obstacle detection device 120 is composed of a track detection unit 121 that can detect information about the track (track information) on which the train 100 is traveling based on sensor information from the external sensor 110, a track information recording unit 122 that can record the track information detected by the track detection unit 121 in chronological order, an ROI processing unit 123 that can determine an ROI based on the track information from the track detection unit 121 and past track information from the track information recording unit 122 and generate an image by cutting out one or more ROI regions from an input image, and an object detection unit 124 that can detect objects based on the ROI region cut-out image from the ROI processing unit 123.

[0016] An object detected by the object detection unit 124 is determined to be an obstacle if the object's three-dimensional position is near the track detected by the track detection unit 121. When the obstacle detection device 120 detects an obstacle near the track, it determines whether or not it is necessary to immediately output a brake operation request based on the current train speed. When there is an obstacle ahead of the train and immediate brake operation is required, the obstacle detection device 120 outputs a brake operation request to a safety device (not shown) or a driver (not shown).

[0017] The external sensor 110 senses the conditions around the train 100 (particularly ahead) and transmits the sensing data to the track detection unit 121 and the ROI processing unit 123. The external sensor 110 includes a camera, a LIDAR (laser range finder), a millimeter wave radar, etc. The camera includes a monocular camera, a stereo camera, an infrared camera, etc. It is common for multiple sensors to be mounted for redundancy. When the train 100 moves forward in the direction of travel, the external sensor installed in the leading car on the side facing the direction of travel is used. In this embodiment, a case will be described in which a camera is used as the external sensor.

[0018] 2 is a diagram showing a schematic example of an input image. The input image 200 includes a track 210 extending forward from the train 100. The track 210 includes a left rail 211 and a right rail 212, which define the width of the track 210. The upper left corner of the input image 200 is taken as the origin, with the vertical direction being the v coordinate and the horizontal direction being the u coordinate.

[0019] The track detection unit 121 detects the track along which the train 100 is traveling from the input image. Track detection may be performed by image processing, or the coordinates of the track within the input image may be calculated using GNSS (Global Navigation Satellite System) or map data. For example, a track over a long distance that cannot be detected by a camera may be calculated using GNSS or map data. Methods for detecting rails through image processing include, for example, a method for searching for rails using brightness gradient information within an image or a method for detecting rails using semantic segmentation using DNN. In the present invention, any method is acceptable as long as the rails can be detected. Because the track extends from the front end of the train 100, it can be easily detected by comparing it with obstacles, etc. Track coordinates, for example, indicate the center line between the left and right rails. The upper left corner of the input image is the origin, and the vertical direction is the v coordinate and the horizontal direction is the u coordinate. The track coordinates indicate the u coordinate and v coordinate of each point on the center line. For example, the v coordinate value increases from top to bottom of the image and the u coordinate value increases from left to right.

[0020] The trajectory information recording unit 122 records the trajectory information detected by the trajectory detection unit 121. The trajectory information recording unit 122 transmits to the ROI processing unit 123 the trajectory information detected by the trajectory detection unit 121 one cycle before the current calculation cycle.

[0021] The ROI processing unit 123 generates one or more ROI clipped images, which serve as input images for object detection processing by image processing, from the track information for the current calculation cycle received from the track detection unit 121, the track information for the cycle before the current calculation cycle received from the track information recording unit 122, and the input image from the external sensor 110. The generated ROI clipped images are sent to the object detection unit 124. Note that although this embodiment describes an example in which multiple ROI regions are defined, a single ROI region may be used. For example, if the speed of the train 100 is slow, an ROI region may be defined only for a short distance. The number of ROI regions defined may vary depending on the region in which an object is desired to be detected (focused on).

[0022] The object detection unit 124 performs object detection using the ROI cropped image received from the ROI processing unit 123. Object detection processing using image processing includes a method of creating a parallax image using a stereo camera and detecting an object ahead from the parallax image, and a method of detecting an object within an image using a DNN. A DNN is one of the methods used in machine learning, and detects various objects by extracting and learning the features of the object. Methods other than those described above may also be used for object detection processing using image processing, and the present invention does not limit the method.

