Safety device used to reduce the risk of reversing collisions during autonomous driving testing.

Abstract

This invention discloses a safety device for reducing the risk of reversing collisions during autonomous driving testing, belonging to the field of autonomous driving testing safety technology. This device addresses the collision risk caused by rapidly moving dynamic obstacles and communication delays when vehicles are reversing in closed testing areas. It acquires millimeter-level position coordinates and heading angles through positioning units deployed on the test vehicle and the target object. Utilizing the C-V2X protocol, it establishes a real-time data link to share vehicle position, speed, reversing trajectory prediction data, and the coordinates of target object's outer contour feature points. Based on future minimum distance calculations and dynamic threshold identification, it generates a cooperative collision avoidance path using a distributed optimization algorithm. This allows high-risk vehicles to prioritize communication channels and issue deceleration, heading adjustment, and braking commands to surrounding vehicles. This device effectively improves reversing safety in multi-vehicle cooperative testing scenarios and is suitable for real-time collision risk warning and active intervention in closed testing areas for autonomous vehicles.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving testing safety technology, and in particular to a safety device for reducing the risk of reversing collisions during autonomous driving testing. Background Technology

[0002] In closed-course autonomous driving testing, collision risk control in reversing scenarios is one of the key technical challenges. Current collision avoidance systems for multi-vehicle cooperative reversing tests still have several shortcomings. First, positioning accuracy is insufficient for high-dynamic scenarios. Traditional positioning solutions often rely on a single sensor (such as GPS or low-cost LiDAR), but in closed environments, GPS signals are easily interfered with by building obstruction or reflection, leading to positioning errors increasing to the decimeter level, making it impossible to accurately capture the real-time relative position of the vehicle and obstacles. For example, during reversing, deviations in the distance calculation between the vehicle and nearby dynamic obstacles may cause misjudgments, especially under millimeter-level avoidance requirements, where existing systems struggle to support accurate trajectory prediction. Second, communication latency and channel contention are prominent issues. In dense testing scenarios, existing vehicle-to-everything (V2X) communication is prone to channel congestion due to multiple vehicles simultaneously transmitting data, and conventional priority allocation mechanisms cannot dynamically respond to sudden risks. For example, when the reversing trajectories of multiple test vehicles intersect, communication latency exceeding 30ms may cause cooperative commands to fail to execute synchronously, exacerbating the collision risk. Furthermore, existing collision prediction models are insufficiently adaptable to dynamic obstacles. Traditional methods employ fixed safety thresholds, failing to consider dynamic changes in target object dimensions (such as length and width parameters W and L), leading to threshold settings that are either too conservative or too lenient. For example, when obstacle sizes vary significantly, fixed thresholds may fail to accurately reflect actual safety boundaries, resulting in missed or false alarms. Furthermore, the global optimization capabilities of multi-vehicle cooperative collision avoidance strategies are limited. Existing systems are mostly based on single-vehicle decisions, lacking distributed coordination mechanisms, which can cause new conflicts to arise from local avoidance paths. For instance, after vehicle A adjusts its path to avoid an obstacle, it may encroach on vehicle B's pre-defined safety area, and existing technologies struggle to coordinate such chain reactions in real time. Finally, the rapid movement of dynamic obstacles places higher demands on system real-time performance. Traditional trajectory prediction algorithms have high computational complexity, making it difficult to complete multi-target trajectory calculation and risk assessment within milliseconds in dense target scenarios, resulting in response delays. The root causes of these problems are: the data fusion algorithm for high-precision multi-source positioning lacks robustness in complex environments (such as varying lighting conditions and dense target objects); communication resource allocation lacks a risk-adaptive dynamic scheduling mechanism; the collision risk model does not deeply integrate the physical parameters and motion states of the target objects; and the distributed collaborative optimization algorithm struggles to balance real-time performance and global convergence. These deficiencies expose existing systems to risks such as missed collision warnings and asynchronous collision avoidance commands in multi-vehicle reversing tests in closed environments, hindering improvements in test safety and efficiency. Summary of the Invention

[0003] One objective of this invention is to address the risk of reversing collisions caused by insufficient positioning accuracy, communication delays, and poor adaptability to dynamic obstacles during multi-vehicle testing in closed environments. Existing systems cannot achieve millimeter-level positioning, dynamic threshold adaptive adjustment, and real-time response for multi-vehicle cooperative collision avoidance.

