High-speed rail plug door safety control method and device integrated with TOF camera

Abstract

The present application relates to the technical field of rail transit vehicle door safety control, and discloses a high-speed rail sliding door safety control method and device integrated with a TOF camera, comprising a door opening and closing unit, a depth data acquisition unit, a safety area setting unit, a motion state sensing unit, an in-area detection unit, an out-of-area intention detection unit, a time sequence verification unit and a control execution unit; the depth image of the door area is acquired by the TOF camera, a dynamic safety area is set in combination with the door closing path, whether there is a person or object in the area is detected in the door closing process, the target outside the area is tracked across frames to determine the intention of getting on or off, and the door is controlled to stop closing after the track rationality of the detection result is analyzed in combination with the time sequence verification, so that the active identification and protection of the target that has entered or is about to enter the dangerous area are realized.

Description

Technical Field

[0001] This invention relates to the field of rail transit vehicle door safety control technology, and more specifically, to a method and device for safety control of high-speed rail sliding doors with integrated TOF camera. Background Technology

[0002] As a crucial mode of rail transportation, high-speed rail's operational safety directly impacts passenger safety and property security, as well as train operation order. The train door system, as the direct interface for passenger boarding and alighting, must balance efficiency and safety during station opening and closing. Especially in scenarios with dense passenger flow, complex passenger movement, and a large amount of carry-on luggage, there is a risk that limbs, personal belongings, or other objects may enter the door closing path during the closing process, posing safety risks such as people or objects being trapped. Therefore, high demands are placed on the real-time performance, accuracy, and reliability of door safety control.

[0003] Among existing anti-pinch solutions for train doors, one type relies primarily on door-edge triggered detection methods, such as pressure sensing, changes in clamping current, or contact-based anti-pinch structures, to respond after the door comes into contact with a person or object. This type of solution typically triggers control only after contact has occurred, making it difficult to provide early warning and proactive anti-pinch protection. Another type uses infrared beams, light curtains, or ordinary image acquisition to monitor the door area. However, in complex platform environments, these methods are easily affected by factors such as obstructions, changes in ambient light, dense crowds, and irregular object shapes, leading to misjudgments or missed detections, making it difficult to reliably identify the actual occupancy status within the door area.

[0004] Furthermore, most existing non-contact detection solutions only focus on whether there are people or objects within the door closing path. For targets located outside the closing path but with a clear tendency to get on or off the vehicle, there is usually a lack of effective identification methods, making it difficult to take timely measures to stop the door closing before the target enters the danger zone. Especially in scenarios where passengers quickly enter or exit the door area as the door is about to close, relying solely on the detection of targets within the closing path often cannot meet the safety requirements of active anti-pinch protection.

[0005] Meanwhile, some existing visual detection solutions do not make sufficient use of spatial information, relying more on two-dimensional image contours or single-frame threshold judgments, and lacking comprehensive analysis of target depth information, spatial positional relationships, and continuous temporal motion trajectories. When people's postures change, luggage shapes become irregular, or targets pause briefly and then approach the train door again, existing solutions struggle to accurately distinguish between safe and dangerous states, easily leading to doors stopping or closing incorrectly, or being missed, affecting passenger experience and train operation efficiency.

[0006] Therefore, it is necessary to provide a safety control method and device for high-speed rail sliding doors that integrates a TOF camera. By setting a safety zone along the door closing path and detecting people and objects within the safety zone based on depth images acquired by the TOF camera during the door closing process, and simultaneously identifying the boarding and alighting intentions of targets outside the safety zone, the door can be stopped from closing in a timely manner when a dangerous situation is detected, thereby improving the active anti-pinch capability and operational safety of the door. Summary of the Invention

[0007] The purpose of this invention is to provide a safety control method and device for high-speed rail sliding doors that integrates a TOF camera, in order to solve the above-mentioned problems.

[0008] On one hand, the present invention provides a safety control device for high-speed rail sliding doors that integrates a TOF camera, comprising:

[0009] The depth data acquisition unit is used to acquire a depth image matrix of the high-speed train door area using a TOF camera.

[0010] The safety zone setting unit is used to determine the safety zone corresponding to the door closing path based on the door's mechanical structure parameters, the door closing path, and the door's real-time position.

[0011] The motion state sensing unit is used to acquire the motion state signal of the door and identify whether the door is in the closing process;

[0012] The detection unit within the safe area is connected to the depth data acquisition unit, the safe area setting unit, and the motion state sensing unit. It is used to detect whether there are people or objects in the safe area based on the depth image matrix or the three-dimensional point cloud generated by the depth image matrix during the door closing process, and output a stop closing control command when the detection result is that there are people or objects.

[0013] An outside-area intent detection unit, connected to the depth data acquisition unit and the safe zone setting unit, is used to detect and track targets outside the safe zone across frames, determine whether the target has the intent to get on or off the vehicle towards the door area, and output a stop door closing control command when the determination result is that the target has the intent to get on or off the vehicle.

[0014] A timing verification unit, connected to the detection unit within the safe area and the intent detection unit outside the area, is used to perform trajectory physical rationality analysis on the detection results within a continuous time window;

[0015] The control execution unit is connected to the detection unit within the safe area, the intention detection unit outside the area, the timing verification unit, and the door opening and closing unit, and is used to control the door to stop closing when the timing verification unit confirms that the detection result is valid.

[0016] Furthermore, the security zone setting unit determines the security zone in the following manner:

[0017] The key area boundaries on the door closing path are determined based on the mechanical structure characteristics of the door; the safety sensitivity parameters of each area are dynamically adjusted according to the real-time movement speed of the door; a spatial coordinate-safety sensitivity mapping table is generated; and the door closing area is divided into high-sensitivity sub-regions, medium-sensitivity sub-regions, and low-sensitivity sub-regions according to the spatial coordinate-safety sensitivity mapping table, wherein the high-sensitivity sub-region is the core monitoring area of ​​the safety area.

[0018] Furthermore, the detection unit within the secure area and the intent detection unit outside the area include:

[0019] A point cloud generation subunit is used to generate a three-dimensional point cloud based on the depth image matrix.

[0020] The spatial region mapping sub-unit is used to determine the spatial sub-region to which each point cloud data belongs based on the spatial coordinate-security sensitivity mapping table.

[0021] The target extraction subunit is used to perform cluster analysis or target detection processing on the three-dimensional point cloud to extract personnel targets and object targets;

[0022] The trajectory association subunit is used to associate data between personnel targets and object targets in consecutive frames to form the target motion trajectory;

[0023] The result output subunit is used to output the judgment result of the presence of the target within the safe area and the judgment result of the intention to get on or off the vehicle outside the area based on the target's motion trajectory.

[0024] Furthermore, the target extraction subunit performs the following operations:

[0025] Construct a hierarchical spatial index structure for the depth image matrix; add hierarchical index depth to the point cloud data in the high-sensitivity sub-region according to the spatial coordinate-security sensitivity mapping table; identify candidate change regions based on the spatial gradient change features of the depth value; perform fine segmentation on regions exceeding the preset gradient threshold; perform 3D point cloud clustering on the candidate change regions to obtain the spatial boundary parameters, volume parameters, and position parameters of the candidate targets.

[0026] Furthermore, the preset gradient threshold is determined through statistical analysis of gradient distribution in historical scenes; areas exceeding the preset gradient threshold are synchronously marked as key anti-pinch areas of concern; when the number of cluster points, target volume parameters, or target occupied space range in the key anti-pinch areas of concern exceed the corresponding preset threshold, it is determined that there are people or objects.

[0027] Furthermore, the trajectory association subunit performs the following operations:

[0028] The system correlates the centroid positions of the target in consecutive frames; calculates the target displacement vector, velocity vector, and motion components toward the door area; predicts the target's motion path within a preset time period based on the target's current pose and historical trajectory; and determines that the target intends to get on or off the vehicle when the predicted motion path intersects with the safe area.

[0029] Furthermore, the criteria for determining the intention to get on or off the vehicle outside the designated area include:

[0030] The target's initial position is outside the safe zone; the target maintains a movement trend toward the door area within consecutive preset frames; the minimum distance between the target's predicted trajectory and the safe zone is less than a preset distance threshold; or the target will enter the safe zone within a preset time.

[0031] Furthermore, the detection unit within the safe area determines the presence of people or objects within the safe area in the following manner:

[0032] The point cloud clustering results within the safe area are compared with the background model to identify new targets; or target detection is performed on the depth image matrix to obtain the target region; when the overlap between the new target or the target region and the safe area exceeds a preset threshold, it is determined that there are people or objects within the safe area.

[0033] Furthermore, during the door closing process, if there are people or objects in the safe area, or if a target outside the safe area intends to get on or off the vehicle, the control execution unit controls the door to stop closing; after the door stops closing, if no people, objects, or intentions to get on or off the vehicle are detected in consecutive preset frames, the control unit controls the door to resume closing or remain open and ready.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows: by acquiring a depth image matrix of the door area using a TOF camera, and combining the door mechanical structure parameters, door closing path, and real-time door position to set a safety zone, the present invention detects whether there are personnel or objects within the safety zone during the door closing process, and simultaneously detects and tracks targets outside the safety zone across frames to determine whether they have the intention to get on or off the vehicle towards the door area. By combining the analysis of the trajectory physical rationality of the detection results within a continuous time window, the present invention controls the door to stop closing. This allows for the identification of personnel and objects that have entered the door closing path, as well as targets that have not yet entered but have an approaching trend, which is beneficial to improving the safety control effect and detection reliability during the door closing process.

[0035] On the other hand, the present invention also provides a safety control method for high-speed rail sliding doors integrating a TOF camera, comprising the following steps:

[0036] S1: Acquire depth image matrix of the high-speed train door area using a TOF camera;

[0037] S2: Determine the safe zone corresponding to the door closing path based on the door's mechanical structure parameters, door closing path, and real-time door position;

[0038] S3: During the closing process of the car door, the presence of personnel or objects in the safe area is detected based on the depth image matrix or the three-dimensional point cloud generated by the depth image matrix.