[0023] Next, the processing of the ROI processing unit 123 according to the first embodiment of the present invention will be described with reference to Fig. 3 to Fig. 5. In the first embodiment, an example will be described in which there are two ROI regions, one for long distances and one for short distances, but as mentioned above, the number of ROI regions is not important.

[0024] FIG. 3 is a flowchart showing the processing executed by the ROI processing unit 123.

[0025] Step 301: An image is received as sensor information from the external sensor 110. The input image is, for example, an image as shown in Figure 2. The process proceeds to step 302.

[0026] Step 302: Determining the ROI region. Figure 4 shows the relationship between the entire input image 400 from the external sensor and the ROI cropped image. To perform object detection using image processing across all distances, from close to long distances, it is necessary to define ROI regions of various sizes. For example, when detecting close-range objects, the detection target occupies a larger proportion of the image. Furthermore, even for objects located at the same distance from the center of trajectory in the real world, the deviation (number of pixels) from the center of trajectory is greater at close range than at long range. Therefore, a larger ROI region 401 must be defined to detect close-range objects. On the other hand, even for objects of the same size in the real world, the number of pixels is smaller at long range than at close range. In such cases, object detection performance can be improved by cropping the region near the detection target and generating a pseudo-long-range image (increasing the number of pixels). Therefore, a smaller ROI region 402 must be defined to detect long-range objects. In step 302, it is determined whether the ROI region to be cropped is for close range (large ROI region) or long range (small ROI region).

[0027] Step 303: If the ROI region to be extracted is for a short distance, the process proceeds to step 304; if the ROI region to be extracted is for a long distance, the process proceeds to step 305.

[0028] Step 304: When the ROI region is for short distances, track information from one cycle before the current calculation cycle is obtained from the track information recording unit 122 as the track information used in the ROI determination process. Figure 5 shows how the track appears for each calculation cycle when the train 100 is traveling on a curved section. Compared to track 510 from one cycle before the current calculation cycle, track 520 in the current calculation cycle is observed such that the distant track 520 is further away from the center of the input screen 500 as the train approaches the curved section. ROI 512 is for short distances and ROI 513 is for long distances, selected based on the information on track 510 from one cycle before the current calculation cycle. ROI 522 is for short distances and ROI 523 is for long distances, selected based on the information on track 520 in the current calculation cycle. There is no significant change in the positions of short distance ROIs 512 and 522. On the other hand, the positions of long distance ROIs 513 and 523 change significantly.

[0029] As described above, at short distances, the change in the positions (u, v coordinates) of the trajectory centers 511, 521 on the image for each calculation cycle is small, and the ROI area is large. Therefore, even if the ROI is determined based on the trajectory information from one calculation cycle before the current calculation cycle, the trajectory will be included in the ROI cropped image. Therefore, even if the ROI cropped image is generated using the trajectory information from one calculation cycle before the current calculation cycle, it is possible to detect objects near the trajectory. Furthermore, since the trajectory information from one calculation cycle before the current calculation cycle is used, there is no need to wait for trajectory information from the trajectory detection unit 121 in the current calculation cycle, and object detection processing using image processing for short distances can begin at the start of the current calculation cycle. Therefore, it is possible to shorten the processing time for object detection processing using image processing. Proceed to step 306.

[0030] In Figure 5, there are parts of the track 520 for the current calculation cycle that are not included in either the short-distance ROI 522 or the long-distance ROI 523. This is an exaggerated representation of the large changes in the positions of the long-distance ROIs 513 and 523 as the train approaches a curved section.

[0031] Step 305: If the ROI region is for long distances, the orbit information for the current calculation cycle is obtained from the orbit detection unit 121 as the orbit information to be used in the ROI determination process. As shown in Figure 5, for long distances, the positions (u, v coordinates) of the orbit centers 511 and 521 on the image change significantly for each calculation cycle. Also, since the ROI region is small, if the ROI is determined based on the orbit information for the previous calculation cycle (ROI for long distances 513), the orbit will not be included in the ROI cropped image. Therefore, it is necessary to generate an ROI cropped image corresponding to the ROI (ROI for long distances 523) determined using the orbit information for the current calculation cycle. Proceed to step 306.