[0004] Another objective of this invention is to ensure the synchronous execution of coordinated deceleration, heading adjustment, and emergency braking commands in multi-vehicle scenarios, thereby solving the problem that traditional control commands lack a graded response mechanism, leading to delayed collision avoidance actions or escalated conflicts.

[0005] Another objective of this invention is to address the problem that existing path planning methods, which do not incorporate the vehicle's minimum turning radius and dynamic safety threshold, often generate infeasible paths by defining constraints for cooperative collision avoidance paths that balance safety and vehicle kinematic limitations.

[0006] Another objective of this invention is to improve positioning accuracy in complex environments through multi-source fusion positioning, thereby solving the problem that single sensors (such as GPS or LiDAR) are prone to failure when signals are blocked or environmental interference occurs, leading to positioning deviations.

[0007] Another objective of this invention is to address the problem that existing positioning systems lack dynamic anti-interference mechanisms and cannot quickly switch to backup positioning modes when GPS signals are lost or sensors fail, thus failing to maintain positioning stability.

[0008] Another objective of this invention is to optimize the robustness of visual SLAM algorithms in scenes with varying lighting conditions and dense target objects, and to address the problem that traditional visual SLAM relies on fixed feature weights and is prone to tracking failures or mapping errors in complex environments.

[0009] Another objective of this invention is to enable rapid recovery of localization when visual SLAM continuous tracking fails, thereby addressing the problem that existing systems rely on single sensor data and lack a multi-source collaborative repair mechanism, leading to the risk of localization interruption.

[0010] Another objective of this invention is to ensure communication reliability and real-time performance in high-density testing scenarios, and to solve the problem that the traditional C-V2X protocol cannot dynamically allocate resources when the channel is congested, resulting in data transmission delay or loss.

[0011] Another objective of this invention is to realize a mechanism for dynamically adjusting communication priority and command synchronization based on collision risk, thereby solving the problem that existing algorithms do not quantify risk values ​​as a basis for channel preemption and thus cannot achieve strong synchronization response at the millisecond level.

[0012] Another objective of this invention is to ensure strict timing of instruction execution through hardware-level interrupts and preload parameters, thereby solving the problem that traditional control signals rely on software layer processing and are difficult to meet the 20-25ms synchronization error tolerance.

[0013] To achieve the above objectives, the present invention adopts the following solution:

[0014] Safety devices used to reduce the risk of reversing collisions during autonomous driving testing include:

[0015] Positioning units are deployed on all test vehicles and risk targets. The positioning units acquire millimeter-level position coordinates and heading angles of each vehicle in real time.

[0016] The communication unit based on the C-V2X protocol constructs real-time data links between vehicles and between vehicles and targets based on the position coordinates and heading angles obtained by the positioning unit to share the following information: the real-time position, speed, and reversing trajectory prediction data of each vehicle, as well as the coordinates of the outer contour feature points and safety boundary parameters of the risky target.

[0017] The communication unit receives the reversing trajectory prediction data uploaded by all vehicles, as well as the coordinates of the outer contour feature points and safety boundary parameters of the risk targets, and performs the following operations:

[0018] Based on the reversing trajectory prediction data, the minimum future distance between each vehicle and between a vehicle and a risky target is calculated. If the minimum future distance is less than the dynamic threshold D, it is marked as a potential conflict. The dynamic threshold D ranges from 1.3 to 1.5 times the larger of the width and length of the risky target.

[0019] Using a distributed optimization algorithm, a cooperative collision avoidance path is dynamically generated based on the position and speed data of each vehicle. At the same time, a dynamic priority algorithm is used to calculate the collision risk value of each vehicle in real time based on the dynamically generated cooperative collision avoidance path and potential conflict data. Communication channel resources are preferentially allocated to the vehicle with the highest collision risk value, and synchronization adjustment commands are issued to its neighboring vehicles. The synchronization adjustment commands include deceleration, adjustment of heading angle and braking, and the neighboring vehicles are forced to respond and execute the synchronization adjustment commands within 20-25ms.

[0020] As a preferred approach, the specific response method for issuing synchronization adjustment instructions is as follows:

[0021] When issuing a deceleration control command, a coordinated deceleration command is sent to all associated vehicles via V2X communication, so that the driving speed of each vehicle is reduced synchronously to the preset safety value.

