Active safety braking method for battery car in rail-mounted area of urban rail transit

Abstract

The invention relates to the technical field of rail traffic construction rail-mounted area safety, and discloses an active safety braking method for a battery car in an urban rail traffic rail-mounted area, and automatic braking is achieved by installing an active safety device comprising a binocular camera, an AI calculation unit and the like at the front end of the battery car. The device collects front images, depth information is calculated by combining a stereo matching algorithm with camera parameters after the front images are processed, the AI unit recognizes personnel (including standing postures, sitting postures and prone postures) and obstacles by combining the depth information with a target detection algorithm, and analyzes the positions and movement tracks of the personnel and the obstacles and whether the personnel and the obstacles are on a driving path or not; and the AI unit sends a brake instruction and triggers an alarm, so that the battery car decelerates to stop. Manual intervention is not needed, the device can be conveniently installed, the device is compatible with various vehicle types, an original vehicle system is not affected, cost is reduced, and construction safety is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of safety technology for rail traffic construction rail area, and particularly relates to a method for actively braking a trolley car in a rail area of urban rail transit. BACKGROUND

[0002] With the rapid advancement of urban rail transit construction, the rail area, as the core area of construction, undertakes key operations such as track laying, equipment installation, and material transportation, and its safety management is directly related to the construction progress and personnel safety. The rail transport vehicle (also known as the rail trolley car) is the core tool for transporting materials and equipment in the rail area, and its operational safety is particularly important. However, the current safety control of the rail transport vehicle in the rail area still has significant shortcomings, and urgent technical breakthroughs are needed.

[0003] The construction environment in the rail area has the characteristics of strong closure, complex operation scene, and large personnel mobility: on the one hand, the construction area is mostly a tunnel or viaduct section, with relatively narrow space and variable light conditions (such as uneven lighting in the tunnel and external natural light interference); on the other hand, operating personnel need to perform operations such as track laying, welding, and equipment debugging around the track, often in various postures such as standing, sitting, and lying, and obstacles such as various construction materials and temporary supports are commonly placed randomly. In this environment, the driving of the rail transport vehicle faces a high risk of collision, and once a collision occurs, it may not only cause personnel casualties but also cause the project to stall, resulting in significant economic losses.

[0004] The safety control of the existing rail transport vehicle mainly relies on the "manual observation + manual braking" mode, and its safety completely depends on the concentration and reaction speed of the driver. However, this mode has the following defects that are difficult to overcome: first, the driver is prone to fatigue after a long period of work, and it is difficult to quickly identify personnel or obstacles in front when the attention is dispersed; second, the line of sight in the rail area is easily affected by factors such as curves and equipment obstructions, and the reaction time of the driver to unexpected situations is often insufficient; third, the construction personnel may ignore the approaching vehicle due to their focus on the work at hand, further amplifying the collision risk. According to industry statistics, about 60% of safety accidents in the rail area are related to the failure of the rail transport vehicle to avoid in time, highlighting the limitations of the manual dependence mode.

[0005] To improve this situation, some projects have attempted to introduce personnel positioning systems. By deploying base stations and assigning positioning chips to workers, they enable dynamic tracking of vehicles and personnel. However, this solution faces multiple challenges in practical application: First, the high cost of deploying base stations and positioning chips on a large scale requires substantial investment, and chips are prone to wear and loss during construction, leading to duplicate procurement. Second, management is difficult, as the high mobility of construction workers makes it difficult to ensure proper chip wear, leading to positioning data deviations. Third, system compatibility is poor, as different brands of positioning equipment struggle to coordinate with vehicle control systems, resulting in delayed early warning responses. Consequently, this solution has been difficult to widely adopt within the industry.