[0032] Step 306: The ROI is determined based on the trajectory information. As mentioned above, when determining an ROI for a short distance, the trajectory information received is the trajectory information from one calculation cycle before the current calculation cycle, and when determining an ROI for a long distance, the trajectory information for the current calculation cycle is received. In the ROI determination process, the v coordinate of the detection start point in the input image (the point farthest from the vehicle on the detected trajectory) is determined. If the ROI region to be extracted is for a long distance, the v coordinate of the point farthest from the vehicle in the coordinates of the detected trajectory (coordinates of the trajectory centerline) is set to v1, and the u coordinate of the trajectory corresponding to v1 is set to u1. Using u1,v1 as the base, a predetermined range in the u and v directions is determined as the long distance ROI. If the ROI region to be extracted is for a short distance, the v coordinate of the bottom edge point on the image in the coordinates of the detected trajectory (coordinates of the trajectory centerline) is set to v2, and the u coordinate of the trajectory corresponding to v2 is set to u2. Using u2,v2 as the base, a predetermined range in the u and v directions is determined as the short distance ROI.

[0033] Although an example has been described in which an ROI is determined within a fixed range (number of pixels) from the reference u and v coordinates, the range may be dynamically changed. For example, the ROI required for object detection may depend on the required detection accuracy for the object detection model and the clarity of the input image. The clarity of the input image may change depending on the surrounding environment, such as weather or time of day. The ROI may be dynamically changed (enlarged or reduced) depending on these conditions. Furthermore, the v coordinate serving as the reference for the ROI may be defined by a method other than that based on detected trajectory information. For example, the v coordinate may be fixed for short, medium, and long distances. In the present invention, the u coordinate serving as the reference for the ROI may be defined based on trajectory information, and the method for defining the reference v coordinate does not matter.

[0034] Step 307: Based on the determined ROI, the input image from the external sensor is cut out to generate an ROI cut-out image.

[0035] Step 308: It is determined whether ROI cutout images have been generated for all of the target ROI regions. If ROI cutout images have been generated for all ROI regions, proceed to step 309. If there are ROI regions for which ROI cutout images have not been generated, proceed to step 302.

[0036] Step 309: The generated ROI cut-out image is sent to the object detection unit 124.

[0037] In this embodiment, the trajectory information from one cycle before the current calculation cycle is used to generate a short-distance ROI cutout image, so there is no need to wait for the trajectory information from the trajectory detection unit 121 in the current calculation cycle, and the object detection process by image processing for short distances can start at the same time as the current calculation cycle starts. Therefore, it is possible to shorten the processing time for the object detection process by image processing.

[0038] In this embodiment, the ROI regions are described as two regions, one for long distances and one for short distances. However, as mentioned above, the number of ROI regions is not limited. For example, three ROI regions may be used, one for long distances, one for medium distances, and one for short distances. In this case, it is preferable to use the trajectory information from the trajectory information recording unit 122 one cycle before the current calculation cycle as the trajectory information used in the ROI determination process for short distances and medium distances, in which the deviation in the trajectory position on the image between the current calculation cycle and the calculation cycle one cycle before the current calculation cycle is considered to be relatively small. In this way, object detection processing by image processing for short distances and medium distances can be performed in parallel with the trajectory detection processing by the trajectory detection unit 121 in the current calculation cycle. This makes it possible to shorten the processing time for object detection processing by image processing, even when the number of ROI clipped images to be processed increases. [Example]

[0039] The processing of the ROI processing unit 123 according to the second embodiment of the present invention will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the processing executed by the ROI processing unit 123. In the first embodiment, the size of the ROI region is used as a criterion for determining the trajectory information to be used for determining the ROI, but in this embodiment, an example will be described in which the trajectory information to be used for determining the ROI is selected according to the trajectory shape.

[0040] Step 601: An image is received as sensor information from the external sensor 110. The input image is, for example, an image as shown in Figure 2. The process proceeds to step 602.

[0041] Step 602: The track shape ahead of the train 100 is determined. The method for determining the track shape will be described later. The process proceeds to step 603.