[0022] When issuing control commands to adjust the heading angle, the heading angle of each vehicle is dynamically adjusted based on the optimized path to ensure that its trajectory deviates from the safety boundary of the risky target.

[0023] When a braking control command is issued, if the path conflict cannot be eliminated, a cascaded braking protocol is triggered, and emergency braking is performed sequentially according to the distance of the vehicles from the conflict point, from closest to furthest.

[0024] As a preferred option, the constraints used when dynamically generating a cooperative collision avoidance path for each vehicle include:

[0025] The minimum spacing between the reversing paths of each vehicle is greater than or equal to the dynamic threshold D; and

[0026] The radius of curvature of the collision avoidance path is greater than or equal to the vehicle's minimum turning radius.

[0027] Preferably, the positioning unit includes a multi-source fusion positioning submodule, which integrates GPS, LiDAR point cloud matching, and visual SLAM algorithms to perform millimeter-level positioning through the following steps:

[0028] The initial position coordinates are obtained through GPS, and the surrounding environmental feature point cloud is scanned by LiDAR. The data is then matched in real time with a high-precision preset site map to correct position deviations. Based on the image data collected by the vehicle-mounted camera, a visual SLAM algorithm is used to generate a local dense map to calibrate the heading angle.

[0029] Preferably, the positioning unit also includes a dynamic anti-interference submodule, which performs the following operations when the signal is interfered with:

[0030] When GPS signal obstruction is detected, a cooperative positioning mode based on vehicle-to-vehicle (V2X) communication is initiated to receive positioning data from neighboring vehicles and calculate its own relative position using the least squares method.

[0031] If the lidar point cloud fails to match for at least 3 consecutive frames, it switches to pure visual inertial odometry (VIO) mode, fusing IMU data with visual feature tracking results to maintain positioning accuracy.

[0032] Real-time self-calibration is performed, and a multi-source data consistency check is executed every 30 seconds. When the positioning deviation between GPS and LiDAR exceeds 2 cm, a global relocation process is triggered to correct the positioning deviation.

[0033] Preferably, the method of using visual SLAM algorithms includes the following steps:

[0034] A dynamic feature weight allocation submodule is set up, which dynamically adjusts the extraction weights of visual feature points based on ambient light intensity and the density of risky targets. The specific operation mode is as follows:

[0035] If the ambient light intensity is strong light or low light, high-contrast edge feature points are extracted first, and the extraction ratio and weight of texture features are reduced to reduce texture dependence; when the density of risky objects reaches the high density threshold, the semantic segmentation network is enabled to identify passable areas, and only ground plane feature points are retained for SLAM mapping.

[0036] Configure a multi-sensor tightly coupled calibration submodule, which performs the following operations:

[0037] The local dense map generated by visual SLAM is registered frame by frame with the LiDAR point cloud, and the heading angle deviation is calculated by the ICP algorithm. When the heading angle deviation exceeds 0.5°, the extended Kalman filter (EKF) is constructed by fusing IMU angular velocity data for coupling calibration, and the calibrated heading angle is output.

[0038] Preferably, the visual SLAM algorithm also includes an abnormal state self-repair mechanism in its operating logic, which operates as follows:

[0039] When visual SLAM tracking fails for 5 consecutive frames, it switches to cooperative localization mode based on V2X communication, receives heading angle data from neighboring vehicles and performs weighted averaging, reverse-calculates its own pose using the outer contour feature points of the risky target, and resets the SLAM initialization parameters.

[0040] Preferably, the communication unit includes:

[0041] The dynamic channel optimization submodule performs the following operations: real-time monitoring of the communication channel load status and interference intensity; dynamic allocation of time-frequency resource blocks based on reinforcement learning algorithms; prioritizing data transmission for vehicles with collision risk values ​​above the risk threshold based on real-time calculated vehicle collision risk values; and in scenarios where there are ≥20 test vehicles and / or ≥5 risk targets within a 100-meter radius, initiating a multi-hop relay transmission mode: setting a central decision node, and transmitting vehicle data more than 200 meters away from the central decision node in segments through relay nodes of adjacent vehicles, with a maximum delay of ≤10ms per hop; the relay node selection criteria are a signal strength ≥-90dBm and remaining battery power >30%.

