A Low Event Rate Sensing and Interception Method for Vehicle Interception

Abstract

This invention discloses a low-event-rate perception and interception method for vehicle interception, relating to the field of edge intelligent vehicle interception technology. The method includes the following steps: S1, collecting and filtering datasets using sensors; S2, performing target detection and tracking using deep learning based on the filtered datasets; S3, outputting interception results based on the filtered datasets and the target detection and tracking results; S4, forming a closed-loop control link and executing the interception results. This invention employs a multimodal low-power initial screening and hierarchical wake-up mechanism. Normally, only milliwatt-level sentinel links are maintained to avoid the AI ​​core / high-performance core being constantly running. Upon triggering, early warning preheating and hierarchical wake-up reduce startup overhead, achieving an optimal balance between long-term endurance and rapid response, reducing the risk of false interception due to target jumps, and ensuring the stability of closed-loop interception.

Description

Technical Field

[0001] This invention relates to the field of edge intelligent vehicle interception technology, specifically to a low-event-rate perception and interception method for vehicle interception. Background Technology

[0002] With the increasing demands for urban traffic safety, vehicle detection and interception systems are playing an increasingly important role in road monitoring and vehicle control. Traditional vehicle detection and interception systems mainly rely on single-modal sensing devices, such as video surveillance and radar detection, which often face problems such as insufficient detection accuracy, high system power consumption, and response delays in complex environments.

[0003] Vehicle interception systems are typically deployed in critical areas for security, protection of vital facilities, or specific controlled zones. Their operational characteristics are characterized by "long-term standby time + instantaneous high-load response." Traditional solutions either continuously maintain high-performance visual inference links to achieve lower response latency, but this leads to consistently high average power consumption and insufficient battery life; or they rely on deep sleep and rule-based triggering to reduce power consumption, but wake-up, initialization, and inference startup after triggering introduce additional latency. Furthermore, single-modal or single-threshold triggering has a high false alarm rate in complex environments, easily resulting in resource waste and even the risk of false interception.

[0004] At the same time, the output of vehicle interception is not a simple identification result, but a control command that directly drives the actuator. It is required to maintain the target identity and state estimation stably under conditions such as occlusion, nonlinear motion, and multi-target interference, and form a perception-decision-execution closed loop. Summary of the Invention

[0005] To address the above problems, this invention proposes a low-event-rate perception and interception method for vehicle interception.

[0006] The technical solution of this invention is: a low-event-rate perception and interception method for vehicle interception, comprising the following steps:

[0007] S1. Collect data using sensors and then filter it;

[0008] S2. Based on the selected dataset, use deep learning to perform object detection and tracking;

[0009] S3. Based on the filtered dataset and the target detection and tracking results, output the interception results;

[0010] S4. Form a closed-loop control link and execute the interception result.

[0011] Further, step S1 includes the following sub-steps:

[0012] S11. Collect thermal infrared data and microwave radar data using sensors;

[0013] S12. Weighted fusion of thermal infrared data and microwave radar data is performed to obtain the fusion intensity;

[0014] S13. When the fusion intensity is less than the wake-up threshold, proceed to S2.

[0015] Furthermore, in S12, the fusion intensity The expression is:

[0016] ;

[0017] in, This represents the weighting coefficients for the thermal infrared sensor channels. This represents the weighting coefficients for the microwave radar sensor channels. This represents the normalized output intensity value of the passive infrared sensor. This represents the normalized output intensity value of the microwave radar sensor.

[0018] Furthermore, S2 includes the following sub-steps:

[0019] S21. Preprocess the image frames captured by the RGB camera, including image size normalization, pixel value normalization and channel standardization, to obtain a standardized input image;

[0020] S22. Obtain the pixel-level depth map of the current frame, perform noise reduction and hole filling processing, and generate a depth mask based on the effective depth range.

[0021] S23. Input the standardized input image into the target detection model based on a deep neural network, and output a set of candidate target detection boxes;

[0022] S24. Based on the depth mask, extract the depth mask coverage rate within the corresponding area of ​​each detection box, and remove detection boxes with a depth mask coverage rate lower than the depth validity threshold to obtain a set of depth valid detection boxes.