[0039] S4: Detect and track targets outside the safe area across frames to determine whether the target has the intention to get on or off the vehicle towards the door area;

[0040] S5: Perform trajectory physical rationality analysis on the detection results within a continuous time window;

[0041] S6: When there are people or objects in the safe area, or when a target outside the safe area intends to get on or off the vehicle, and the trajectory physical rationality analysis result is valid, control the door to stop closing.

[0042] It should be noted that the high-speed rail sliding door safety control method with integrated TOF camera provided by this invention has the same beneficial effects as its control device, and will not be described in detail here. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A functional framework diagram of a high-speed rail sliding door safety control device integrating a TOF camera is provided in an embodiment of the present invention.

[0045] Figure 2 A flowchart illustrating a safety control method for a high-speed rail sliding door integrating a TOF camera, provided as an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] See Figure 1 As shown, this application proposes a safety control device for a high-speed rail sliding door with an integrated TOF camera, including a door opening and closing unit for controlling the door, and further including:

[0049] The depth data acquisition unit is used to acquire a depth image matrix of the high-speed train door area using a TOF camera.

[0050] The safety zone setting unit is used to determine the safety zone corresponding to the door closing path based on the door's mechanical structure parameters, the door closing path, and the door's real-time position.

[0051] The motion state sensing unit is used to acquire the motion state signal of the door and identify whether the door is in the closing process;

[0052] The detection unit within the safe area is connected to the depth data acquisition unit, the safe area setting unit, and the motion state perception unit. It is used to detect whether there are people or objects in the safe area based on the depth image matrix or the three-dimensional point cloud generated by the depth image matrix during the door closing process, and output a stop closing control command when the detection result indicates that there are people or objects.

[0053] The out-of-area intent detection unit, connected to the depth data acquisition unit and the safe area setting unit, is used to detect targets outside the safe area and track them across frames, determine whether the target has the intent to get on or off the vehicle towards the door area, and output a stop door closing control command when the determination result is that the target has the intent to get on or off the vehicle.

[0054] The timing verification unit, connected to the detection unit within the safe area and the intent detection unit outside the area, is used to perform trajectory physical rationality analysis on the detection results within a continuous time window;

[0055] The control execution unit is connected to the detection unit within the safe area, the intention detection unit outside the area, the timing verification unit, and the door opening and closing unit. It is used to control the door to stop closing when the timing verification unit confirms that the detection result is valid.

[0056] Specifically, the Time-of-Flight (TOF) camera in the depth data acquisition unit emits modulated light and receives the reflected echo from the target. By calculating the time of flight or phase difference between the emitted light and the echo, the distance values ​​of each sampling point within the high-speed rail door area relative to the camera are obtained, thus forming a depth image matrix. This depth image matrix can be understood as a set of depth values ​​arranged according to the row and column coordinates of an image. Each pixel in the matrix corresponds to the depth information of a spatial sampling point within the door area. In practical applications, to ensure subsequent detection accuracy, it is preferable to first perform installation pose calibration and depth compensation calibration on the TOF camera. Installation pose calibration is used to determine the transformation relationship between the TOF camera coordinate system and the door body coordinate system. Compensation calibration is used to correct system ranging errors caused by targets with different distances and reflectivities. These calibration parameters can be obtained by fitting depth data from standard calibration boards, standard pillars, or standard planar targets at multiple known locations during the vehicle's factory testing phase. The safety zone setting unit determines the safety zone based on the door's mechanical structure parameters, door closing path, and real-time door position. The door's mechanical structure parameters include at least the door width, thickness, door leaf trajectory, door frame boundary, threshold position, and the maximum intrusion space of the door in the closing direction. The real-time door position can be obtained from the door encoder, limit switch, or displacement sensor output. Based on this, a spatial geometric model of the door closing process is established. The model essentially maps the spatial range that the car door might sweep through at various closing moments onto a unified coordinate system, thus forming a dangerous occupancy area, or safe zone, corresponding to the door's closing path. To improve applicability, the safe zone can be set as a fixed boundary area or a time-varying area that dynamically changes with the door's position. For example, when the door approaches the closing end of the door frame, the highly sensitive range near the door gap can be appropriately expanded to improve the detection sensitivity of high-risk locations where people or objects might be trapped. The motion state sensing unit is used to determine whether the door is currently in the closing process. Its output can be a closing start signal, a closing in progress signal, or a closing termination signal. Only when the door is detected to be in the closing process will the detection unit within the safe zone... Only when the intention detection unit outside the area can it enter the high-frequency detection state, thereby avoiding the introduction of invalid calculations when the car door is fully open or fully closed; the detection unit within the safe area performs detection based on the depth image matrix or the 3D point cloud converted from it. Here, generating the 3D point cloud from the depth image matrix means combining the intrinsic parameters of the TOF camera to restore the 2D pixel coordinates and corresponding depth values ​​to the 3D spatial coordinates to obtain the spatial point set of the car door area. Then, point cloud clustering, connected component segmentation or target detection algorithms based on depth features are used to identify whether there are people or objects in the safe area. People or objects include the whole human body, parts of the human body, suitcases, backpacks, hand-pushed objects, and other obstacles that may enter the closing path of the car door.To avoid mistaking depth noise, dust reflections, or the edges of the door itself for targets, it is usually necessary to set a target detection threshold. This threshold can include at least a cluster point count threshold, a target spatial volume threshold, a target duration frame count threshold, and a target intrusion into the safe zone percentage threshold. The cluster point count threshold can be determined based on statistical results of empty-field depth noise. For example, in an empty car door scene, several sets of depth data are continuously collected, the maximum number of pseudo-clusters formed by random noise is counted, and a safety margin is added to this to prevent normal noise from triggering target detection. The spatial volume threshold can be calculated based on the minimum actual size of human limbs, common luggage, and small debris, combined with the camera's spatial resolution. For example, the minimum identifiable obstacle volume can be set... The threshold values ​​are empirical values ​​ranging from 0.5L to 2L, fine-tuned based on experimental data. The continuous frame rate threshold can be set according to the camera frame rate and the door operator response time. For example, at a sampling frequency of 20 to 30 frames per second, the same target is only considered a valid obstacle when it is detected for 2 to 5 consecutive frames to suppress transient noise. The intrusion ratio threshold can be calculated by the overlap volume or overlap area ratio between the target point cloud or the target bounding box and the safe area. For example, when the overlap ratio reaches 10% to 30%, the target is considered to have actually entered the danger zone. The out-of-area intent detection unit is used to detect targets outside the safe area and track them across frames. The so-called boarding / alighting intent is not based on the passenger's subjective psychology, but on the target's continuous Objective behavioral judgments based on spatial position changes, direction of motion, speed of motion, and tendency to approach the door area in a time series can be made by first extracting the target from a depth image matrix or 3D point cloud, and then establishing the target trajectory using centroid position, contour similarity, volume continuity, or nearest neighbor association methods between adjacent frames, thus forming a target motion model. The target motion model can be understood as a mathematical expression describing the changes in the target's position, velocity, and direction over multiple consecutive time points. Based on this, the velocity component of the target towards the door area, the rate of change of distance between the target and the safety zone boundary, and the predicted trajectory within a preset time window can be calculated. When the target continues to move towards the door, its predicted trajectory intersects with the safety zone. Alternatively, if the estimated time to reach the safe area is less than a preset time threshold, it can be determined that the target intends to board or alight. The direction determination threshold, speed threshold, and time threshold should have a clear source. For example, the direction determination can be based on the angle between the target's motion direction vector and the door's normal vector, with an angle threshold ranging from 20° to 45°. The speed threshold can be set based on statistics of normal passenger walking speeds on the platform, for example, a range of 0.2 m / s to 2.0 m / s, used to exclude stationary or randomly swaying targets. The estimated arrival time threshold should be set in conjunction with the current door closing speed, braking response delay, and safety margin, for example, a range of 0.2 s to 1.0 s, to ensure the system has sufficient time to issue a stop-closing control command.The aforementioned thresholds can all be obtained through joint statistics of measured samples from test lines, passenger flow scenario samples from platforms, and simulated samples. Specifically, sample data can be collected in empty scenarios, single-person boarding / alighting scenarios, multi-person confluence scenarios, and scenarios with luggage, and the false detection rate, false negative rate, and response time can be statistically analyzed. Then, a parameter set that balances safety and usability can be selected as the final set value. The timing verification unit is used to perform trajectory physical rationality analysis on the detection results within a continuous time window. Its purpose is to eliminate false targets or false intentions that do not conform to objective motion laws, such as instantaneous abnormal point clouds caused by depth jumps, abnormal reflections, or partial occlusion. The trajectory physical rationality analysis can be based on the continuity of the target's position, velocity, and acceleration in consecutive frames. Position continuity means that the target should not undergo abrupt displacements exceeding the physically possible range between adjacent frames; velocity continuity means that the target's velocity changes should be within a reasonable range for human movement or the movement of common objects; and acceleration rationality means that the target should not be within a reasonable range for human movement or the movement of common objects. If an acceleration change significantly exceeds the actual movement capacity of a human or object within a very short period of time, the relevant judgment threshold can be set based on experimental data of human walking, brisk walking, and boarding / alighting actions. For example, the maximum displacement threshold per frame can be set to 0.05m to 0.3m, matching the frame rate, and the maximum acceleration threshold can be set to 1m / s² to 5m / s², and further corrected during the line commissioning phase. After the timing verification unit confirms the validity of the detection result, the control execution unit sends a stop-closing control command to the door opening and closing unit. Stopping the closing can be done by immediately cutting off the door closing drive and maintaining the current position, or by stopping and then performing a small-distance retraction or reopening. The specific control strategy can be set according to the door operator controller's capabilities and the overall vehicle safety logic. When no personnel or objects are detected within the safe area for several consecutive frames, and targets outside the area no longer have the intention to board or alight towards the door area, the stop-closing state is lifted, thus ensuring both safety and door closing efficiency and train operation order.