[0042] The data priority encryption submodule performs hierarchical encryption on transmitted data: collision warning data is encrypted using quantum key distribution (QKD), and position and velocity status data are encrypted using the lightweight national cryptographic algorithm SM4.

[0043] The central decision-making node dynamically adjusts the number of iterations of the path optimization algorithm based on communication latency and packet loss rate; if the packet loss rate is greater than 5%, a local collision avoidance mode is triggered, allowing each vehicle to perform emergency braking solely based on local sensor data.

[0044] As a preferred method, the collision risk value of the vehicle and the allocation of communication channel resources are as follows:

[0045] Collision risk value (Risk) is calculated based on dynamic weighting.

[0046]

[0047] Where θ is the vehicle heading angle deviation, R is the path curvature radius, α, β, γ are environmental adaptive weight coefficients with values ​​of α=8, β=2, γ=1, TTC is the collision time, v is the vehicle speed, and D is the dynamic threshold.

[0048] Communication channel resources are allocated through a channel preemption hierarchical strategy, specifically as follows:

[0049] Vehicles with a collision risk value of Risk ≥ 8 preempt the dedicated emergency channel, which has a bandwidth of 10MHz and a command transmission delay of ≤ 15ms; vehicles with a collision risk value of 5 ≤ Risk < 8 share the high-priority channel, which has a bandwidth of 5MHz and a delay of ≤ 25ms.

[0050] When forcing adjacent vehicles to execute synchronization adjustment commands, a strong synchronization mechanism is used. The mechanism operates as follows: after receiving the command, the adjacent vehicle triggers the control signal through a hardware interrupt, preloads coordination parameters in the reversing motor controller and steering system, sets the trajectory correction or braking response time to within 20-25ms, and the error tolerance is ≤0.2 meters.

[0051] As a preferred option, the specific operating modes of the strong synchronization mechanism include:

[0052] The hardware-level interrupt triggering module is directly connected to the underlying hardware interface between the reversing motor controller and the steering system. Priority interrupt signals are generated through programmable logic devices to forcibly interrupt the current control command and load the cooperative parameters.

[0053] The cooperative parameters are preloaded in the following way: the cooperative parameters generated by the dynamic priority algorithm are written into the controller and steering system before the command is issued by the cooperative parameter preloading module. The cooperative parameters include the target deceleration gradient, heading angle correction and braking pressure threshold.

[0054] The timing constraint module is set up to align the start time of instruction execution based on the inter-vehicle clock synchronization protocol, and monitors the entire delay from instruction reception to execution completion through a hardware timer, strictly controlling the delay within the range of 20-25ms.

[0055] An error tolerance verification module is set up, which provides real-time feedback of the actual displacement through the positioning unit after trajectory correction or braking is completed. If the deviation from the target displacement exceeds 0.2 meters, the emergency braking lock protocol is triggered.

[0056] The present invention has at least the following beneficial effects: (1) By using multi-source fusion positioning (GPS, LiDAR, visual SLAM) and dynamic anti-interference mechanisms (cooperative positioning, VIO mode switching), millimeter-level positioning accuracy is achieved in complex environments (such as GPS signal blockage, changes in illumination), and the heading angle deviation is calibrated in real time, which significantly improves the accuracy of the relative position calculation of vehicles and obstacles, and provides a reliable data foundation for collision prediction and collision avoidance path planning; (2) Based on the dynamic channel optimization and multi-hop relay transmission mode of the C-V2X protocol, combined with reinforcement learning algorithms, high-risk vehicle communication resources are allocated first, ensuring that the instruction transmission delay is ≤25ms in high-density test scenarios (≥20 vehicles / 100 meters), and data transmission security is ensured through hierarchical encryption (QKD, SM4), effectively solving the channel congestion and data loss problems of traditional V2X systems; (3) Multi-vehicle cooperative collision avoidance paths are dynamically generated using distributed optimization algorithms, and collision risk values ​​are quantified by dynamic priority algorithms, which prioritize the use of communication channels to issue instructions, and through Through layered control strategies such as cascading braking and heading angle correction, global synchronization of multi-vehicle actions is achieved (20-25ms response), avoiding chain conflicts caused by single-vehicle collision avoidance; (4) Based on the target object size (W, L), the safety threshold is dynamically adjusted, and combined with the collision time (TTC), heading angle deviation and other multi-parameter risk value calculation model, potential conflicts are accurately identified, reducing misjudgments caused by fixed thresholds and improving the adaptability and reliability of collision warning; (5) The underlying controller is directly controlled through the hardware interrupt module, and the collaborative parameters (deceleration gradient, braking pressure threshold) are preloaded. Combined with the clock synchronization protocol and hardware timer monitoring, the delay of the entire instruction execution is strictly controlled within 20-25ms, and the error tolerance is ≤0.2 meters, which greatly improves the real-time control and execution consistency in emergency scenarios; When visual SLAM fails or the sensor is abnormal, it quickly switches to V2X collaborative positioning mode and reverse calculates the pose based on the outer contour features of the risk target object to achieve self-recovery of positioning, ensuring the continuity of the test process and the fault tolerance of the system, and reducing the need for manual intervention.