[0023] S25. Perform non-maximum suppression on the set of effective depth detection boxes to obtain the refined detection result set of the current frame;

[0024] S26. Perform multi-target tracking on the current frame's refined detection result set and perform cross-frame ID association;

[0025] S27. Assign a unique target ID to the trajectories that are associated within several consecutive frames, and calculate the three-dimensional spatial position estimate of the target to complete multi-target detection.

[0026] Furthermore, in S22, the depth mask The expression is:

[0027] ;

[0028] in, This represents the preset minimum effective detection distance threshold. Represents a pixel-level depth map. This indicates the preset maximum effective detection distance threshold. The horizontal coordinate of an image pixel. The vertical coordinate of an image pixel;

[0029] In S23, the expression for the candidate object detection box set B is:

[0030] B={ , ,..., };

[0031] in, This indicates the first detection box. This indicates the second detection box. Indicates the first One detection box;

[0032] In S24, depth mask coverage The expression is:

[0033] ;

[0034] in, Indicates the first One detection box, Indicates the first The width of each detection box, Indicates the first The height of each detection box.

[0035] Furthermore, in S26, trajectory prediction based on Kalman filtering and the Hungarian method are used to perform cross-frame ID association to complete multi-target tracking;

[0036] Trajectory state vector of trajectory prediction The expression is:

[0037] ;

[0038] in, The x-coordinate of the target center. The ordinate represents the center of the target. Represents the horizontal velocity component. Represents the vertical velocity component. Indicates the width of the target bounding box. Indicates the height of the target bounding box. Indicates transpose;

[0039] The Kalman prediction equation for Kalman filtering is expressed as follows:

[0040] ;

[0041] in, express Time-state pair The prior state prediction at time (i.e., the state estimate in the prediction phase). express The posterior state estimate at time (the best estimate after updating from the previous frame). This represents the prior error covariance matrix at time step of the state transition matrix. express The posterior error covariance matrix at time (a measure of uncertainty in the prediction phase). express The posterior error covariance matrix at time step (covariance updated in the previous frame). Indicates transpose. Represents the process noise covariance matrix;

[0042] The Kalman update equation for Kalman filtering is expressed as follows:

[0043] ;

[0044] ;

[0045] in, Indicates Kalman gain, Represents the observation matrix. Represents the observation noise covariance matrix. This represents the observations detected in the current frame.

[0046] Furthermore, in S27, the expression for the three-dimensional spatial location estimation is:

[0047] ;

[0048] ;

[0049] in, The x-coordinate represents the estimated position in three-dimensional space. The ordinate represents the estimated position in three-dimensional space. Indicates the first The x-coordinate of the center point of each target detection box in the image (pixel coordinates). Indicates the first The pixel ordinate of the center point of each target detection box in the image. Represents a pixel-level depth map. Indicates the x-coordinate of the camera's optical center. Indicates the ordinate of the camera's optical center. Indicates the mean depth of the target. Indicates the target bounding box region. The horizontal coordinate of an image pixel. The vertical coordinate of an image pixel. This is represented by the focal length of the first camera. This indicates the focal length of the second camera.

[0050] Furthermore, S3 includes the following sub-steps:

[0051] S31. Generate an interception decision output based on the filtered dataset and the target detection and tracking results;

[0052] S32. Based on the interception decision output, convert the corresponding target position, speed and threat level into motor servo drive commands;

[0053] S33. Use the motor servo drive command and the unique interception target ID as the interception result.

[0054] Furthermore, in S31, the interception decision output... The expression is:

[0055] ;

[0056] in, This represents the fusion weight of the AI ​​detection results. The weights represent the fusion weights of depth information. The fusion weights represent the fusion results of the sensor fusion. This represents the inference result from the deep learning object detection and tracking module. This indicates the output of depth information acquired and processed by a depth or distance sensor. This represents the real-time output result indicating the sensor fusion strength.

[0057] Furthermore, in S4, the closed-loop control link includes standby monitoring, suspected early warning, fusion detection, and execution interception.