[0042] Step 603: If the track shape ahead of the train 100 is straight, the process proceeds to step 604, and if the track shape ahead of the train 100 is curved, the process proceeds to step 605.

[0043] Step 604: When the track shape ahead of the train is straight, the amount of change in track coordinates on the image over time (especially in the u-coordinate direction) is small. Therefore, the difference between the ROI determined based on the track position on the image in the current calculation cycle and the ROI determined based on the track position on the image in the calculation cycle one cycle before the current calculation cycle is also small. Therefore, the ROI cropped image, which is cropped based on the ROI determined based on the track position on the image in the calculation cycle one cycle before the current calculation cycle, includes the area where object detection is desired (near the track). In straight sections, the processing time can be further reduced by using track information from the calculation cycle one cycle before the current calculation cycle to determine the ROIs for short distances and long distances. Using the reduced time to generate new ROI cropped images and perform image processing can also improve object detection performance through image processing. In step 604, track information from the calculation cycle one cycle before the current calculation cycle is obtained from the track information recording unit 122. Proceed to step 607.

[0044] Step 605: When the track shape ahead of the train is curved, the amount of change in track coordinates on the image over time (especially in the u-coordinate direction) is large even for short distances. Therefore, there is a large difference between the ROI determined based on the track position on the image in the current calculation cycle and the ROI determined based on the track position on the image in the calculation cycle one cycle before the current calculation cycle. Therefore, even for short distances, the ROI cropped based on the ROI determined based on the track position on the image in the calculation cycle one cycle before the current calculation cycle may not include the area where object detection is desired (near the track). Therefore, in curved sections, by using track information for the current calculation cycle in the ROI determination process for short and medium distances, it is possible to ensure that the area where object detection is desired (near the track) is included in the ROI cropped image. In step 605, track information for the current calculation cycle is obtained from the track detection unit 121. Proceed to step 606.

[0045] Step 606: If the track shape ahead of the train is curved, the track information for the current calculation cycle is used, which increases the processing time for object detection processing using image processing. In curved sections, the influence of trackside structures often makes it impossible to see the track area in the image in the first place. Therefore, by skipping the long-distance object detection processing and using the track information for the current calculation cycle for the short- and medium-distance ROI determination processing, it is possible to shorten the processing time for object detection processing using image processing while maintaining object detection performance in curved sections. In step 606, an ROI region for object detection processing is selected. Specifically, it is decided to skip the long-distance ROI determination processing, where visibility is thought to be poor in curved sections, and to perform only the short- and medium-distance ROI determination processing. Proceed to step 607.

[0046] Step 607: The ROI is determined based on the ROI region information that was selected as the processing target in step 606 and the trajectory information acquired in steps 604 and 605. The process proceeds to step 608.

[0047] Step 608: Based on the determined ROI, the input image from the external sensor is cut out to generate an ROI cut-out image. The process proceeds to step 609.

[0048] Step 609: It is determined whether ROI cutout images have been generated for all of the target ROI regions. If ROI cutout images have been generated for all ROI regions, proceed to step 610. If there are ROI regions for which ROI cutout images have not been generated, proceed to step 602.

[0049] Step 610: The generated ROI cut-out image is sent to the object detection unit 124.

[0050] Next, the method of determining whether the section is a straight line section or a curved line section in step 602 will be described with reference to FIG.

[0051] The left diagram in Figure 7(a) shows an example of a straight section. In the left diagram in Figure 7(a), the actual trajectory is represented by a dotted line 700, the trajectory detection result detected by image processing is represented by a trajectory detection result 710, and the trajectory center position 720 of the trajectory coordinates (u) on the left and right of the trajectory detection result 710. In this case, the relationship between the ordinate (v) of the trajectory detection result 710 on the image and the trajectory center position 720 is expressed by a linear function 730 as shown in the right diagram in Figure 7(a). In such a case, it is determined to be a straight section.