[0057] These effects collectively address core challenges in multi-vehicle reversing tests in closed environments, such as positioning deviations, communication delays, path conflicts, and sudden environmental changes, significantly improving the safety and efficiency of autonomous driving testing. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of a safety device for reducing the risk of reversing collisions during autonomous driving testing, provided by the present invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0060] Safety devices used to reduce the risk of reversing collisions during autonomous driving testing include:

[0061] Positioning units are deployed on all test vehicles and risk targets. The positioning units acquire millimeter-level position coordinates and heading angles of each vehicle in real time.

[0062] The communication unit based on the C-V2X protocol constructs real-time data links between vehicles and between vehicles and targets based on the position coordinates and heading angles obtained by the positioning unit to share the following information: the real-time position, speed, and reversing trajectory prediction data of each vehicle, as well as the coordinates of the outer contour feature points and safety boundary parameters of the risky target.

[0063] The communication unit receives the reversing trajectory prediction data uploaded by all vehicles, as well as the coordinates of the outer contour feature points and safety boundary parameters of the risk targets, and performs the following operations:

[0064] Based on the reversing trajectory prediction data, the minimum future distance between each vehicle and between a vehicle and a risky target is calculated. If the minimum future distance is less than the dynamic threshold D, it is marked as a potential conflict. The dynamic threshold D ranges from 1.3 to 1.5 times the larger of the width and length of the risky target.

[0065] Using a distributed optimization algorithm, a cooperative collision avoidance path is dynamically generated based on the position and speed data of each vehicle. At the same time, a dynamic priority algorithm is used to calculate the collision risk value of each vehicle in real time based on the dynamically generated cooperative collision avoidance path and potential conflict data. Communication channel resources are preferentially allocated to the vehicle with the highest collision risk value, and synchronization adjustment commands are issued to its neighboring vehicles. The synchronization adjustment commands include deceleration, adjustment of heading angle and braking, and the neighboring vehicles are forced to respond and execute the synchronization adjustment commands within 20-25ms.

[0066] By employing millimeter-level positioning and real-time C-V2X communication, a vehicle-target end-to-end data sharing mechanism was established. Combined with dynamic thresholds (adaptively adjusted based on target size) for accurate identification of potential conflicts, and multi-vehicle cooperative collision avoidance was achieved through distributed optimization and dynamic priority algorithms. Its advantages are: 1) The dynamic threshold D = 1.3~1.5·max(W,L) comprehensively considers target size, avoiding false alarms or missed detections caused by fixed thresholds; 2) The distributed optimization algorithm generates collision avoidance paths globally, avoiding cascading conflicts caused by single-vehicle decisions; 3) The dynamic priority algorithm preempts channel resources based on risk values, ensuring high-risk commands are transmitted first, compressing command response time to 20-25ms, significantly reducing the impact of communication latency on safety.

[0067] In another technical solution, the specific response method for issuing synchronization adjustment instructions is as follows:

[0068] When issuing a deceleration control command, a coordinated deceleration command is sent to all associated vehicles via V2X communication, so that the driving speed of each vehicle is reduced synchronously to the preset safety value.

[0069] When issuing control commands to adjust the heading angle, the heading angle of each vehicle is dynamically adjusted based on the optimized path to ensure that its trajectory deviates from the safety boundary of the risky target.

[0070] When a braking control command is issued, if the path conflict cannot be eliminated, a cascaded braking protocol is triggered, and emergency braking is performed sequentially according to the distance of the vehicles from the conflict point, from closest to furthest.