[0058] The beneficial effects of this invention are:

[0059] (1) Controllable trade-off between average power consumption and response latency: Through multimodal low-power initial screening and graded wake-up mechanism, only milliwatt-level sentinel links are kept running under normal conditions to avoid AI cores / high-performance cores being always on. After triggering, the startup overhead is reduced through early warning warm-up and graded wake-up, so that the system achieves the optimal balance between long-term battery life and fast response.

[0060] (2) Significant improvement in the stability of false triggering and interception: Multimodal consistency, voting / hysteresis / cooling / timeout fallback and other mechanisms reduce repeated wake-ups and false alarms caused by environmental interference; the target ID and state estimation of the detection / tracking output are entered into the fusion decision and control layer to reduce the risk of false interception caused by target jump and ensure the stability of closed-loop interception. Attached Figure Description

[0061] Figure 1 A flowchart for a low-event-rate perception and interception method for vehicle interception;

[0062] Figure 2 This is a diagram of the overall architecture for a four-layer progressive perception and interception system.

[0063] Figure 3 This is an event-driven state transition diagram. Detailed Implementation

[0064] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0065] like Figure 1 As shown, the present invention provides a low-event-rate perception and interception method for vehicle interception, comprising the following steps:

[0066] S1. Collect data using sensors and then filter it;

[0067] S2. Based on the selected dataset, use deep learning to perform object detection and tracking;

[0068] S3. Based on the filtered dataset and the target detection and tracking results, output the interception results;

[0069] S4. Form a closed-loop control link and execute the interception result.

[0070] This invention is based on always-on monitoring with low-power sensors and achieves hierarchical wake-up and task layering through multimodal triggering and an event-driven state machine: During standby, only the low-power sentinel link is maintained; after triggering, the visual link and AI acceleration domain are activated level by level for detection and tracking; after fusion decision confirms the threat, the system enters the execution control closed loop and issues an interception command; when the target disappears or the risk decreases, it returns to the low-power state according to timeout and cooling rules. This method reduces average power consumption while ensuring controllable response latency after triggering and suppressing false triggering and false interception.

[0071] like Figure 2 As shown, this invention organizes the vehicle interception task into a four-level processing flow, with each level decoupled according to function and resource requirements to form a closed-loop feedback.

[0072] The first layer is a low-power sensor initial screening layer, using a PIR (passive infrared) sensor for thermal infrared sensing and MR (microwave radar) for coarse distance measurement. These two sensors fuse multi-sensor data to form a hierarchical wake-up criterion. When the weighted fusion result... The system status feedback and closed-loop process are triggered in a timely manner.

[0073] The second layer is the deep learning precision perception layer. The system wakes up the RGB input link by triggering a threshold, performs target detection and multi-target tracking based on deep neural networks, and simultaneously performs target feature extraction and trajectory prediction analysis. It outputs target ID association and trajectory management results, realizing ID-level tracking and jump suppression of suspicious targets.

[0074] The third layer is the multimodal fusion decision layer, which integrates AI detection results, sensor fusion output, and depth information into the decision module. Through reliability assessment and weight / threshold adaptation, it generates an interception decision output. It outputs the unique target ID and threat level.

[0075] The fourth layer is the execution control layer. Based on the decision output, it generates path planning (MPC strategy), motion control (PID hierarchical control), and execution drive (motor servo drive) commands. Through status feedback and anomaly handling modules (real-time pose correction, anomaly termination handling), it forms a closed-loop control link to ensure the safety and reliability of the interception process. This four-layer architecture achieves cross-layer linkage through system status feedback and closed-loop links. In standby mode, only the first layer remains operational to reduce average power consumption. Upon triggering, subsequent layers are woken up sequentially to ensure response latency and interception accuracy.

[0076] This invention provides an event-driven hierarchical wake-up state machine scheme. For example... Figure 3 As shown, the system defines four core states: S0 (standby monitoring), S1 (suspected warning), S2 (AI fusion detection), and S3 (execution interception). The states are connected in a closed loop through triggering conditions and fallback strategies.

[0077] In the S0 standby monitoring state, the system only runs the low-power core and the PIR / MR sensor. When the sensor fusion output meets the initial screening threshold, it enters the S1 state.