[0052] Next, an example of a section where a curved section exists will be described. In the left diagram of Figure 7(b), the actual trajectory is represented by a dotted line 701, the trajectory detection result detected by image processing is represented by a trajectory detection result 711, and the trajectory center position 721 of the trajectory coordinates (u) on the left and right of the trajectory detection result 711. In this case, the relationship between the ordinate (v) of the trajectory detection result 711 on the image and the trajectory center position 721 is expressed by a curve (a quadratic or higher degree function) 741, as shown in the right diagram of Figure 7(b).

[0053] If the relationship between the ordinate (v) of the track detection result 711 and the track center position 721 is a curve (a function of second degree or higher) 741 as shown in the right diagram of FIG. 7(b), it is determined to be a curved section.

[0054] In addition to the relationship between the ordinate (v) of the trajectory detection result 711 and the trajectory center position 721 being a quadratic or higher function, if the difference between the farthest trajectory center position 751, which is the trajectory center position of the point in the trajectory detection result 711 with the smallest ordinate (v) value (the farthest part of the trajectory detection result), and the straight section trajectory center position 761, which is one of the track center positions 730 in the straight section and has the same v coordinate as the farthest trajectory center position 751, is equal to or greater than a predetermined threshold, the section may be recognized as a curved section. This makes it possible to determine whether the section is straight or curved with a simpler calculation.

[0055] In this embodiment, the processing time can be reduced by using the trajectory information from the calculation cycle one cycle before the current calculation cycle to determine the ROIs for short distances and long distances in the straight section. The reduced time is used to generate a new ROI cutout image and perform image processing, which also makes it possible to improve the object detection performance through image processing.

[0056] In addition, by skipping the long-distance object detection process in curved sections and using the trajectory information in the current calculation cycle for the ROI determination process for short and medium distances, it is possible to shorten the processing time for object detection process using image processing while maintaining object detection performance in curved sections.

[0057] According to the above-described embodiment, in an obstacle detection system installed in a rail transportation system, it is possible to select an appropriate ROI to improve obstacle detection performance, while reducing the time required for obstacle detection processing by image processing.

[0058] The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and are not necessarily limited to those including all of the described configurations. Furthermore, part of the configuration of one embodiment can be replaced with the configuration of another embodiment, or the configuration of another embodiment can be added to the configuration of one embodiment. Furthermore, part of the configuration of each embodiment can be added, deleted, or replaced with other configurations. Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be implemented in hardware, in part or in whole, by designing, for example, an integrated circuit. Furthermore, the above-described configurations, functions, etc. may be implemented in software, by a processor interpreting and executing a program that realizes each function. Information such as programs, tables, and files that realize each function can be stored in memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

[0059] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present invention.

[0060] The following are examples of the present invention, but the present invention is not limited to these examples.

[0061] (Aspect 1) a track detection unit that detects track information from data from an external sensor for recognizing the environment around the train; an ROI processing unit that extracts a region of interest to be subjected to object detection processing from the trajectory information; an object detection unit that detects an object; An obstacle detection system comprising: a trajectory information recording unit that records the trajectory information detected by the trajectory detection unit, An obstacle detection system characterized in that the ROI processing unit changes the track information used to extract the region of interest based on information on either the size of the region of interest or the shape of the track ahead of the train.

[0062] (Aspect 2) 2. The obstacle detection system of claim 1, wherein the ROI processing unit extracts the region of interest from the trajectory information from the trajectory information recording unit when the region of interest is large, and extracts the region of interest from the trajectory information from the trajectory detection unit when the region of interest is small.

[0063] (Aspect 3) 2. The obstacle detection system of claim 1, The ROI processing unit extracts a region of interest from the track information from the track information recording unit when the track shape ahead of the train is straight, and extracts a region of interest from the track information from the track detection unit when the track shape ahead of the train is curved.

[0064] (Aspect 4) 4. The obstacle detection system according to claim 3, An obstacle detection system characterized in that the ROI processing unit does not perform processing to extract the smallest region of interest when the track shape ahead of the train is curved.

[0065] (Aspect 5) The obstacle detection system according to any one of aspects 1 to 4, Obstacle detection system with external sensors.

[0066] (Aspect 6) An obstacle detection system according to any one of aspects 1 to 5, An obstacle detection system in which the external sensor is a camera.