[0006] In summary, current safety control methods for rail transport vehicles are either manually reliant and risky, or costly and impractical. The industry urgently needs a low-cost, easy-to-install, and human-independent active safety technology that can automatically identify and brake for people and obstacles ahead, fundamentally improving construction safety in tracked areas. Summary of the Invention

[0007] The present invention aims to provide an active safety braking method for electric vehicles in urban rail transit track areas, so as to solve the problems that existing rail transport vehicles rely on the driver to detect obstacles or personnel, which may easily cause safety hazards due to negligence, and the traditional personnel positioning system is costly and difficult to operate.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions: An active safety braking method for a battery vehicle in an urban rail transit track area, comprising: An active safety device is installed at the front end of the battery vehicle. The active safety device includes a binocular camera, a fill light, an image processing unit, an AI computing unit, a DC-DC converter, a communication interface, an alarm horn, and a portable fixing bracket. The portable fixing bracket is detachably connected to the front end of the battery vehicle. The binocular camera synchronously captures the left and right images of the scene in front of the battery vehicle. After being processed by the image processing unit, it is transmitted to the AI ​​computing unit. The stereo matching algorithm finds the common feature points in the left and right images, and uses the pixel offset, camera baseline distance and focal length to calculate the depth information of each point in the scene. The AI ​​computing unit combines depth information with target detection algorithms to identify and determine targets, including people and obstacles, analyze the target's position and movement trajectory, and determine whether the target is in the battery vehicle's driving path; When it is detected that the target is within the preset safety distance and on the driving path, the AI ​​computing unit sends a braking command to the electric vehicle controller and triggers the alarm horn to sound an alarm; After receiving the brake command, the battery vehicle controller controls the brake actuator to slow down the battery vehicle until it stops completely.

[0009] The principle and advantages of the present scheme are: in actual application, the active safety device is connected through a detachable connection, making the active safety device compatible with different models of track battery cars, and after installation, it does not affect the normal operation of the original vehicle circuit and control system. Through the binocular camera, image processing unit and AI calculation unit, depth information calculation is realized, and the driver does not need to estimate the distance of the front target through vision, but automatically calculates the depth information through the stereo matching algorithm and the principle of triangulation, avoiding distance misjudgment caused by human negligence, visual error or fatigue, and ensuring more accurate and reliable distance detection of personnel and obstacles. Compared with deploying a personnel positioning system, the binocular camera combined with the algorithm does not require additional positioning equipment, and can calculate depth information through image data only, greatly reducing the construction cost and maintenance difficulty of track area safety management. Combined with the pre-processing of the light supplement lamp and the image processing unit, even in complex environments such as low light and dust, the depth information of the front target can still be accurately identified through feature point matching and distance calculation, and various postures of personnel and various obstacles can be covered, avoiding the risk of missing detection. The AI calculation unit combines depth information and target detection algorithm to realize target recognition, analysis and path judgment, avoiding misjudgment of non-target objects (such as shadows and debris) as risk sources, and accurately identifying personnel in different postures to provide differentiated basis for subsequent risk assessment. Analyzing the position and motion trajectory of the target, the dynamic change of the target is mastered in real time, and the trend of the distance between the target and the battery car is judged combined with the depth information. Compared with only detecting static targets, this dynamic analysis can predict potential collision risks in advance, providing more sufficient reaction time for brake decision-making. By judging whether the target is on the driving path, only the target that really constitutes a threat triggers the brake, avoiding false braking for targets outside the path and reducing interference with construction efficiency.

[0010] Preferably, as an improvement, the power supply of the active safety device comes from the battery of the battery car, and the voltage is adjusted to 12V DC through a DC-DC converter, independent of the original vehicle circuit system.

[0011] Technical effect: independent voltage conversion for safe power supply, adaptation to different models of battery car battery voltage, avoiding electromagnetic interference or load pressure on the original vehicle circuit.

[0012] Preferably, as an improvement, the communication interface is connected with the battery car controller through CAN bus or adaptive protocol to receive brake instructions and drive the brake actuator to complete automatic braking.

[0013] Technical effect: ensure real-time transmission of brake instructions and normal operation of the original vehicle control system.

[0014] Preferably, as an improvement, the distance between the binocular cameras remains fixed.

[0015] Technical effects: Ensure the stability of the baseline distance parameter, provide accurate basic data for the calculation of depth information based on the principle of triangulation, and reduce the distance calculation error caused by the fluctuation of hardware parameters.

[0016] Preferably, as an improvement, the finding of the same feature points in the left and right images by the stereo matching algorithm comprises: Extracting feature points from left and right images respectively to generate feature descriptors containing position, scale and direction; Conducting similarity comparison of feature descriptors of left and right images by K-Nearest Neighbor matching algorithm, and screening feature point pairs meeting matching degree as same feature points; Using Random Sample Consensus algorithm to eliminate mis-matched feature points and retain effective feature point pairs meeting geometric constraints.