[0057] In some embodiments of this application, the security area setting unit determines the security area in the following ways:

[0058] The key area boundaries on the door closing path are determined based on the mechanical structure characteristics of the door; the safety sensitivity parameters of each area are dynamically adjusted according to the real-time movement speed of the door; a spatial coordinate-safety sensitivity mapping table is generated; and the door closing area is divided into high-sensitivity sub-regions, medium-sensitivity sub-regions, and low-sensitivity sub-regions based on the spatial coordinate-safety sensitivity mapping table, with the high-sensitivity sub-region serving as the core monitoring area of ​​the safety zone.

[0059] Specifically, the safety zone setting unit first establishes a spatial boundary model of the door during the closing process based on the mechanical structural characteristics of the door. These mechanical structural characteristics include the door leaf's outer dimensions, door thickness, door frame edge position, sill position, door guide rail position, door closing stroke, and the door leaf's motion trajectory relative to the vehicle body during closing. The critical area boundary on the door closing path refers to the boundary range with safety risks formed between the door leaf edge, door seam clamping position, door sweep area, and areas that passengers may approach when the door moves from the open position to the closed position. This boundary can be obtained directly from the door structural design drawings or by fitting a prototype after multi-condition position calibration. For example, the coordinates of door edge points are collected when the door is in multiple typical positions such as fully open, half-open, nearly closed, and fully closed, and a continuous boundary model is established through interpolation or curve fitting. Based on this, the safety zone setting unit dynamically adjusts the safety sensitivity parameters of each area in conjunction with the real-time movement speed of the door. These safety sensitivity parameters can be understood as... This is a quantitative indicator characterizing the priority and triggering severity of obstacle detection response at a specific spatial location. Higher safety sensitivity indicates that the area is closer to a high-risk location for people or objects to be trapped, and the system sets a stricter detection threshold and responds faster for targets within that area. The real-time movement speed of the car door can be obtained from feedback signals from the door operator encoder, displacement sensor, or controller. When the door closing speed increases, the door braking distance and response time margin decrease relatively. Therefore, the safety sensitivity near the door gap, the front edge of the door, and the relatively closed area between the door and the door frame should be appropriately increased. When the door closing speed is low, the sensitivity of the outer low-risk area can be reduced accordingly to avoid over-triggering. The dynamic adjustment process can be implemented using a linear mapping method or a piecewise function method. For example, the door speed can be divided into three levels: low speed, medium speed, and high speed, each corresponding to a different sensitivity gain coefficient. The threshold values ​​for each level can be determined by the door operator design parameters and actual debugging data. For example, low speed can be set to 0 m / s to 0.1 m / s, medium speed to 0.1 m / s to 0.25 m / s, and high speed to 0.Speeds above 25 m / s are acceptable, but specific values ​​can be adjusted based on the door operator response performance of different vehicle models. After completing boundary modeling and sensitivity adjustment, a spatial coordinate-safety sensitivity mapping table is generated. This mapping table establishes a correspondence between spatial location points or spatial grid units and corresponding safety sensitivity values ​​under a unified door coordinate system. It can be generated by dividing the door area into multiple three-dimensional voxel units or two-dimensional projected grid units according to a preset spatial resolution, and then calculating the sensitivity gain of each unit based on its relative distance and orientation to the door closing trajectory, door gap position, common passenger passage areas, and the current door speed. The safety sensitivity value should be determined by the fact that the spatial resolution should match the ranging accuracy and field-of-view resolution of the TOF camera, for example, at the level of 10mm, 20mm, or 50mm. The safety sensitivity value can be obtained by a weighted model, for example, by weighting the trajectory proximity weight, the door gap clamping risk weight, the speed correction weight, and the historical false trigger correction weight. Each weight parameter can be determined by circuit debugging, simulation testing, and sample statistics. If it is difficult to accurately model the value at once, empirical values ​​can be used first, for example, assigning a high sensitivity of 0.8 to 1.0 to the central area of ​​the door gap and assigning a sensitivity of 0 to the outer area of ​​the door leaf sweep edge. A medium sensitivity of 5 to 0.8 is assigned to the outermost observation area, while a low sensitivity of 0.2 to 0.5 is applied. The parameters are then iteratively adjusted based on measured false positive and false negative rates. According to the spatial coordinate-safety sensitivity mapping table, the door closing area can be divided into high-sensitivity, medium-sensitivity, and low-sensitivity sub-regions. The high-sensitivity sub-region typically includes the area near the final closing position of the door leaf, the door seam clamping strip, the near-field range of the door's leading edge sweep, and the area where human limbs and personal belongings are most likely to intrude. This area serves as the core monitoring area for safety, and during detection, a lower target trigger threshold, a higher sampling update frequency, or more stringent requirements can be used. The strategy involves continuous frame confirmation; medium-sensitive sub-regions typically include the periphery of high-sensitive sub-regions and transition areas that may briefly intersect with the door closing path, used to identify approach risks in advance; low-sensitive sub-regions serve as peripheral warning areas, mainly used to assist in judging the target's movement trend and intention to get on or off the vehicle; the thresholds for classifying high, medium, and low-sensitive sub-regions should be clearly defined, and are usually determined jointly based on the spatial sensitivity statistical distribution, vehicle safety specifications, and field test results. For example, areas with a sensitivity value greater than or equal to 0.75 can be defined as high-sensitive sub-regions, 0.45 to 0.75 as medium-sensitive sub-regions, and less than 0.75 as low-sensitive sub-regions.45 is defined as a low-sensitivity sub-region, or a percentile method based on risk ranking can be used for partitioning, for example, defining the top 20% of high-risk grids as high-sensitivity sub-regions. Through these methods, the safety zone setting unit allows the division of safety zones to move beyond simple fixed rectangles or strips, creating dynamic risk zones that match the vehicle door's mechanical structure, operating status, and actual risk distribution. This provides a reliable spatial constraint basis for the accurate detection of personnel, objects, and intentions to get on and off the vehicle.

[0060] In some embodiments of this application, the detection unit within the secure area and the intent detection unit outside the area include:

[0061] The point cloud generation subunit is used to generate a 3D point cloud based on the depth image matrix.

[0062] The spatial region mapping sub-unit is used to determine the spatial sub-region to which each point cloud data belongs based on the spatial coordinate-security sensitivity mapping table.

[0063] The target extraction subunit is used to perform cluster analysis or target detection processing on the 3D point cloud to extract personnel and object targets;

[0064] The trajectory association subunit is used to associate data between personnel targets and object targets in consecutive frames to form the target motion trajectory;

[0065] The result output sub-unit is used to output the judgment result of the presence of the target within the safe area and the judgment result of the intention to get on or off the vehicle outside the area, based on the target's movement trajectory.