[0071] This solution achieves refined execution of multi-vehicle coordinated actions through tiered control commands (deceleration, heading adjustment, and braking). Coordinated deceleration commands ensure that vehicle speeds are synchronously reduced to a safe range, avoiding secondary collisions caused by speed differences. Dynamic heading angle adjustment corrects the trajectory in real time based on the optimized path, ensuring vehicles deviate from the target's safe boundary. A cascaded braking protocol triggers braking based on the distance to the conflict point, reducing the interference of emergency braking on the overall test scenario. This solution addresses the lack of layered response in traditional commands. Especially when path conflicts cannot be eliminated, cascaded braking maximizes the reduction of collision energy and improves collision avoidance reliability in complex scenarios.

[0072] The constraints used when dynamically generating a collaborative collision avoidance path for each vehicle include: the minimum spacing between the reversing paths of each vehicle ≥ the dynamic threshold D; and the radius of curvature of the collision avoidance path ≥ the minimum turning radius of the vehicle.

[0073] By introducing path constraints (minimum spacing and radius of curvature), the generated collision avoidance path is ensured to meet both dynamic safety thresholds and vehicle kinematic constraints. The minimum spacing constraint avoids close-range risks caused by multiple vehicle paths intersecting; the radius of curvature constraint ensures path feasibility and prevents control failures due to insufficient turning radius. This solution overcomes the deficiency of separating safety and feasibility in traditional path planning, and is particularly suitable for high-precision reversing scenarios, reducing the frequency of path replanning and improving the overall system efficiency.

[0074] In another technical solution, the positioning unit includes a multi-source fusion positioning submodule, which integrates GPS, LiDAR point cloud matching, and visual SLAM algorithms to achieve millimeter-level positioning through the following steps:

[0075] The initial position coordinates are obtained through GPS, and the surrounding environmental feature point cloud is scanned by LiDAR. The data is then matched in real time with a high-precision preset site map to correct position deviations. Based on the image data collected by the vehicle-mounted camera, a visual SLAM algorithm is used to generate a local dense map to calibrate the heading angle.

[0076] Millimeter-level accuracy is achieved in complex environments through multi-source fusion positioning (GPS, LiDAR, and visual SLAM). GPS provides initial coordinates, LiDAR point cloud matching corrects map deviations, and visual SLAM further calibrates the heading angle; these three technologies complement each other to overcome the limitations of a single sensor. For example, when GPS signals are blocked, LiDAR and visual SLAM can maintain positioning continuity; in low-light environments, LiDAR point cloud matching prioritizes accuracy. This solution significantly improves the robustness of the positioning system, making it particularly suitable for scenarios with dense dynamic obstacles in enclosed spaces.

[0077] The positioning unit also includes a dynamic anti-interference submodule, which performs the following operations when the signal is interfered with:

[0078] When GPS signal obstruction is detected, a cooperative positioning mode based on vehicle-to-vehicle (V2X) communication is initiated to receive positioning data from neighboring vehicles and calculate its own relative position using the least squares method.

[0079] If the lidar point cloud fails to match for at least 3 consecutive frames, it switches to pure visual inertial odometry (VIO) mode, fusing IMU data with visual feature tracking results to maintain positioning accuracy.

[0080] Real-time self-calibration is performed, and a multi-source data consistency check is executed every 30 seconds. When the positioning deviation between GPS and LiDAR exceeds 2 cm, a global relocation process is triggered to correct the positioning deviation.

[0081] A dynamic anti-interference mechanism ensures the stability of the positioning system when sensors fail. The cooperative positioning mode utilizes V2X communication to acquire neighboring vehicle data and calculates relative position using the least squares method to compensate for GPS signal loss. The VIO mode fuses IMU and visual data when the LiDAR fails to maintain positioning accuracy. Real-time self-calibration logic periodically checks the consistency of multi-source data and triggers relocation to eliminate accumulated errors. This solution addresses the downtime risk of traditional positioning systems during signal interruptions, ensuring continuous and reliable operation in highly dynamic scenarios.