[0017] Technical effects: Through feature point extraction, similarity comparison and mis-matching elimination, the accuracy of feature point matching of left and right images is improved, and reliable pixel displacement data is provided for subsequent depth information calculation.

[0018] Preferably, as an improvement, the calculation of depth information using pixel displacement, camera baseline distance and focal length comprises: Calculating the horizontal pixel displacement of the same feature points in the left and right images; Calculating the scene depth information corresponding to the feature points based on the principle of triangulation; Smoothing and filtering the calculated depth information to eliminate isolated noise points.

[0019] Technical effects: Improve the stability of distance data.

[0020] Preferably, as an improvement, the image processing unit further performs dynamic exposure adjustment processing on the image, comprising: Real-time detection of the average brightness value of the image, when the brightness value is lower than the threshold value, triggering the light supplement lamp to open and adjusting the camera exposure time; When the brightness value is higher than the threshold value, shorten the exposure time and reduce the power of the light supplement lamp.

[0021] Technical effects: Ensure the clarity of image feature point extraction.

[0022] Preferably, as an improvement, the AI calculation unit combines depth information and target detection algorithm to identify and determine the target, comprising: Based on the pre-trained deep learning target detection model, feature extraction is performed on the pre-processed image to generate candidate regions containing target bounding boxes; Extracting the depth information corresponding to the candidate regions, combining the correlation between the pixel size and the depth of the target, and screening the regions whose actual physical size meet the characteristics of personnel or obstacles; The target type is determined by performing feature matching on the candidate region through a preset target feature library.

[0023] Technical effects: The depth information is combined with the feature library matching to accurately distinguish between personnel (including different postures) and obstacles, reduce misjudgment (such as avoiding misjudgment of a person in a sitting posture as an obstacle), and improve the pertinence and reliability of target identification.

[0024] Preferably, as an improvement, the determination of whether the target is on the path of the electric vehicle includes: When the lateral offset distance of the target is within the path range, a person in a sitting or lying posture and an obstacle are directly determined to be on the path, and a person in a standing posture and a moving obstacle are determined to be on the path when the lateral offset amount will enter the path range within a future preset time length.

[0025] Technical effects: Different targets are differentiated according to their risk avoidance ability and motion characteristics, ensuring timely response to high-risk targets (such as a person in a lying posture and a fixed obstacle), and early warning through prediction of the trajectory of a moving target, balancing safety protection and construction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A structure diagram of a city rail transit track area electric vehicle active safety brake method according to an embodiment of the present application; Figure 2 A structure diagram of an active safety device according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will be further described in detail through specific embodiments: The embodiment is basically as shown in the accompanying Figure 1 , and Figure 2 : A city rail transit track area electric vehicle active safety brake method includes the following steps: An active safety device is installed at the front end of the electric vehicle, and the active safety device includes a binocular camera, a fill light, an image processing unit, an AI computing unit, a DC-DC converter, a communication interface, an alarm horn, and a portable fixing bracket. The portable fixing bracket is detachably connected to the front of the electric vehicle by a magnet or a bolt, and the distance between the binocular cameras is fixed.

[0028] Specifically, the portable fixing support in the embodiment is a hollow metal frame structure, which is provided with a binocular camera mounting position at the front end face, the spacing is fixed at 15-30 cm, the top is integrated with a light supplement lamp, the included angle between the light supplement lamp and the camera optical axis is ≤10°, the inner cavity accommodates an image processing unit and an AI computing unit, and heat dissipation is realized by adhering the heat dissipation fins to the inner wall of the support. The DC-DC converter and the communication interface are integrated in a separate sealed box at the rear end of the portable fixing support, and the alarm speaker is connected to the side of the portable fixing support through a flexible cable. The components are connected through internal wires, among which the binocular camera and the image processing unit transmit image data through an LVDS interface, the image processing unit and the AI computing unit interact through a PCIe bus, and the AI computing unit and the communication interface and the alarm speaker are connected through a UART serial port and a GPIO interface respectively. The positive electrode of the storage battery is connected through the input end of the DC-DC converter, and the output end is independently powered for the binocular camera, the light supplement lamp, the image processing unit, and the AI computing unit. The signal port of the communication interface and the controller of the electric bicycle is connected through a CAN bus or an adaptive protocol, supporting bidirectional data transmission to facilitate brake command sending and vehicle state feedback. The magnet of the portable fixing support is a neodymium iron boron strong magnet (attractive force ≥50 kg) embedded in the groove at the bottom of the support; the bolt connection hole is a general M8 specification, which is suitable for different vehicle mounting hole distances, and the two fixing methods can be used alone or in combination.