[0066] Specifically, the point cloud generation subunit is used to convert the depth image matrix output by the TOF camera into point cloud data in three-dimensional space. Each pixel in the depth image matrix corresponds to a depth value. Combining the intrinsic parameters of the TOF camera, the installation pose parameters, and the pixel coordinates, the three-dimensional spatial coordinates of each pixel in the door coordinate system can be calculated, thereby forming a three-dimensional point cloud to characterize the spatial distribution of the door area. To ensure the quality of the point cloud, it is preferable to perform preprocessing on the original depth data before point cloud generation. Preprocessing may include invalid depth point removal, flying point filtering, isolated noise point removal, depth smoothing, and static background correction of the ground or door. The relevant filtering parameters should have a clear source, for example, through empty... Multiple sets of depth data are continuously collected in the scene to statistically analyze the spatial distribution density of random noise points, depth fluctuation amplitude, and number of isolated points. This allows for the setting of neighborhood radius thresholds, minimum neighborhood point thresholds, and smoothing window sizes. For example, the neighborhood radius can be set from 20mm to 80mm, and the minimum neighborhood point count can be set from 3 to 10. The spatial region mapping sub-unit is used to map each point in the 3D point cloud to its corresponding spatial sub-region based on a pre-established spatial coordinate-security sensitivity mapping table. This determines whether each point is located in a high-sensitivity, medium-sensitivity, or low-sensitivity sub-region. Essentially, the mapping process matches the point cloud coordinates with pre-divided spatial grids, voxel units, or boundary models, thus providing spatial support for subsequent target recognition. Risk attributes; in other words, for targets of the same size or category, if they are located in a high-sensitivity sub-region, their trigger threshold can be set lower, and if they are located in a low-sensitivity sub-region, the trigger threshold can be appropriately increased to balance safety and false detection control; the target extraction sub-unit is used to perform cluster analysis or target detection processing on 3D point clouds to extract personnel targets and object targets. Among them, the so-called cluster analysis refers to aggregating point clouds belonging to the same entity into point cloud clusters based on the spatial proximity relationship between points. Common methods can be Euclidean distance clustering, density-based clustering, or region growing clustering. The clustering model is not set out of thin air, but is established based on the spatial resolution of the TOF camera, the expected target size range, and the actual scene noise level. For example, the clustering distance threshold can be determined based on the common spacing between adjacent human body surface points and the camera ranging resolution, and can be taken from 30mm to 150mm. The minimum clustering point threshold can be set by combining the minimum number of effective points corresponding to local human limbs, luggage edges, or small objects under the current ranging conditions. If target detection processing is used, a depth target detection model can be built based on the depth image matrix. The model can be obtained by collecting depth sample data under different passenger postures, different item types, and different platform conditions in the door area and then annotating and training it. The training samples can be labeled with categories such as human limbs, luggage, and other obstacles. Then, a target detector can be established through a convolutional neural network or other classification and detection networks suitable for depth map processing.If the amount of data is insufficient, a combination of rule-based models and small-sample classification models can be used. For example, candidate targets can be screened based on point cloud volume, length, width, height, ground clearance, and contour continuity, and then a shallow classifier can be used to distinguish between personnel targets and object targets. Personnel targets refer to targets corresponding to the whole human body or parts of the human body, while object targets refer to suitcases, backpacks, handbags, trolley parts, or other obstacles that may affect the safety of closing the car door. The trajectory association subunit is used to perform data association between personnel targets and object targets in consecutive frames to form target motion trajectories. Data association refers to the process of associating the targets detected in the current frame with the target data. Matching targets with those in the previous frame or several previous frames to determine if they belong to the same real object. Common association criteria can include the target's centroid position, bounding box size, point cloud volume, direction of motion, and region affiliation consistency. For example, initial matching can be performed based on the Euclidean distance between the centroids of targets in adjacent frames, followed by secondary screening based on the target size change rate and velocity continuity, thus avoiding confusing two close but different passengers as the same target. Related association thresholds should also have a clear source; for example, the centroid displacement threshold can be determined based on the camera frame rate and the maximum reasonable human body movement speed. At 20 frames per second, if the human body moves quickly... If the approach speed is estimated to be no more than 2 m / s, the maximum displacement between adjacent frames can be initially set to about 100 mm, and then corrected based on experimental samples. To improve the correlation stability, a target state prediction model can also be established, such as a uniform motion model or a Kalman filter prediction model. Based on the position and velocity of the previous few frames, the possible position of the target in the current frame can be predicted, and then matching can be performed near the predicted position. The initial noise covariance, measurement noise covariance, and other parameters of the prediction model can be obtained by statistically analyzing a large number of real boarding and alighting motion samples collected during the line commissioning phase. The result output subunit is used to output a judgment on the existence of the target within the safe area based on the target's motion trajectory. The results and the results of judging the intention to get on and off the vehicle outside the area are as follows. The judgment result of the existence within the safe area refers to the judgment of whether there are real and effective personnel or object targets that have entered or occupied the safe area. The judgment should not be based solely on the detection results of a single frame. Instead, it is preferable to comprehensively determine the results based on factors such as whether the target has entered the high-sensitivity sub-region, whether the number of consecutive frames has reached a preset value, whether the target volume exceeds the minimum obstacle threshold, and whether the difference from the static background is significant. For example, when a target enters the high-sensitivity sub-region for more than 3 consecutive frames and the number of cluster points is greater than the set value and the space occupied is greater than the minimum volume threshold, the judgment result of the existence within the safe area can be output.The so-called "outside-area boarding / alighting intention judgment result" refers to the situation where, although the target has not yet entered the safe zone, its movement trajectory indicates a tendency to approach the door and potentially enter the closing path. This judgment is determined by comprehensively considering factors such as the rate of change of the distance from the target to the safe zone, the velocity component towards the door, the intersection of the predicted trajectory and the boundary of the safe zone, and the duration of the approach. For example, if the target continues to move towards the door for several consecutive frames and is predicted to enter a highly sensitive sub-region within 0.3s to 1.0s, or if its predicted trajectory intersects with the safe zone, it can be determined that the target has an intention to board / alight. The aforementioned time threshold, velocity threshold, and intersection... The optimal threshold for fork detection is obtained through a combination of sampling of actual passenger flow scenarios on the platform, joint debugging tests under different door closing speed conditions, and simulation verification. If an example is required in the application documents, it can be stated that the threshold for the speed component facing the door can be taken as 0.2 m / s to 1.5 m / s, the number of consecutive approach frames can be taken as 2 to 5 frames, and the prediction time window can be taken as 0.3 s to 1.0 s. Through the collaborative work of the above sub-units, a continuous processing chain can be realized from depth image matrix to 3D spatial perception, from target recognition to trajectory tracking, and then to safe area occupancy judgment and boarding / alighting intention recognition, thereby improving the accuracy, real-time performance, and stability of the active anti-pinch control of the door.

[0067] In some embodiments of this application, the target extraction subunit performs the following operations:

[0068] Construct a hierarchical spatial index structure for the depth image matrix; add hierarchical index depth to the point cloud data in the high-sensitivity sub-regions according to the spatial coordinate-security sensitivity mapping table; identify candidate change regions based on the spatial gradient change features of the depth values; perform fine segmentation on regions exceeding the preset gradient threshold; perform 3D point cloud clustering on the candidate change regions to obtain the spatial boundary parameters, volume parameters, and position parameters of the candidate targets.

[0069] Specifically, the target extraction subunit first constructs a hierarchical spatial index structure for the depth image matrix. This hierarchical spatial index structure refers to a data structure that organizes depth data hierarchically according to spatial location. Its purpose is to reduce the computational load during subsequent candidate region retrieval and target segmentation, and to improve the accuracy of locating local anomalies. This index structure can recursively partition the two-dimensional depth image plane using a quadtree, or recursively partition the three-dimensional spatial point cloud recovered from the depth image matrix using an octree. Alternatively, it can use a multi-layer mesh partitioning method to divide the car door area into coarse-grained and fine-grained regions. The upper-layer structure is used to quickly locate areas where targets may exist, while the lower-layer structure is used to refine key areas. Detailed analysis is needed; the establishment of a hierarchical index structure is not arbitrary but should be determined based on the TOF camera's field of view, pixel resolution, ranging accuracy, and the actual size of the vehicle door area. For example, for a detection scenario covering a vehicle doorway with a width of approximately 1m to 2m and a depth range of 0.5m to 2m, a first-level division can be performed using a larger grid size. Then, areas with significant depth changes or located in highly sensitive sub-regions within the first-level grid can be further subdivided into smaller sub-grids to balance computational efficiency and local recognition accuracy. Based on the spatial coordinate-security sensitivity mapping table, increasing the hierarchical index depth for point cloud data within highly sensitive sub-regions means using a higher or smaller division level than for medium-sensitive and low-sensitive sub-regions for spatial data falling into highly sensitive sub-regions. The retrieval units are designed to allow for more refined representation and analysis of data near door seams, the front edge of doors, or locations with a higher risk of being caught. In other words, within highly sensitive sub-regions, more index nodes or higher spatial sampling densities will be allocated within the same area or volume range, thereby improving the ability to identify human limb edges, protruding parts of small luggage, and locally intrusive targets. This increased hierarchical indexing depth can be achieved by setting the maximum number of division layers for highly sensitive sub-regions to 1.5 to 3 times that of medium-sensitive sub-regions, or by setting the minimum voxel side length for highly sensitive sub-regions to half or less of that for medium-sensitive sub-regions. For example, when medium-sensitive sub-regions use 40mm to 60mm spatial units, highly sensitive sub-regions can use 10mm to 30mm units. The specific value of the m-level spatial unit should be determined based on the actual spatial resolution of the TOF camera at the corresponding distance. In principle, it should not be smaller than the minimum spatial scale that the system can stably resolve. After establishing a hierarchical index, candidate change regions are identified based on the spatial gradient change features of depth values. The so-called spatial gradient change features refer to the magnitude and direction of the change in depth values ​​between adjacent pixels, adjacent grids, or adjacent voxels. They are used to reflect the degree of abrupt change in the local spatial surface. If a human body, luggage, or other obstacle intrudes into a certain area, there will usually be a relatively obvious depth transition between it and the background plane, door surface, or ground. Therefore, indicators such as the first-order difference of depth, local gradient magnitude, normal change rate, or neighborhood depth dispersion can be used to locate suspected target areas.For example, the depth difference between adjacent pixels can be calculated along both row and column directions in the depth image matrix, and the gradient magnitude can be obtained from this. Regions with larger gradient magnitudes can then be selected as candidate change regions. Alternatively, the surface normal vector changes and curvature changes of local neighborhood points can be calculated in the point cloud space to identify regions with significantly different geometric continuity from the background. Fine segmentation is performed on regions exceeding a preset gradient threshold. This means that when the depth gradient change of a region reaches a level sufficient to indicate the existence of a real object boundary or abrupt shape change, the target boundary extraction processing of that region is performed with finer granularity than the initial index layer. Fine segmentation can employ region growing, edge constraint segmentation, local adaptive threshold segmentation, or connectivity-based point segmentation. The method is cloud segmentation; the preset gradient threshold must have a clear source and cannot be merely described in principle. It is preferably determined based on the statistical difference between the depth noise fluctuation range in the background scene and the depth transition amplitude at the edge of the real target. Specifically, multiple sets of depth samples can be collected in empty scenes, static door scenes, and typical passenger / luggage intrusion scenes. The upper bound of the gradient distribution in the pure noise region and the lower bound of the gradient distribution at the boundary of the real target are statistically analyzed. Then, a boundary value that balances the false detection rate and the false negative rate is selected as the preset gradient threshold. If an example is required in the application, it can be stated as: when the standard deviation of the depth measurement by the TOF camera in the door area is within the range of 5mm to 15mm... The local gradient threshold can be initially set to 20mm to 80mm and further adjusted during actual vehicle testing based on the platform environment, door material reflection characteristics, and noise levels. Furthermore, the gradient threshold can be designed as an adaptive threshold related to regional sensitivity, using a lower threshold in highly sensitive sub-regions and a higher threshold in medium- to low-sensitivity sub-regions to achieve earlier triggering in high-risk areas and reduced false detections in low-risk areas. Three-dimensional point cloud clustering is performed on candidate change regions to obtain the spatial boundary parameters, volume parameters, and position parameters of candidate targets. Three-dimensional point cloud clustering refers to grouping discrete spatial points belonging to the same potential entity according to spatial proximity and geometric continuity. A cluster is formed to determine whether it constitutes a real person or object target. The clustering method used can be Euclidean distance clustering, density clustering, or constrained region growing clustering. The clustering parameters should be determined in combination with the minimum target size, camera distance resolution, and scene noise level. For example, the inter-point aggregation distance threshold can be set to 20mm to 100mm, and the minimum number of cluster points can be set to 5 to 50, preferably obtained through statistical analysis of typical obstacle samples. Spatial boundary parameters refer to parameters used to characterize the external spatial contour range of candidate targets, such as the length, width, and height of the minimum circumscribed cuboid, the boundary coordinates of the bounding box in the door coordinate system, or the maximum and minimum value range of the target in each coordinate axis direction.Volume parameters are parameters used to characterize the spatial occupancy of candidate targets, and can include bounding box volume, convex hull volume, voxel occupancy volume, or effective point cloud volume estimates. Volume parameters help distinguish between human body parts, small debris, and invalid noise clusters. Position parameters are parameters used to characterize the relative positional relationship of candidate targets within the door region, such as the target centroid coordinates, lowest point coordinates, closest distance to the door seam boundary, overlap ratio with the safety zone boundary, and projection position relative to the door closing direction. These parameters can be extracted directly by calculating the geometric center, bounding box, and spatial distribution statistics of the clustered point cloud clusters. Furthermore, to improve the realism assessment capability, the target extraction subunit can also use spatial boundary parameters, volume parameters, and position parameters to perform preliminary screening of candidate targets. For example, when… If a candidate target's volume is smaller than the preset minimum obstacle volume, its duration is only one frame, and its location is far from the highly sensitive sub-region, it can be classified as noise or a non-dangerous target. The minimum obstacle volume threshold can be determined based on the minimum recognizable cross-sectional size of human limbs, the size of common luggage edges, and the system's anti-pinch safety requirements. For example, it can be set to the voxel occupancy scale within the range of 20mm×20mm×50mm to 50mm×50mm×100mm, with the specific value based on actual measured recognition capabilities. Through the above processing, the target extraction subunit can progressively complete spatial index establishment, key region enhancement, candidate change detection, fine segmentation, and candidate target parameter extraction from the original depth image matrix, thereby providing a clear geometric and physical input basis for subsequent target classification, trajectory association, and safety determination.