[0078] In the S1 suspected warning state, the system preheats the camera and input path and buffers sensor data for a short time. If the threshold conditions are met continuously, the AI ​​core is woken up and enters S2 through the "initial screening passed" path. If the conditions are not met continuously or environmental interference is detected, the system returns to S0 through the "false alarm fallback" path, thereby avoiding energy waste caused by a single noise trigger.

[0079] In the AI ​​fusion detection state of S2, the system wakes up the deep learning precise perception layer to perform target detection and tracking. When there is a stable target ID within K consecutive frames and the confidence level meets the threshold, it enters S3 through the dual paths of "wake up AI kernel" and "target feature". If no effective target is detected within the timeout period or the target confidence level decreases, it falls back to S0 through the "target lost" path.

[0080] In the S3 interception state, the system wakes up the high-performance core to execute multimodal fusion decision and control command generation, and sends it to the motor servo system through the execution control layer; when the interception task is completed or the target stops moving (ΔS=0), the system falls back to S0 to resume standby through the "risk removal" path.

[0081] This invention reduces the cold start latency of the AI ​​link through the early warning and warm-up mechanism of S1, suppresses false triggering through multi-level thresholds and persistence constraints (time window voting, hysteresis threshold, cooling time), and ensures that the system automatically degrades when there is no effective target to save energy through timeout fallback and target loss detection, thereby achieving the optimal trade-off between power consumption, response latency and false triggering rate.

[0082] In this embodiment of the invention, step S1 includes the following sub-steps:

[0083] S11. Collect thermal infrared data and microwave radar data using sensors;

[0084] S12. Weighted fusion of thermal infrared data and microwave radar data is performed to obtain the fusion intensity;

[0085] S13. When the fusion intensity is less than the wake-up threshold, proceed to S2.

[0086] In this embodiment of the invention, in S12, the fusion strength The expression is:

[0087] ;

[0088] in, This represents the weighting coefficients for the thermal infrared sensor channels. This represents the weighting coefficients for the microwave radar sensor channels. This represents the normalized output intensity value of the passive infrared sensor. This represents the normalized output intensity value of the microwave radar sensor.

[0089] The weighting coefficients of the thermal infrared sensor channel are used to adjust the contribution ratio of the PIR signal in the fusion result; the weighting coefficients of the microwave radar sensor channel are used to adjust the contribution ratio of the microwave radar signal in the fusion result. and satisfy And all are greater than zero. The normalized output intensity value of a passive infrared (PIR) sensor reflects the change in thermal infrared radiation within the monitoring area; the normalized output intensity value of a microwave radar (MR) sensor reflects the motion intensity and distance information of targets within the monitoring area.

[0090] In this embodiment of the invention, S2 includes the following sub-steps:

[0091] S21. Preprocess the image frames captured by the RGB camera, including image size normalization, pixel value normalization and channel standardization, to obtain a standardized input image;

[0092] S22. Obtain the pixel-level depth map of the current frame, perform noise reduction and hole filling processing, and generate a depth mask based on the effective depth range.

[0093] S23. Input the standardized input image into the target detection model based on a deep neural network, and output a set of candidate target detection boxes;

[0094] S24. Based on the depth mask, extract the depth mask coverage rate within the corresponding area of ​​each detection box, and remove detection boxes with a depth mask coverage rate lower than the depth validity threshold to obtain a set of depth valid detection boxes.

[0095] S25. Perform non-maximum suppression on the set of effective depth detection boxes to obtain the refined detection result set of the current frame;

[0096] S26. Perform multi-target tracking on the current frame's refined detection result set and perform cross-frame ID association;

[0097] S27. Assign a unique target ID to the trajectories that are associated within several consecutive frames, and calculate the three-dimensional spatial position estimate of the target to complete multi-target detection.