[0029] The left and right images of the scene in front of the electric bicycle are synchronously collected by the binocular camera, processed by the image processing unit, and transmitted to the AI computing unit. The same feature points in the left and right images are found through a stereo matching algorithm.

[0030] Specifically: the left and right images collected by the binocular camera are denoised, random noise in the images is removed through a Gaussian filter algorithm, and image clarity is improved; image grayscale and contrast enhancement are realized, color images are converted to grayscale images to reduce data volume, and histogram equalization is used to adjust the light and dark contrast of the images, so that the edges and details of objects in the scene are more easily identified; the processed images are standardized in format, resolution and pixel arrangement, and are subsequently transmitted to the AI computing unit for stereo matching operation.

[0031] Finding the same feature points in the left and right images through a stereo matching algorithm includes: extracting feature points from the left and right images respectively to generate feature descriptors containing position, size and direction; performing similarity comparison on the feature descriptors of the left and right images through a K nearest neighbor matching algorithm, and selecting feature point pairs that meet the matching degree as the same feature points; using a random sample consensus algorithm to remove false matching feature points and retain valid feature point pairs that meet geometric constraints.

[0032] In this embodiment, the ORB or SIFT algorithm is used to scan the preprocessed left and right images to locate areas with significant grayscale changes, edge turning points or texture differences, and the key pixels in these areas are determined as feature points; for each feature point, a local neighborhood window is constructed with it as the center, and the gradient direction, scale change and grayscale distribution law of the pixels in the window are calculated to generate a feature descriptor containing the two-dimensional coordinates (position) of the feature point, the neighborhood size (scale), and the main gradient direction (direction), to ensure that the descriptor can uniquely characterize the local image characteristics of the feature point. The process of screening feature point pairs that meet the matching degree through the K-nearest neighbor matching algorithm is as follows: taking the descriptor of a feature point in the left image as the benchmark, retrieve K (K ≥ 2) most similar candidate descriptors from all feature point descriptors in the right image; calculate the similarity score between the benchmark descriptor and the first candidate descriptor, as well as the similarity score with the second-best candidate descriptor. If the ratio of the two is less than a preset threshold (such as 0.75), the feature point corresponding to the first candidate descriptor and the benchmark feature point are determined to be a matching pair; traverse all feature points in the left image, repeat the above process, and screen out all feature point pairs that meet the similarity conditions as the initial matching results. The specific method of using the random sampling consistency algorithm to eliminate mismatched feature points is as follows: based on the inherent parameters of the binocular camera such as the baseline distance and focal length, determine the geometric constraints (such as epipolar constraints) that the feature points in the left and right images satisfy; randomly extract a small number of samples from the initial matched feature point pairs, and fit a mathematical model that conforms to the geometric constraints; count the number of inliers that conform to the model in all initial matching point pairs, and repeat the sampling and fitting process until the model with the largest number of inliers is found; eliminate outliers outside the model (mismatched point pairs), and retain the inliers as valid feature point pairs that conform to the geometric constraints.

[0033] The image processing unit also performs dynamic exposure adjustment on the image, including: real-time detection of the average brightness value of the image. When the brightness value is lower than a threshold, such as lower than 50 lux, the fill light is triggered to turn on and the camera exposure time is adjusted to 50-100ms; when the brightness value is higher than the threshold, such as higher than 500 lux, the exposure time is shortened to 1-10ms and the power of the fill light is reduced.

[0034] The time difference is used to calculate the pixel offset of the same feature point in the left and right images. According to the principle of triangulation, the pixel offset, camera baseline distance and focal length are used to calculate the depth information of each point in the scene, that is, the distance between the electric vehicle and each point.