[0070] In some embodiments of this application, the preset gradient threshold is determined by statistical analysis of gradient distribution in historical scenes; areas exceeding the preset gradient threshold are synchronously marked as key anti-pinch areas of concern; when the number of cluster points, target volume parameters, or target occupied space range within the key anti-pinch areas of concern exceed the corresponding preset threshold, it is determined that there are people or objects.

[0071] Specifically, the preset gradient threshold is not arbitrarily set, but determined through statistical analysis of gradient distribution in historical scenarios. Historical scenarios refer to a set of multiple depth detection samples pre-collected and stored in the actual or simulated high-speed rail door operating environment. Preferably, these samples include at least empty scene samples, normal passenger passage samples, passenger limbs approaching the door gap samples, samples of passengers boarding and alighting with luggage samples, samples of multiple people converging, and samples of door movement interference. This ensures that the obtained threshold can cover common platform conditions. In practical implementation, the spatial gradient values ​​can be calculated frame-by-frame on the depth image matrix of the aforementioned historical samples, statistically analyzing the background area, door edge area, and real people / objects. The gradient distribution characteristics corresponding to each intrusion area are analyzed, such as mean, variance, quantile, and peak distribution. The upper bound of the gradient caused by background noise is compared with the lower bound of the gradient caused by the actual target boundary to determine a boundary value that effectively distinguishes noise changes from actual obstacle changes as a preset gradient threshold. Furthermore, to improve adaptability, preset gradient thresholds can be set by region and working condition. For example, a lower threshold can be used for highly sensitive sub-regions, a medium threshold for moderately sensitive sub-regions, and a higher threshold for low-sensitive sub-regions. Alternatively, threshold tables can be established based on changes in daytime, nighttime, platform reflection conditions, and different door closing speeds. (If the instruction manual...) The implementation example needs to be provided, which can be described as follows: After statistical analysis of historical samples, when the 95th percentile of the background noise gradient is less than 25mm, while the gradients of most real people or object boundaries are concentrated above 40mm, the initial preset gradient threshold can be set to 30mm to 40mm, and further adjusted based on the false detection rate and false negative rate during the actual vehicle debugging phase; when the gradient change in a certain area exceeds the preset gradient threshold, it indicates that the area has undergone a significant spatial abrupt change relative to the surrounding background, so the area is simultaneously marked as a key area of ​​focus for anti-pinch detection. The key area of ​​focus for anti-pinch detection refers to the area that is given priority for detailed analysis by the system in the current detection frame or several consecutive detection frames. The system continuously tracks high-risk local areas, which are not necessarily equivalent to the final personnel or object targets. Instead, they indicate that there is a high probability of real obstacle intrusion, requiring further clustering analysis and parameter judgment. After being marked as a key area of ​​concern for anti-pinch measures, the system further counts the number of cluster points, target volume parameters, and target occupied space within the area, and compares them with their respective preset thresholds. The number of cluster points refers to the number of valid points belonging to the same candidate target after 3D point cloud clustering. Its purpose is to reflect whether the target has stable and continuous spatial support, avoiding misjudging a small number of discrete points formed by random noise as real targets.The cluster point threshold is preferably determined based on the statistical distribution of points in noise samples and real obstacle samples in an empty scene. For example, the maximum number of points in pseudo-clustering in an empty scene can be counted first, and then a safety margin can be added to this maximum number of points to serve as the minimum effective cluster point threshold. In practical applications, this threshold can be set to 10 to 50 points, with the specific value depending on the camera resolution, installation distance, and detection area size. The target volume parameter refers to the volume occupied by the candidate target in three-dimensional space, which can be estimated from its bounding box volume, voxel volume, or convex hull volume. This parameter is used to exclude non-dangerous targets that, although having a certain number of points, are too small in actual size to pose a risk of trapping people or objects. Examples of obstacles include dust clumps, abnormal local reflections, or extremely small debris. The preset threshold for the target volume parameter can be determined based on the minimum identifiable size of human limbs, the dimensions of common luggage corners, and door anti-pinch safety requirements. For instance, the minimum effective obstacle volume can be set to a volume range corresponding to 20mm×20mm×50mm, 30mm×30mm×80mm, or larger. Specific values ​​can be determined after verifying the system's stable recognition capability for small objects and limb edges in experiments. The target's occupied space range refers to the actual space covered by the candidate target within the safe area, which can be determined by its bounding box size in the door coordinate system and its overlap length with the safe area boundary. The degree, overlap area, or overlap volume represent the extent to which a target has substantially entered the door closing path or a critical danger zone. Correspondingly, the threshold for the target's occupied space can be set based on the door clamping boundary width, the door gap risk zone width, and the target intrusion depth requirement. For example, when the target's intrusion depth in the closing direction exceeds 10mm to 30mm, or its overlap with a highly sensitive sub-region exceeds 10% to 25%, it can be considered to have created an actual anti-pinch risk. In terms of judgment logic, the aforementioned cluster point count, target volume parameter, and target occupied space range can all be determined using a parallel triggering method where satisfying any one of them indicates existence, thereby improving the sensitivity of safety protection. Alternatively, a combination of at least two conditions or a main condition plus an auxiliary condition can be used to reduce the false detection rate. For example, it can be stipulated that when the number of cluster points in a candidate region exceeds a threshold and the space occupied by the target exceeds a threshold, the presence of a person or object can be directly determined; or when the target volume parameter significantly exceeds a threshold, even if the number of cluster points is slightly lower, it can be regarded as a valid target. The above combination rules can be optimized based on actual false detection samples. Furthermore, to enhance realism, a continuous frame confirmation condition can be added, that is, only when the above judgment result is continuously true within 2 to 5 consecutive frames, the final judgment result of the presence of a person or object will be output, thereby suppressing false triggering caused by abnormal fluctuations in a single frame.Through the above methods, the system can reasonably determine the preset gradient threshold based on historical sample statistical results. After detecting local spatial abrupt changes, it combines the number of cluster points, target volume parameters, and the target's occupied space range to perform multi-dimensional constraint judgments on candidate regions. This allows for a more reliable distinction between real people and real objects and background noise or non-dangerous disturbances, improving the accuracy and engineering feasibility of active anti-pinch detection.

[0072] In some embodiments of this application, the trajectory association subunit performs the following operations:

[0073] The system correlates the centroid positions of the target in consecutive frames; calculates the target displacement vector, velocity vector, and motion components toward the door area; predicts the target's motion path within a preset time period based on the target's current pose and historical trajectory; and determines that the target intends to get on or off the vehicle when the predicted motion path intersects with the safe area.