[0098] In this embodiment of the invention, in S22, the depth mask The expression is:

[0099] ;

[0100] in, This represents the preset minimum effective detection distance threshold. Represents a pixel-level depth map. This indicates the preset maximum effective detection distance threshold. The horizontal coordinate of an image pixel. The vertical coordinate of an image pixel;

[0101] In S23, the expression for the candidate object detection box set B is:

[0102] B={ , ,..., };

[0103] in, This indicates the first detection box. This indicates the second detection box. Indicates the first One detection box;

[0104] In S24, depth mask coverage The expression is:

[0105] ;

[0106] in, Indicates the first One detection box, Indicates the first The width of each detection box, Indicates the first The height of each detection box.

[0107] In this embodiment of the invention, in S26, trajectory prediction based on Kalman filtering and the Hungarian method are used to perform cross-frame ID association to complete multi-target tracking;

[0108] Trajectory state vector of trajectory prediction The expression is:

[0109] ;

[0110] in, The x-coordinate of the target center. The ordinate represents the center of the target. Represents the horizontal velocity component. Represents the vertical velocity component. Indicates the width of the target bounding box. Indicates the height of the target bounding box. Indicates transpose;

[0111] The Kalman prediction equation for Kalman filtering is expressed as follows:

[0112] ;

[0113] in, express Time-state pair The prior state prediction at time (i.e., the state estimate in the prediction phase). express The posterior state estimate at time (the best estimate after updating from the previous frame). This represents the prior error covariance matrix at time step of the state transition matrix. express The posterior error covariance matrix at time (a measure of uncertainty in the prediction phase). express The posterior error covariance matrix at time step (covariance updated in the previous frame). Indicates transpose. Represents the process noise covariance matrix;

[0114] The Kalman update equation for Kalman filtering is expressed as follows:

[0115] ;

[0116] ;

[0117] in, Indicates Kalman gain, Represents the observation matrix. Represents the observation noise covariance matrix. This represents the observations detected in the current frame.

[0118] In this embodiment of the invention, in S27, the expression for the three-dimensional spatial position estimation is:

[0119] ;

[0120] ;

[0121] in, The x-coordinate represents the estimated position in three-dimensional space. The ordinate represents the estimated position in three-dimensional space. Indicates the first The x-coordinate of the center point of each target detection box in the image (pixel coordinates). Indicates the first The pixel ordinate of the center point of each target detection box in the image. Represents a pixel-level depth map. Indicates the x-coordinate of the camera's optical center. Indicates the ordinate of the camera's optical center. Indicates the mean depth of the target. Indicates the target bounding box region. The horizontal coordinate of an image pixel. The vertical coordinate of an image pixel. This is represented by the focal length of the first camera. This indicates the focal length of the second camera.

[0122] In this embodiment of the invention, S3 includes the following sub-steps:

[0123] S31. Generate an interception decision output based on the filtered dataset and the target detection and tracking results;

[0124] S32. Based on the interception decision output, convert the corresponding target position, speed and threat level into motor servo drive commands;

[0125] S33. Use the motor servo drive command and the unique interception target ID as the interception result.

[0126] In this embodiment of the invention, in S31, the interception decision output is... The expression is:

[0127] ;

[0128] in, This represents the fusion weight of the AI ​​detection results. The weights represent the fusion weights of depth information. The fusion weights represent the fusion results of the sensor fusion. This represents the inference result from the deep learning object detection and tracking module. This indicates the output of depth information acquired and processed by a depth or distance sensor. This represents the real-time output result indicating the sensor fusion strength.

[0129] The inference results from the deep learning object detection and tracking module include the bounding box, confidence score, ID, and trajectory status.

[0130] In this embodiment of the invention, S4 includes standby monitoring, suspected early warning, fusion detection, and execution interception.

[0131] After the system starts up, it first enters the normal low-power initial screening phase, where the low-power core continuously operates at the sampling period. Read the outputs of passive infrared and microwave radar, and calculate the sensor fusion strength. ;when Below the wake-up threshold When there are no suspicious targets, the system remains in S0 standby state for cyclic sampling. During this time, the visual link and AI acceleration domain are either off or in reduced-frequency standby mode, and the average power consumption remains in the milliwatt range. And it satisfies persistence constraints (e.g., within a time window). Internal trigger count ≥ When the system detects a suspicious target, it enters a rapid wake-up path: first, it activates the camera and input path for warm-up, and simultaneously wakes up the AI ​​acceleration core and high-performance decision core (this process corresponds to...). Figure 3(In the "initial screening passed" transfer from S0 to S1), the warm-up stage reduces the subsequent AI inference startup latency from hundreds of milliseconds during cold start to less than 100 milliseconds by pre-initializing the visual link and caching historical sensor data.