[0035] Specifically, the horizontal pixel offset of the same feature points in the left and right images is calculated; the binocular camera ensures that the time difference of image acquisition is less than or equal to 1 ms through a hardware synchronization trigger mechanism, and the difference corresponds to the imaging difference of the same physical scene under left and right viewing angles; for the same feature points determined by the stereo matching algorithm in the left and right images, the horizontal coordinates of the feature points in the left image and the horizontal coordinates of the feature points in the right image are obtained through a pixel coordinate positioning tool; the pixel offset is calculated, and the formula is: d = x1-x2, d is the parallax, x1 and x2 are the horizontal coordinates in the left image and the horizontal coordinates in the right image respectively, and the pixel offset is directly quantified by the horizontal coordinate difference of the feature points in the left and right images, reflecting the imaging offset caused by the distance between the binocular cameras. The scene depth information corresponding to the feature points is calculated based on the principle of triangulation; the intrinsic parameters of the binocular camera, including the baseline distance and the focal length, are calibrated in advance and stored in the parameter database of the AI computing unit, and based on the principle of triangulation, Z=(Bxf) / d, Z is the actual distance between the target and the camera, B is the baseline distance, i.e. the distance between the optical centers of the left and right cameras, f is the focal length, and d is the parallax. The scene depth corresponding to the feature points is calculated using the formula Z. The calculated depth information is subjected to smoothing filtering processing to eliminate isolated noise points.

[0036] The AI computing unit combines the depth information and the target detection algorithm to identify and determine the target, determine the target including personnel and obstacles, analyze and determine the position, type and motion trajectory of the target, and determine whether the determined target is on the electric vehicle driving path.

[0037] Specifically, a pre-trained deep learning target detection model is called, and in this embodiment, a YOLO model adapted to the track area scene is used to perform pixel-level scanning on the images collected by the binocular camera to generate candidate regions containing target bounding boxes; the depth information corresponding to the candidate regions is extracted, and the regions whose actual physical size meets the characteristics of personnel or obstacles are screened out in combination with the correlation between the pixel size of the target and the depth; feature comparison is performed on the candidate regions: personnel features include the contour ratio and joint point distribution of the head, torso and limbs, and obstacle features include regular geometric shapes (such as cubes and cylinders) or non-biological textures (such as metal and concrete surfaces), and the target type is distinguished through feature matching.

[0038] A two-dimensional coordinate system is established with the center of the electric vehicle head as the origin, the longitudinal direction as the driving direction, and the transverse direction as the vertical driving direction. The depth information of the target is converted into a longitudinal distance Z, and the transverse offset distance X is calculated by combining the transverse pixel offset of the target in the image and the focal length, to obtain the target coordinates (X, Z). When the target is determined to be a person, the standing posture, sitting posture, and lying posture are identified through a skeleton key point detection algorithm. For example, in the standing posture, the head, torso, and lower limbs are vertically distributed, and the knee joint angle is ≥160°. In the sitting posture, the angle between the torso and the lower limbs is 30°-90°, the hips are in contact with the ground, and in the lying posture, the angle between the torso and the ground is ≤30°, and the whole body profile is horizontally distributed. When the target is determined to be an obstacle, it is subdivided into fixed obstacles and moving obstacles according to shape characteristics. Fixed obstacles, such as trackside tool piles, have regular edges and no motion trajectory, while moving obstacles, such as out-of-control material trucks, have continuous frame position changes.

[0039] The same target in 10 consecutive frames of images is tracked, the coordinate change sequence is recorded, the motion direction vector is obtained through a linear fitting algorithm, and the moving speed is calculated based on the frame interval time.

[0040] The driving path range is set based on the width of the electric vehicle body. For example, the body width is 1.5-2 meters, the longitudinal direction is the detection direction of the camera, and the transverse direction is the area within ±0.75-1 meters of the center of the vehicle head to set the driving path range.

[0041] For a person target: if the transverse offset distance (X) is within the path range and the posture is sitting or lying, the autonomous avoidance ability is weak, and it is directly determined to be on the path; if a standing person will enter the path range within a predetermined time in the future, it is also determined to be at risk. For an obstacle target: if the transverse offset distance (X) is within the path range and the depth (Z) ≤ a predetermined safety distance (such as 30 meters), it is determined to be on the path; for moving obstacles, combined with their motion trajectory prediction, if they will enter the path range within a predetermined time in the future, an early warning will be triggered.