[0074] Specifically, the trajectory association subunit first associates the centroid positions of targets in consecutive frames. Consecutive frames refer to adjacent depth data frames continuously output by a TOF camera at a predetermined sampling frequency. The centroid position refers to the three-dimensional coordinate value of the spatial center of a person or object target obtained through cluster analysis or target detection in each frame, calculated in the vehicle door coordinate system. This centroid position can be obtained by averaging the coordinates of each point in the target point cloud or by converting the center position of the target bounding box. Associating the centroid positions of targets in consecutive frames essentially involves determining which target in the current frame belongs to the same real object as a target in the previous frame or several previous frames, in order to form a continuous and consistent motion trajectory. This can be achieved through... A target matching model based on distance constraints, size constraints, and motion continuity constraints is established. For example, the Euclidean distance between the centroid of the target in the current frame and the centroids of all targets in historical frames is first calculated. The target with the smallest distance that is less than a preset correlation distance threshold is used as the initial matching object. Then, a secondary screening is performed by combining the target volume change rate, the bounding box size change rate, and the continuity of the sub-region in the space, so as to avoid mismatching two passengers that are close to each other but exist independently as the same target. The correlation distance threshold should have a clear source, preferably determined by comprehensively considering the TOF camera frame rate, the typical passenger movement speed, and the scale of the safe area. For example, at a sampling frequency of 20 frames / second to 30 frames / second, the upper limit of the possible fast movement speed of passengers is set to 1.5 m / s to 2 m / s.If the speed is 0 m / s, the maximum reasonable displacement per frame can be initially set to 50 mm to 100 mm, and then corrected based on the false correlation rate under different passenger flow density scenarios in actual vehicle testing. After completing the continuous frame correlation, the target displacement vector, velocity vector, and motion component towards the door area are further calculated. The displacement vector refers to the direction and length quantities formed by the change in the centroid position of the target between two or several adjacent frames, used to characterize the direction and amplitude of the target's movement. The velocity vector is the motion state parameter obtained by dividing the displacement vector by the corresponding time interval, used to characterize the target's speed and direction of movement. The motion component towards the door area refers to the... The component value obtained by projecting the velocity vector along the normal direction of the door closing path or the preset door entry direction is significant in separating the effective component from the overall motion of the target to determine whether it is approaching the danger zone of the door. For example, although the target is moving as a whole, if its main direction of motion is parallel to the edge of the door rather than towards the doorway, its motion component towards the door area will be smaller, and it is usually not appropriate to directly determine that it intends to get on or off the vehicle. The projection direction towards the door area can be predetermined based on the normal vector perpendicular to the boundary of the safe area in the door coordinate system, or it can be calibrated according to the door structure of different vehicle models, the platform layout direction, and the normal passenger entry and exit direction. To reduce the impact of single-frame noise on velocity estimation, it is preferable not to calculate the velocity solely based on the difference between two frames, but to use multi-frame sliding window averaging, weighted moving average, or state estimation models for smoothing. For example, the centroid trajectory can be fitted within the most recent 3 to 8 frames to obtain more stable displacement and velocity vectors. Based on this, the target's motion path within a preset time period is predicted based on its current pose and historical trajectory. The current pose can be understood as the target's current position and orientation. For regular objects, this can be described by the centroid position, principal axis direction, and bounding box pose; for human targets, it can be described by at least... The target's centroid position, overall motion direction, and height distribution characteristics are used for approximate characterization. The historical trajectory refers to the target's position sequence, velocity sequence, and regional change sequence in several consecutive frames. The future preset duration is a time window used for short-term behavior prediction. Its setting should match the door closing speed, controller response delay, and braking distance. It should not be too short so that the door closing action cannot be taken in advance, nor too long so that too much prediction error is introduced. It is preferably set to 0.2s to 1.0s. For example, in scenarios where the door control response is fast, it can be taken to 0.3s to 0.5s. In scenarios where a larger safety margin is required, it can be taken to 0.5s to 0.The specific value of 8s can be obtained by combining the time required for the door to decelerate from its current speed to a stop with the common approach speed of passengers; the motion path prediction can adopt a uniform linear prediction model, a uniform acceleration prediction model, or a state prediction model based on Kalman filtering. Among these, considering the system's real-time performance and implementation complexity, a uniform or weak acceleration model based on the estimation of position and velocity from several historical frames is preferred. This assumes that the target maintains its current main motion trend for a short period of time, and extrapolates the position points at several future moments to form the predicted trajectory; the initial velocity in the uniform model is directly obtained from the aforementioned velocity vector, while the uniform acceleration model... The acceleration parameters in the model can be calculated based on historical trajectory differences, while the state transition matrix, process noise covariance, and observation noise covariance in the Kalman filter model should be established using measured samples. For example, a large amount of trajectory data can be collected under typical platform conditions, including normal boarding and alighting, turning back near closing doors, walking side-by-side, and approaching the door area with luggage. The speed fluctuations and measurement jitter characteristics can be statistically analyzed to calibrate the model parameters. If a more specific and feasible description is needed in the instruction manual, it can be stated that: the historical trajectory prediction model is established by collecting target motion samples under various passenger flow conditions, and based on the displacement distribution and velocity distribution of adjacent frames in the samples... Initial motion parameters are obtained by fitting the distribution of changes in orientation and direction, and then further corrected during the actual vehicle debugging phase. When the predicted motion path intersects with the safe area, it is determined that the target intends to get on or off the vehicle. Intersection is not limited to the predicted trajectory centerline strictly crossing the boundary of the safe area; it can also include situations where the predicted trajectory envelope overlaps with the safe area, or where the target bounding box contacts the safe area after expanding outwards by a safety margin. The determination method can be either calculating whether discrete sampling points on the predicted trajectory fall within the safe area, or calculating whether the minimum distance between the target's future bounding box and the safe area is less than a preset distance threshold. The distance threshold should be based on clear criteria, and can usually be set comprehensively based on the TOF ranging error, trajectory prediction error, and door braking safety margin, for example, it can be 20mm to 100mm. Furthermore, in order to avoid misjudging a target that briefly passes by the doorway but does not enter the door as having the intention to get on or off the vehicle, auxiliary judgment conditions can also be set, such as requiring the target's motion component toward the door area to be positive for several consecutive frames, the distance between the target and the boundary of the safe area to continuously decrease, and the predicted time to enter the safe area to be less than a preset time threshold. The consecutive frame threshold can be, for example, 2 to 5 frames, and the time threshold can be, for example, 0.3s to 0.The parameters mentioned above, all within 8 seconds, can be optimized through passenger flow sample statistics and playback of misjudgment cases. Through this processing, the trajectory association subunit does not merely make static judgments about the target based on a single frame position, but rather establishes a consistent trajectory based on the target's spatial evolution over continuous time. It further extracts the target's true movement trend as it approaches the danger zone near the door, and outputs a judgment result indicating the intention to board or alight when the predicted movement path intersects with the safe zone. This provides a temporally continuous and physically reasonable basis for the door stopping and closing control.

[0075] In some embodiments of this application, the criteria for determining the intention to get on or off the vehicle outside the designated area include:

[0076] The target's initial position is outside the safe zone; the target maintains a movement trend toward the door area within consecutive preset frames; the minimum distance between the target's predicted trajectory and the safe zone is less than a preset distance threshold; or the target will enter the safe zone within a preset time.

[0077] Specifically, the criteria for determining the intention to get on or off the vehicle outside the designated area are used to identify targets that have not yet entered the safe zone but have a possibility of entering the door closing path in the short term. Here, "the target's initial position is outside the safe zone" means that when the target is first detected and its trajectory is established, its centroid position, bounding box boundary, or main point cloud occupancy area does not overlap with the safe zone, but is located outside the high-sensitivity sub-region, medium-sensitivity sub-region, or within the observation area surrounding the safe zone. This condition distinguishes targets that have entered the anti-pinch risk zone from targets that have not yet entered but have an approaching trend, thus ensuring that the detection of intentions outside the area and the occupancy detection within the safe zone are functionally interconnected and do not cause confusion. The target maintains a motion trend towards the door area within consecutive preset frames. The potential does not merely indicate that the target is moving, but rather that the projection component of the target's velocity vector in the direction of the door continuously satisfies the approach condition across several consecutive detection frames, and the distance between the target and the boundary of the safe area generally decreases. In other words, the target should exhibit an objective motion state of approaching the door area, preparing to enter the door or get on or off the vehicle, rather than just randomly swaying, passing laterally, or briefly stopping outside the door area. To avoid misinterpreting instantaneous noise or accidental attitude changes as intentions to get on or off the vehicle, the number of consecutive preset frames should have a clear source. It is preferable to combine the TOF camera frame rate, the target's normal walking speed, and the door control response delay. For example, under sampling conditions of 20 frames / second to 30 frames / second, the number of consecutive preset frames can be set to 2 to 5 frames, so that the system is based on at least 0.1 to 0.The judgment is made based on continuous motion information of about 25 seconds. The specific value can be obtained by statistically analyzing the false positive rate and false negative rate under different frame count conditions during actual vehicle testing. The minimum distance between the predicted trajectory of the target and the safe area is less than the preset distance threshold. This means that after predicting the future motion path based on the target's current pose and historical trajectory, the shortest spatial distance between the predicted path and the boundary of the safe area is calculated. When the minimum spatial distance is less than the preset distance threshold, it means that although the target has not yet entered the safe area at the current moment, its subsequent motion is close enough to the danger zone of the door, and it should be regarded as having a high tendency to get on or off the vehicle. The minimum distance can be calculated by the shortest Euclidean distance from the target's centroid trajectory to the boundary of the safe area, or by the minimum distance between the target's bounding box, circumscribed ellipsoid, or point cloud envelope and the safe area. The preset distance threshold should not be determined subjectively or arbitrarily, but should be set in combination with the TOF ranging error, trajectory prediction error, door braking response distance, and safety margin. For example, the average deviation range of the predicted trajectory of the target and the camera ranging error can be statistically analyzed in a typical platform scenario. The range is then determined based on the minimum safe distance required for the door to switch from its current closed state to a stopped state. In practice, this preset distance threshold can be set to, for example, 20mm to 100mm. It is preferable to use a smaller safety margin when approaching highly sensitive sub-regions and to increase it appropriately when the door speed is high. The target entering the safe zone within a preset time means that, based on the target's current speed, direction, and short-term motion model prediction, the target is expected to cross the boundary of the safe zone and enter the range corresponding to the door closing path within a future time window starting from the current moment. The preset time should be matched with the door operator response time, the stop closing execution time, and the total system processing delay. It must be ensured that there is still enough time to output and execute the stop closing control command after the system makes an intent determination. Therefore, the preset time can usually be obtained by superimposing the perception processing delay + controller response delay + door operator execution delay + safety margin. For example, under the condition that the overall delay of the system perception and control link is short, the preset time can be set to 0.3s to 0.8s. When a higher safety margin is required, it can also be set to 1.The specific value is preferably determined through a combination of door mechanism braking tests and passenger approach speed sample statistics. The condition that the target will enter the safe zone within a preset time reflects a parallel judgment relationship. That is, even if the target's predicted trajectory does not form a clear geometric intersection with the safe zone, but based on the current speed and distance analysis, the target will enter the safe zone in a very short time, it can still be directly determined that the target intends to get on or off the vehicle, thus avoiding missed judgments caused by relying solely on a single geometric intersection condition. Furthermore, to enhance the realism and stability of the judgment, the above conditions can be designed as a combination of basic and enhanced conditions. For example, requiring the target's initial position to be outside the safe zone as a prerequisite, and on this basis, simultaneously satisfying at least one of the following: maintaining a movement trend towards the door area within consecutive preset frames and the minimum distance between the predicted trajectory and the safe zone being less than a preset distance threshold; or simultaneously satisfying the condition that the distance continuously decreases and the target will enter the safe zone within a preset time. When two conditions are met for entering the safe zone, the system outputs a judgment result indicating that the intention to get on or off the vehicle outside the zone is valid. Correspondingly, the determination of relevant thresholds and setpoints can be accomplished jointly using historical passenger flow samples, simulation samples, and real vehicle debugging samples. Specifically, target trajectory data can be collected under various scenarios such as normal passage, lingering outside the door, lateral passage, rapid entry near closing time, and approaching with luggage. The differences in continuous frame trends, minimum distances, and expected entry times between samples with actual boarding / alighting behavior and those without are statistically analyzed. Then, a parameter set that balances anti-pinch safety and false trigger control effects is selected as the final setpoint. Through these judgment conditions, the system can make an advance judgment on whether a target intends to board or alight before it actually enters the safe zone, based on its initial spatial position, continuous approach trend, the spatial relationship between the predicted trajectory and the safe zone, and the expected entry time, thus providing a forward-looking basis for door closing control.