[0132] After entering the deep learning precision perception layer, the system performs three processing steps in parallel: the depth information input module collects depth / distance data and outputs distance results. The multi-target tracking module performs detection and tracking on RGB input and outputs AI inference results. (Including target bounding box, confidence score, ID, and trajectory status), the initial screening results module maintains the sensor fusion output. Real-time updates. The three results enter the multimodal fusion decision layer, which first normalizes the three inputs and calculates the weighted fusion result, and then... With the interception threshold (Waterproofing warning index %) comparison: If If the target is determined to meet the interception conditions, the system enters the execution control layer through the "Identify Dangerous Target" path; if The system will continue to observe in the AI ​​fusion detection state, or fall back to S0 after a timeout. The innovation of this multimodal fusion mechanism lies in the fact that by introducing depth constraints and sensor consistency verification, it reduces the risk of misjudgment when relying solely on visual detection in the case of occlusion, changes in lighting, or target jumps, and significantly improves the robustness of interception decisions.

[0133] Once the system enters the execution control layer, it first executes the serial port transmission motion control command module, which converts the target position, speed, and threat level output by the fusion decision into motor servo drive commands (including steering angle, speed setpoint, and attitude adjustment) and sends them to the underlying actuators via the serial port. Simultaneously, it initiates the target interception module based on the fusion results, generating and continuously updating a real-time control sequence based on target trajectory prediction and path planning (MPC strategy). During the interception execution process, the system periodically checks the target stopping conditions. (Target stops or target has left the monitoring area) If the stopping condition is met, the process will enter the termination procedure via the "YES" branch and fall back to S0 standby; if the stopping condition is not met ("NO" branch), the process will continue to loop through control command updates and status monitoring until the interception task is completed or an abnormal rollback is triggered (e.g., continuous...). (Frame lack of valid target, communication anomaly, actuator failure, etc.). The innovation of this closed-loop control process lies in: through status feedback and anomaly handling mechanisms ( Figure 2 The fourth-level "Status Feedback and Anomaly Handling" module enables real-time monitoring and security degradation of the interception process, avoiding the risk of false interception or loss of control; at the same time, it uses hierarchical decision thresholds. Its configurable design allows the system to flexibly adjust its interception strategy according to different security level requirements, achieving a balance between false alarm rate and missed alarm rate.

[0134] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A low-event-rate perception and interception method for vehicle interception, characterized in that, Includes the following steps: S1. Collect data using sensors and then filter it; S2. Based on the selected dataset, use deep learning to perform multi-target detection and tracking; S3. Based on the filtered dataset and the results of multi-target detection and tracking, output the interception results; S4. Form a closed-loop control link and execute the interception result.

2. The low-event-rate perception and interception method for vehicle interception according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Collect thermal infrared data and microwave radar data using sensors; S12. Weighted fusion of thermal infrared data and microwave radar data is performed to obtain the fusion intensity; S13. When the fusion intensity is less than the wake-up threshold, proceed to S2.

3. The low-event-rate perception and interception method for vehicle interception according to claim 2, characterized in that, In S12, the fusion strength The expression is: ； in, This represents the weighting coefficients for the thermal infrared sensor channels. This represents the weighting coefficients for the microwave radar sensor channels. This represents the normalized output intensity value of the passive infrared sensor. This represents the normalized output intensity value of the microwave radar sensor.