[0082] The steps involved in using visual SLAM algorithms are as follows:

[0083] A dynamic feature weight allocation submodule is set up, which dynamically adjusts the extraction weights of visual feature points based on ambient light intensity and the density of risky targets. The specific operation mode is as follows:

[0084] If the ambient light intensity is strong light or low light, high-contrast edge feature points are extracted first, and the extraction ratio and weight of texture features are reduced to reduce texture dependence; when the density of risky objects reaches the high density threshold, the semantic segmentation network is enabled to identify passable areas, and only ground plane feature points are retained for SLAM mapping.

[0085] Configure a multi-sensor tightly coupled calibration submodule, which performs the following operations:

[0086] The local dense map generated by visual SLAM is registered frame by frame with the LiDAR point cloud, and the heading angle deviation is calculated by the ICP algorithm. When the heading angle deviation exceeds 0.5°, the extended Kalman filter (EKF) is constructed by fusing IMU angular velocity data for coupling calibration, and the calibrated heading angle is output.

[0087] By employing dynamic feature weight allocation and tightly coupled multi-sensor calibration, the performance of visual SLAM in complex environments is optimized. The dynamic feature weight module adaptively adjusts the feature extraction strategy based on illumination and target density; for example, it relies on high-contrast edge features under strong light to reduce texture noise interference. The semantic segmentation network filters redundant features when targets are densely packed, improving mapping efficiency. The tightly coupled calibration module fuses multi-sensor data using the ICP algorithm and EKF to correct heading angle deviation in real time (with an accuracy of 0.5°), preventing SLAM mapping drift. This approach significantly improves the robustness and accuracy of visual SLAM in scenes with varying lighting and dense targets.

[0088] The visual SLAM algorithm also includes an abnormal state self-repair mechanism in its operating logic, which operates as follows:

[0089] When visual SLAM tracking fails for 5 consecutive frames, it switches to cooperative localization mode based on V2X communication, receives heading angle data from neighboring vehicles and performs weighted averaging, reverse-calculates its own pose using the outer contour feature points of the risky target, and resets the SLAM initialization parameters.

[0090] This solution addresses the localization interruption issue caused by continuous tracking failure in visual SLAM through an anomaly self-healing mechanism. When SLAM fails, the system switches to V2X cooperative localization mode, using a weighted average of neighboring vehicle heading angle data to calculate its own pose. Simultaneously, it calculates the pose in reverse based on the outer contour feature points of the risky target, resetting the SLAM parameters to quickly resume tracking. This approach avoids the reliance on manual intervention in abnormal states inherent in traditional SLAM systems, shortens the localization recovery time (<100ms in typical scenarios), and ensures the continuity of the testing process.

[0091] In another technical solution, the communication unit includes:

[0092] The dynamic channel optimization submodule performs the following operations: real-time monitoring of the communication channel load status and interference intensity; dynamic allocation of time-frequency resource blocks based on reinforcement learning algorithms; prioritizing data transmission for vehicles with collision risk values ​​above the risk threshold based on real-time calculated vehicle collision risk values; and in scenarios where there are ≥20 test vehicles and / or ≥5 risk targets within a 100-meter radius, initiating a multi-hop relay transmission mode: setting a central decision node, and transmitting vehicle data more than 200 meters away from the central decision node in segments through relay nodes of adjacent vehicles, with a maximum delay of ≤10ms per hop; the relay node selection criteria are a signal strength ≥-90dBm and remaining battery power >30%.

[0093] The data priority encryption submodule performs hierarchical encryption on transmitted data: collision warning data is encrypted using quantum key distribution (QKD), and position and velocity status data are encrypted using the lightweight national cryptographic algorithm SM4.

[0094] The central decision-making node dynamically adjusts the number of iterations of the path optimization algorithm based on communication latency and packet loss rate; if the packet loss rate is greater than 5%, a local collision avoidance mode is triggered, allowing each vehicle to perform emergency braking solely based on local sensor data.

[0095] By employing dynamic channel optimization and multi-hop relay transmission, the solution addresses communication congestion and latency issues in high-density testing scenarios. Reinforcement learning algorithms dynamically allocate resource blocks, prioritizing data transmission from high-risk vehicles. The multi-hop relay mode expands communication coverage (each hop latency ≤10ms), preventing data loss from distant vehicles. QKD and SM4 hierarchical encryption balance security and transmission efficiency. Furthermore, a local collision avoidance mode is triggered when the packet loss rate exceeds a threshold, ensuring basic security even during communication interruptions. In scenarios with vehicle density ≥20 vehicles / 100 meters, this solution still maintains an average communication latency below 15ms, significantly outperforming traditional V2X protocols.