[0078] In some embodiments of this application, the detection unit within the safe area determines the presence of people or objects within the safe area in the following ways:

[0079] The point cloud clustering results within the safe area are compared with the background model to identify new targets; or target detection is performed on the depth image matrix to obtain the target area; when the overlap between the new target or the target area and the safe area exceeds a preset threshold, it is determined that there are people or objects within the safe area.

[0080] Specifically, the detection unit within the safe area can determine the presence of personnel or objects by employing either a new target detection method based on a background model or a target detection method based on a depth image matrix. Either method can be used individually or in combination to improve robustness. The background model refers to a reference scene model obtained by modeling the spatial distribution of a depth image matrix or 3D point cloud in a normal empty scene with no personnel, luggage, or other obstacles intruding into the door area. Its function is to characterize the depth distribution features of the door, door frame, ground, platform edges, and other fixed structures under normal conditions. The background model can be established by continuously collecting multiple images under conditions where the train is stationary, the door is open or partially open but the door area is unobstructed. Frame depth data is used to establish a background model by performing temporal averaging, median fusion, or statistical analysis of stable regions to reduce the impact of single-frame noise on background modeling. Furthermore, to adapt to changes in door position and slow environmental drift, the background model is preferably established as a dynamic background model that updates in segments according to the door position. This means that corresponding reference backgrounds are saved when the door is at different opening positions, or the current background is obtained by interpolating between adjacent background templates based on the real-time position of the door, thus avoiding misidentification of depth changes caused by the door's own movement as new targets. The point cloud clustering results within the safe area are compared with the background model to determine if a new target is a new target. This involves first generating a point cloud from the depth data within the safe area in the current frame and clustering it, then comparing each clustering result with the corresponding background model. Static spatial distribution is compared using difference methods. When a cluster does not have a corresponding stable structure in the background model, or shows a significant increase in spatial occupancy relative to the background model, it is identified as a new target. Comparison methods can include distance comparison from a point to the background surface, voxel occupancy difference comparison, difference in the number of points before and after, or bounding box difference comparison. For example, the occupancy voxel sets of the current frame and the background model can be counted separately in a unified voxel grid. When the number of newly occupied voxels in the current frame reaches a preset value, a new target is considered to exist. On the other hand, target detection is performed on the depth image matrix. Target regions refer to directly extracting target contours or candidate regions based on the depth image matrix, without entirely relying on background difference. This can be achieved by... Rule-based detection methods based on depth gradients, local depth discontinuities, and region connectivity can also be object detection models trained on depth samples. Preferably, these models are trained using sample data containing only depth information to avoid introducing RGB image dependencies. Training samples can include depth maps of the entire human body, parts of the human body, suitcases, backpacks, handbags, and other common obstacles at different doorway locations, in different poses, and under different platform conditions. Target regions are labeled, enabling the model to output the location and extent of the target region. When using rule-based methods, suspected target regions can be extracted first based on local depth abrupt changes and region connectivity, and then non-real targets can be filtered out by combining factors such as target size, ground clearance, and continuous frame stability.Regardless of whether new target detection or target region detection is used, the final determination must be made as to whether the overlap between the new target or target region and the safe area exceeds a preset threshold. The overlap range refers to the degree to which the target occupies the safe area in space, which can be characterized by the two-dimensional projected overlap area, the three-dimensional voxel overlap volume, the overlap length or overlap ratio between the target bounding box and the safe area. For example, the proportion of points in the target point cloud falling into the safe area can be calculated, or the intrusion depth of the target bounding box and the safe area in the closing direction can be calculated. The preset threshold must have a clear source, preferably determined by a combination of TOF camera ranging error, background noise level, mechanical safety margin of the vehicle door, and the size of the smallest dangerous target. For example, the upper limit of the false overlap range caused by random noise can be statistically determined using empty scene samples, and then combined with the lower limit of the actual overlap range formed by the smallest targets requiring protection, such as human fingers, clothing corners, and bag edges, within the safe area. A threshold that balances safety and false detection rate can be selected between these two. Value; if an implementable example is needed, it can be stated as follows: when the overlap ratio of the two-dimensional projection of the target and the safe area reaches 10% to 25%, or the intrusion depth of the target in the direction of the door closing reaches 10mm to 30mm, or the number of effective voxels occupied by the target in the safe area exceeds a preset voxel threshold, it is determined that there are personnel or objects in the safe area; furthermore, to avoid misjudgment caused by single-frame noise, abnormal reflection, or short-term occlusion, a continuous frame confirmation mechanism can be added, that is, only when the newly added target or target area meets the condition that the overlap range exceeds the preset threshold for 2 to 5 consecutive frames, can the final judgment result be output; through the above method, the detection unit in the safe area can identify the real intrusion target in the door closing path using two technical paths: background model comparison and depth target detection, and distinguish the situation of being close but not dangerous from the situation of having entered the anti-pinch risk zone by the overlap range threshold, thereby more reliably determining whether there are personnel or objects in the safe area.

[0081] In some embodiments of this application, during the door closing process, if there are people or objects in the safe area, or if a target outside the safe area intends to get on or off the vehicle, the control execution unit controls the door to stop closing; after the door stops closing, if no people, objects, or intentions to get on or off the vehicle are detected in consecutive preset frames, the control unit controls the door to resume closing or remain open and ready.

[0082] Specifically, during the door closing process, the control execution unit receives judgment results from the detection unit within the safe area and the intent detection unit outside the area, and combines this with the validity confirmation information output by the timing verification unit to execute control on the door opening and closing unit to stop closing, resume closing, or remain open and standby. Controlling the door to stop closing does not simply mean cutting off the closing action, but rather outputting a stop closing control command to the drive mechanism based on the execution capability of the door operator controller, causing the door to immediately decelerate and stop in the currently closed state. If necessary, it can also perform a small-distance retraction or reopening according to the safety strategy to prevent the door's inertia from continuing to intrude into the danger zone. This can be achieved by cutting off the closing drive signal, outputting a braking control signal, switching the door operator to a holding state, or triggering a reverse drive command. The relevant control methods can be implemented compatible with existing door operator control logic. When there are people or objects within the safe area, it indicates that the target has entered or occupied the danger space corresponding to the door closing path, and the control execution unit should prioritize outputting a stop closing command to prevent the door from continuing to move towards the closed end. When a target outside the safe area has not yet entered the safe area, but trajectory prediction determines that it intends to get on or off the vehicle, the control execution unit... The unit should also stop closing the door in advance to achieve active anti-pinch rather than responding after contact. To avoid frequent false stops caused by false detections in a single frame, the control trigger condition is preferably established as the detection result being valid and confirmed by the timing verification unit. Validity can be understood as the target trajectory conforming to positional continuity, velocity continuity, and reasonable motion in continuous time, eliminating false targets caused by abnormal reflections, noise jumps, or short-term occlusion. After the door stops closing, the control execution unit does not immediately resume action but continuously monitors whether personnel, objects, and intentions to get on or off are still detected in subsequent preset frames. "Not detected for consecutive preset frames" means that within a certain number of consecutive frame detection cycles, no personnel or object targets meeting the presence criteria reappear within the safe area, and no approaching targets meeting the boarding / alighting intent criteria reappear outside the area. Furthermore, the relevant results remain stable and disappear after time-series verification. The number of consecutive preset frames should have a clear source and is preferably set comprehensively based on the TOF camera frame rate, detection algorithm stability, door operator response latency, and false detection suppression requirements. For example, under conditions of 20 frames / second to 30 frames / second, the number of consecutive preset frames can be set to 2 to 8 frames, corresponding to approximately 0.1 to 0 seconds.A 4-second safety confirmation window is provided. If the system environment is noisy, the frame rate can be increased; if faster resumption of traffic is needed, the frame rate can be decreased while ensuring safety. If no personnel, objects, or intentions to board or alight are detected within consecutive preset frames, the control unit can choose to resume closing or remain open and ready based on the current door position, train station status, and upper-level control strategy. Resuming closing means continuing the interrupted door-closing process after confirming the risk has been eliminated, causing the door to move back towards closing from its current stopping position. Remaining open and ready means maintaining the open state before the risk has been eliminated or even if it has been temporarily eliminated. However, if the system determines that there is still a high possibility of re-intrusion, the door should remain in its current open position or its safe retraction position, awaiting further confirmation. The switching conditions for these two strategies should also have clear justifications. For example, if the remaining travel distance after the target disappears is short, no new approaching targets are detected outside the area, and the platform departure conditions are met, the door can be closed first. Conversely, if the target has just left the safe area, there are still crowds outside the area, multiple attempts to board or alight are triggered within a short period, or the system is operating under conditions of high passenger flow, then maintaining the open standby state is preferable. If a more specific implementation plan is needed, it can be provided below. Note: When no target is detected within the safe area for 3 to 5 consecutive frames after the door stops closing, and no valid approach trajectory towards the door area is detected outside the safe area within the same number of consecutive frames, the door will resume closing. If an approaching target is continuously detected outside the safe area within a preset waiting time, or if multiple repeated triggers occur within the last few seconds, the door will remain open and ready to close. The preset waiting time can be set according to the train's station stop time, passenger flow organization needs, and door operator control strategy, for example, it can be 0.5 seconds to 3 seconds. Furthermore, to prevent the door from frequently switching between stopping and resuming, a control mechanism can also be set. The hysteresis mechanism, where the release threshold for resuming closure is higher than the initial threshold for triggering stop, or a minimum holding stop time is added before resuming closure (e.g., 0.2 to 1.0 seconds), improves control stability. Through this method, the control execution unit can promptly prevent the door from continuing to close when it detects a dangerous target within the door closing path or a significant approaching trend for boarding or alighting outside the closing path. After the risk is eliminated, it can conditionally resume closure or remain open based on continuous detection results and the control strategy, thus balancing the safety of the door's active anti-pinch mechanism, control stability, and train operation efficiency.