[0067] (Aspect 7) The track detection unit detects track information from data from external sensors that recognize the environment around the train, The ROI processing unit extracts the region of interest to be processed for object detection from the trajectory information, The object detection unit detects an object. 1. An obstacle detection method, comprising: a trajectory information recording unit that records the trajectory information detected by the trajectory detection unit; An obstacle detection method characterized in that the ROI processing unit changes the track information used to extract the region of interest based on information on either the size of the region of interest or the shape of the track ahead of the train.

[0068] (Aspect 8) 8. The obstacle detection method according to claim 7, further comprising: an obstacle detection method, characterized in that the ROI processing unit extracts the region of interest from the trajectory information from the trajectory information recording unit when the region of interest is large, and extracts the region of interest from the trajectory information from the trajectory detection unit when the region of interest is small.

[0069] (Aspect 9) 8. The obstacle detection method according to claim 7, further comprising: an ROI processing unit that extracts a region of interest from track information from the track information recording unit when the track shape ahead of the train is straight, and extracts a region of interest from track information from the track detection unit when the track shape ahead of the train is curved.

[0070] (Aspect 10) 10. The obstacle detection method according to aspect 9, An obstacle detection method characterized in that the ROI processing unit does not perform processing to extract the smallest region of interest when the track shape ahead of the train is curved. [Explanation of symbols]

[0071] 100 trains 110 External Sensor 120 Obstacle detection device in embodiment 1 121 Trajectory detection unit in the first embodiment 122 Orbit information recording unit in Example 1 123 ROI processing unit in the first embodiment 124 Object detection unit in the first embodiment

Claims

1. a track detection unit that detects track information from data from an external sensor for recognizing the environment around the train; an ROI processing unit that extracts a region of interest to be subjected to object detection processing from the trajectory information; an object detection unit that detects an object; An obstacle detection system comprising: a trajectory information recording unit that records the trajectory information detected by the trajectory detection unit, An obstacle detection system characterized in that the ROI processing unit changes the track information used to extract the region of interest based on information on either the size of the region of interest or the shape of the track ahead of the train.

2. 2. The obstacle detection system according to claim 1, an obstacle detection system characterized in that the ROI processing unit extracts the region of interest from the trajectory information from the trajectory information recording unit when the region of interest is large, and extracts the region of interest from the trajectory information from the trajectory detection unit when the region of interest is small.

3. 2. The obstacle detection system according to claim 1, The obstacle detection system is characterized in that the ROI processing unit extracts a region of interest from track information from the track information recording unit when the track shape ahead of the train is straight, and extracts a region of interest from track information from the track detection unit when the track shape ahead of the train is curved.

4. 4. The obstacle detection system according to claim 3, An obstacle detection system characterized in that the ROI processing unit does not perform processing to extract the smallest region of interest when the track shape ahead of the train is curved.

5. 2. The obstacle detection system according to claim 1, Obstacle detection system with external sensors.

6. 2. The obstacle detection system according to claim 1, An obstacle detection system in which the external sensor is a camera.

7. The track detection unit detects track information from data from external sensors that recognize the environment around the train, An ROI processing unit extracts a region of interest to be subjected to object detection processing from the trajectory information, The object detection unit detects an object. An obstacle detection method, comprising: a trajectory information recording unit that records the trajectory information detected by the trajectory detection unit; An obstacle detection method characterized in that the ROI processing unit changes the track information used to extract the region of interest based on information on either the size of the region of interest or the shape of the track ahead of the train.

8. 8. The obstacle detection method according to claim 7, an ROI processing unit that extracts a region of interest from trajectory information from the trajectory information recording unit when the region of interest is large, and extracts a region of interest from trajectory information from the trajectory detection unit when the region of interest is small.

9. 8. The obstacle detection method according to claim 7, an ROI processing unit that extracts a region of interest from track information from the track information recording unit when the track shape ahead of the train is straight, and extracts a region of interest from track information from the track detection unit when the track shape ahead of the train is curved.

10. 10. The obstacle detection method according to claim 9, An obstacle detection method, characterized in that the ROI processing unit does not perform processing to extract the smallest region of interest when the track shape ahead of the train is curved.

Images