[0042] When a person or obstacle is detected within a predetermined safety distance and on the driving path, the AI computing unit sends a brake command to the electric vehicle controller, and simultaneously triggers the alarm horn to issue a warning; After receiving the brake command, the electric vehicle controller controls the brake actuator to slow down the electric vehicle until it completely stops.

[0043] The above-mentioned are only embodiments of the present application, and common technical solutions and / or common knowledge of the scheme are not described in detail. It should be pointed out that, for those skilled in the art, without departing from the technical solutions of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application, and these will not affect the effect and practicality of the patent. The protection scope claimed in the present application should be subject to the content of its claims, and the specific implementation mode and the like recorded in the specification can be used to explain the content of the claims.

Claims

1. An active safety braking method for battery vehicles in urban rail transit track areas, characterized in that: include: An active safety device is installed at the front end of the battery vehicle. The active safety device includes a binocular camera, a fill light, an image processing unit, an AI computing unit, a DC-DC converter, a communication interface, an alarm horn, and a portable fixing bracket. The portable fixing bracket is detachably connected to the front end of the battery vehicle. The binocular camera synchronously captures the left and right images of the scene in front of the battery vehicle. After being processed by the image processing unit, it is transmitted to the AI ​​computing unit. The stereo matching algorithm finds the common feature points in the left and right images, and uses the pixel offset, camera baseline distance and focal length to calculate the depth information of each point in the scene. The AI ​​computing unit combines depth information with target detection algorithms to identify and determine targets, including people and obstacles, analyze the target's position and movement trajectory, and determine whether the target is in the battery vehicle's driving path; When it is detected that the target is within the preset safety distance and on the driving path, the AI ​​computing unit sends a braking command to the electric vehicle controller and triggers the alarm horn to sound an alarm; After receiving the brake command, the battery vehicle controller controls the brake actuator to slow down the battery vehicle until it stops completely.

2. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The active safety device is powered by the battery of the electric vehicle, and the voltage is adjusted to 12V direct current through a DC-DC converter, and operates independently of the original vehicle circuit system.

3. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The communication interface is connected to the battery vehicle controller via a CAN bus or an adaptation protocol to receive a brake instruction and drive a brake actuator to complete automatic braking.

4. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The distance between the binocular cameras remains fixed.

5. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The method of finding the same feature points in the left and right images by using a stereo matching algorithm includes: Extract feature points from the left and right images respectively and generate feature descriptors including position, scale and direction; The feature descriptors of the left and right images are compared for similarity using the K-nearest neighbor matching algorithm, and feature point pairs that meet the matching degree are selected as the same feature points; A random sampling consensus algorithm is used to eliminate mismatched feature points and retain valid feature point pairs that meet geometric constraints.

6. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: Calculating depth information using pixel offset, camera baseline distance, and focal length includes: Calculate the horizontal pixel offset of the same feature point in the left and right images; Calculate the scene depth information corresponding to the feature points based on the triangulation principle; The calculated depth information is smoothed and filtered to eliminate isolated noise points.

7. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The image processing unit further performs dynamic exposure adjustment processing on the image, including: Detect the average brightness value of the image in real time. When the brightness value is lower than the threshold, the fill light is triggered to turn on and the camera exposure time is adjusted; When the brightness value is higher than the threshold, the exposure time is shortened and the fill light power is reduced.

8. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The AI ​​computing unit combines depth information with the target detection algorithm to identify and determine the target, including: Perform feature extraction on the pre-processed image based on the pre-trained deep learning target detection model to generate candidate regions containing the target bounding box; Extract the depth information corresponding to the candidate area, combine the correlation between the pixel size and depth of the target, and filter out the area whose actual physical size meets the characteristics of people or obstacles; The candidate area is matched with features through the preset target feature library to determine the target type.

9. The active safety braking method for battery vehicles in urban rail transit track areas according to claim 1 is characterized in that: The determination of whether the target is on the battery vehicle's travel path includes: When the lateral offset distance of the target is determined to be within the path range, people and obstacles in a sitting or lying position are directly determined to be on the path. People and moving obstacles in a standing position are determined to be on the path when their lateral offset will enter the path range within a preset time period in the future.

Images