4. The low-event-rate perception and interception method for vehicle interception according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Preprocess the image frames captured by the RGB camera, including image size normalization, pixel value normalization and channel standardization, to obtain a standardized input image; S22. Obtain the pixel-level depth map of the current frame, perform noise reduction and hole filling processing, and generate a depth mask based on the effective depth range. S23. Input the standardized input image into the target detection model based on a deep neural network, and output a set of candidate target detection boxes; S24. Based on the depth mask, extract the depth mask coverage rate within the corresponding area of ​​each detection box, and remove detection boxes with a depth mask coverage rate lower than the depth validity threshold to obtain a set of depth valid detection boxes. S25. Perform non-maximum suppression on the set of effective depth detection boxes to obtain the refined detection result set of the current frame; S26. Perform multi-target tracking on the current frame's refined detection result set and perform cross-frame ID association; S27. Assign a unique target ID to the trajectories that are associated within several consecutive frames, and calculate the three-dimensional spatial position estimate of the target to complete multi-target detection.

5. The low-event-rate perception and interception method for vehicle interception according to claim 4, characterized in that, In S22, the depth mask The expression is: ； in, This represents the preset minimum effective detection distance threshold. Represents a pixel-level depth map. This indicates the preset maximum effective detection distance threshold. Represents the x-coordinate of an image pixel. The vertical coordinate of an image pixel; In step S23, the expression for the candidate target detection box set B is: B={ , ,..., }； in, This indicates the first detection box. This indicates the second detection box. Indicates the first One detection box; In S24, the depth mask coverage The expression is: ； in, Indicates the first One detection box, Indicates the first The width of each detection box, Indicates the first The height of each detection box.

6. The low event rate perception and interception method for vehicle interception according to claim 4, characterized in that, In step S26, trajectory prediction based on Kalman filtering and the Hungarian method are used to perform cross-frame ID association to complete multi-target tracking. The trajectory state vector of the trajectory prediction The expression is: ； in, The x-coordinate of the target center. The ordinate represents the center of the target. Represents the horizontal velocity component. Represents the vertical velocity component. Indicates the width of the target bounding box. Indicates the height of the target bounding box. Indicates transpose; The expression for the Kalman prediction equation of the Kalman filter is as follows: ； in, express Time-state pair The predicted prior state value at time t. express The posterior state estimate at time t. This represents the prior error covariance matrix at time step of the state transition matrix. express The posterior error covariance matrix at time t. express The posterior error covariance matrix at time t. Indicates transpose. Represents the process noise covariance matrix; The expression for the Kalman update equation of the Kalman filter is as follows: ； ； in, Indicates Kalman gain, Represents the observation matrix. Represents the observation noise covariance matrix. This represents the observations detected in the current frame.

7. The low-event-rate perception and interception method for vehicle interception according to claim 4, characterized in that, In step S27, the expression for the three-dimensional spatial position estimation is: ； ； in, The x-coordinate represents the estimated position in three-dimensional space. The ordinate represents the estimated position in three-dimensional space. Indicates the first The x-coordinate of the center point of each target detection box in the image (pixel coordinates). Indicates the first The pixel ordinate of the center point of each target detection box in the image. Represents a pixel-level depth map. Indicates the x-coordinate of the camera's optical center. Represents the ordinate of the camera's optical center. Indicates the average depth of the target. Indicates the target bounding box region. Represents the x-coordinate of an image pixel. The vertical coordinate of an image pixel. This is represented by the focal length of the first camera. This indicates the focal length of the second camera.

8. The low-event-rate perception and interception method for vehicle interception according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Based on the filtered dataset and the target detection and tracking results, generate the interception decision output; S32. Based on the interception decision output, convert the corresponding target position, speed and threat level into motor servo drive commands; S33. Use the motor servo drive command and the unique interception target ID as the interception result.

9. The low event rate perception and interception method for vehicle interception according to claim 8, characterized in that, In S31, the interception decision output The expression is: ； in, This represents the fusion weight of the AI ​​detection results. The weights represent the fusion weights of depth information. The fusion weights represent the fusion results of the sensor fusion. This represents the inference result from the deep learning object detection and tracking module. This indicates the output of depth information acquired and processed by a depth or distance sensor. This represents the real-time output result indicating the sensor fusion strength.

10. The low-event-rate perception and interception method for vehicle interception according to claim 1, characterized in that, In S4, the closed-loop control link includes standby monitoring, suspected early warning, fusion detection, and execution interception.

Images