[0083] See Figure 2 As shown, this application proposes a safety control method for high-speed rail sliding doors integrating a TOF camera, the specific steps of which include:

[0084] S1: Acquire depth image matrix of the high-speed train door area using a TOF camera;

[0085] S2: Determine the safe zone corresponding to the door closing path based on the door's mechanical structure parameters, door closing path, and real-time door position;

[0086] S3: During the closing process of the car door, the presence of people or objects in the safe area is detected based on the depth image matrix or the 3D point cloud generated by the depth image matrix;

[0087] S4: Detect and track targets outside the safe zone across frames to determine whether the target has the intention to get on or off the vehicle towards the door area;

[0088] S5: Perform trajectory physical rationality analysis on the detection results within a continuous time window;

[0089] S6: When there are people or objects in the safe area, or when a target outside the safe area intends to get on or off the vehicle, and the trajectory physical rationality analysis result is valid, control the door to stop closing.

[0090] Understandably, step S1 is used to acquire raw depth perception data of the door area, providing basic input for subsequent spatial analysis; step S2 is used to establish a safe area corresponding to the current door closing state by combining the door's mechanical structure, closing path, and real-time position, giving subsequent detection targets and risk assessments a clear spatial constraint range; step S3 focuses on determining whether there are already people or objects in the dangerous space corresponding to the door closing path, i.e., identifying situations where people have entered the risk zone; step S4 focuses on early identification of targets that are still outside the safe zone but are moving towards the door area, i.e., identifying situations where people are about to enter the risk zone, thus enabling this application to achieve both anti-pinch detection of already intruded targets and forward-looking prediction of potential intruded targets; step S5 is used to perform continuous time-dimensional verification of the detection results obtained in steps S3 and S4, eliminating errors caused by depth noise, abnormal reflections, short-term occlusion, or occasional trajectory jumps. The judgment ensures that the final output detection result conforms to objective motion laws; step S6 is to execute door control based on the aforementioned spatial detection, intent recognition, and timing verification. When there are people or objects in the safe area, or when a target outside the safe area has a clear intention to get on or off the train, and this conclusion is confirmed to be valid by trajectory physical rationality analysis, the door is immediately controlled to stop closing to prevent the door from continuing to move towards the danger zone. Thus, the above method does not only make a passive response after the door comes into contact with an obstacle, but during the door closing process, based on the depth image matrix collected by the TOF camera, the dynamic safe area setting results, the existence of targets in the area, the approach trend of targets outside the area, and the trajectory rationality analysis results within a continuous time window, a complete closed-loop processing process from data acquisition, spatial modeling, target detection, behavior prediction to control execution is formed, thereby improving the active anti-pinch capability, detection accuracy, and control reliability of high-speed rail sliding doors in complex boarding and alighting scenarios.

[0091] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A safety control device for a high-speed rail sliding door integrating a TOF camera, comprising a door opening and closing unit for controlling the opening and closing of the door, characterized in that, Also includes: The depth data acquisition unit is used to acquire a depth image matrix of the high-speed train door area using a TOF camera. The safety zone setting unit is used to determine the safety zone corresponding to the door closing path based on the door's mechanical structure parameters, the door closing path, and the door's real-time position. The motion state sensing unit is used to acquire the motion state signal of the door and identify whether the door is in the closing process; The detection unit within the safe area is connected to the depth data acquisition unit, the safe area setting unit, and the motion state sensing unit. It is used to detect whether there are people or objects in the safe area based on the depth image matrix or the three-dimensional point cloud generated by the depth image matrix during the door closing process, and output a stop closing control command when the detection result is that there are people or objects. An outside-area intent detection unit, connected to the depth data acquisition unit and the safe zone setting unit, is used to detect and track targets outside the safe zone across frames, determine whether the target has the intent to get on or off the vehicle towards the door area, and output a stop door closing control command when the determination result is that the target has the intent to get on or off the vehicle. A timing verification unit, connected to the detection unit within the safe area and the intent detection unit outside the area, is used to perform trajectory physical rationality analysis on the detection results within a continuous time window; The control execution unit is connected to the detection unit within the safe area, the intention detection unit outside the area, the timing verification unit, and the door opening and closing unit, and is used to control the door to stop closing when the timing verification unit confirms that the detection result is valid.

2. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 1, characterized in that, The safe zone setting unit determines the safe zone in the following manner: The key area boundaries on the door closing path are determined based on the mechanical structure characteristics of the door; the safety sensitivity parameters of each area are dynamically adjusted according to the real-time movement speed of the door; a spatial coordinate-safety sensitivity mapping table is generated; and the door closing area is divided into high-sensitivity sub-regions, medium-sensitivity sub-regions, and low-sensitivity sub-regions according to the spatial coordinate-safety sensitivity mapping table, wherein the high-sensitivity sub-region is the core monitoring area of ​​the safety area.

3. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 2, characterized in that, The detection unit within the safe area and the intent detection unit outside the area include: A point cloud generation subunit is used to generate a three-dimensional point cloud based on the depth image matrix. The spatial region mapping sub-unit is used to determine the spatial sub-region to which each point cloud data belongs based on the spatial coordinate-security sensitivity mapping table. The target extraction subunit is used to perform cluster analysis or target detection processing on the three-dimensional point cloud to extract personnel targets and object targets; The trajectory association subunit is used to associate data between personnel targets and object targets in consecutive frames to form the target motion trajectory; The result output subunit is used to output the judgment result of the presence of the target within the safe area and the judgment result of the intention to get on or off the vehicle outside the area based on the target's motion trajectory.

4. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 3, characterized in that, The target extraction subunit performs the following operations: Construct a hierarchical spatial index structure for the depth image matrix; based on the spatial coordinate-security sensitivity mapping table, add hierarchical index depth to the point cloud data within the high-sensitivity sub-region; identify candidate change regions based on the spatial gradient change features of the depth values; Fine segmentation is performed on regions exceeding a preset gradient threshold; three-dimensional point cloud clustering is performed on the candidate change regions to obtain the spatial boundary parameters, volume parameters, and position parameters of the candidate targets.

5. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 4, characterized in that, The preset gradient threshold is determined through statistical analysis of gradient distribution in historical scenes; areas exceeding the preset gradient threshold are synchronously marked as key anti-pinch areas of concern; when the number of cluster points, target volume parameters, or target occupied space range within the key anti-pinch areas of concern exceed the corresponding preset threshold, it is determined that there are people or objects.

6. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 5, characterized in that, The trajectory association subunit performs the following operations: Correlate the centroid positions of the target in consecutive frames; Calculate the target's displacement vector, velocity vector, and motion component toward the door area; predict the target's motion path within a preset time period based on the target's current pose and historical trajectory; when the predicted motion path intersects with the safe area, determine that the target intends to get on or off the vehicle.

7. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 6, characterized in that, The criteria for determining the intention to get on or off the vehicle outside the designated area include: The target's initial position is outside the safe zone; the target maintains a movement trend toward the door area within consecutive preset frames; the minimum distance between the target's predicted trajectory and the safe zone is less than a preset distance threshold; or the target will enter the safe zone within a preset time.

8. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 3, characterized in that, The detection unit within the safe area determines the presence of people or objects within the safe area using the following methods: The point cloud clustering results within the safe area are compared with the background model to determine new targets; or target detection is performed on the depth image matrix to obtain the target region. When the overlap between the newly added target or the target area and the safe area exceeds a preset threshold, it is determined that there are people or objects in the safe area.

9. The high-speed rail sliding door safety control device integrating a TOF camera according to claim 8, characterized in that, During the door closing process, if there are people or objects in the safe area, or if a target outside the safe area intends to get on or off the vehicle, the control execution unit controls the door to stop closing; after the door stops closing, if no people, objects, or intentions to get on or off the vehicle are detected in consecutive preset frames, the control unit controls the door to resume closing or remain open and ready.

10. A method for safety control of high-speed rail sliding doors integrating a TOF camera, applied to the high-speed rail sliding door safety control device integrating a TOF camera as described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Acquire depth image matrix of the high-speed train door area using a TOF camera; S2: Determine the safe zone corresponding to the door closing path based on the door's mechanical structure parameters, door closing path, and real-time door position; S3: During the closing process of the car door, the presence of personnel or objects in the safe area is detected based on the depth image matrix or the three-dimensional point cloud generated by the depth image matrix. S4: Detect and track targets outside the safe area across frames to determine whether the target has the intention to get on or off the vehicle towards the door area; S5: Perform trajectory physical rationality analysis on the detection results within a continuous time window; S6: When there are people or objects in the safe area, or when a target outside the safe area intends to get on or off the vehicle, and the trajectory physical rationality analysis result is valid, control the door to stop closing.

Images