[0096] In another technical solution, the method for calculating the vehicle's collision risk value and allocating communication channel resources is as follows:

[0097] Collision risk value (Risk) is calculated based on dynamic weighting.

[0098]

[0099] Where θ is the vehicle heading angle deviation, R is the path curvature radius, α, β, and γ are environmental adaptive weight coefficients with values ​​of α=8, β=2, and γ=1, TTC is the collision time, v is the vehicle speed, and D is the dynamic threshold; note that the values ​​of α, β, and γ can be finely adjusted according to the actual situation.

[0100] Communication channel resources are allocated through a channel preemption hierarchical strategy, specifically as follows:

[0101] Vehicles with a collision risk value of Risk ≥ 8 preempt the dedicated emergency channel, which has a bandwidth of 10MHz and a command transmission delay of ≤ 15ms; vehicles with a collision risk value of 5 ≤ Risk < 8 share the high-priority channel, which has a bandwidth of 5MHz and a delay of ≤ 25ms.

[0102] When forcing adjacent vehicles to execute synchronization adjustment commands, a strong synchronization mechanism is used. The mechanism operates as follows: after receiving the command, the adjacent vehicle triggers the control signal through a hardware interrupt, preloads coordination parameters in the reversing motor controller and steering system, sets the trajectory correction or braking response time to within 20-25ms, and the error tolerance is ≤0.2 meters.

[0103] Precise control of channel preemption and command synchronization is achieved by quantifying risk values. The risk value formula integrates multiple dimensions of parameters such as TTC, speed, and heading angle deviation to dynamically reflect the urgency of a collision. A hierarchical channel strategy (dedicated / shared channel) allocates bandwidth and latency resources according to risk level, ensuring that high-risk commands are transmitted first. Hardware-level interrupts and parameter preloading mechanisms compress command response time to 20-25ms, with an error tolerance of ≤0.2 meters. In sudden risk scenarios (such as a vehicle suddenly cutting into a reversing path), this solution can reduce the probability of collision to less than 1 / 5 of that of traditional systems.

[0104] The specific operation modes of the instruction strong synchronization mechanism include:

[0105] The hardware-level interrupt triggering module is directly connected to the underlying hardware interface between the reversing motor controller and the steering system. Priority interrupt signals are generated through programmable logic devices to forcibly interrupt the current control command and load the cooperative parameters.

[0106] The cooperative parameters are preloaded in the following way: the cooperative parameters generated by the dynamic priority algorithm are written into the controller and steering system before the command is issued by the cooperative parameter preloading module. The cooperative parameters include the target deceleration gradient, heading angle correction and braking pressure threshold.

[0107] The timing constraint module is set up to align the start time of instruction execution based on the inter-vehicle clock synchronization protocol, and monitors the entire delay from instruction reception to execution completion through a hardware timer, strictly controlling the delay within the range of 20-25ms.

[0108] An error tolerance verification module is set up, which provides real-time feedback of the actual displacement through the positioning unit after trajectory correction or braking is completed. If the deviation from the target displacement exceeds 0.2 meters, the emergency braking lock protocol is triggered.

[0109] Strict timing control of instruction execution is achieved through hardware-level interrupts and preloaded parameter mechanisms. The hardware interrupt module bypasses software layer latency and directly controls the underlying controller, ensuring an interrupt response time of ≤1ms. Preloaded parameters are written to the controller before instruction issuance, reducing processing time during execution. A clock synchronization protocol and hardware timer control the overall latency error within ±1ms. An error tolerance verification module provides real-time feedback on execution results, triggering secondary braking and locking when limits are exceeded. This solution improves the overall reliability of instruction execution to over 99.9%, making it particularly suitable for millisecond-level cooperative collision avoidance scenarios.

[0110] The working process of the positioning unit of this invention:

[0111] 1. GPS Initial Positioning: When the vehicle starts, the initial position coordinates (longitude and latitude) are obtained through the GPS module with an accuracy of centimeters.

[0112] 2. LiDAR point cloud matching: The LiDAR scans the surrounding environment 20 times per second to generate point cloud data, which is then matched in real time with a pre-stored high-precision site map. The ICP (Iterative Closest Point) algorithm is used to correct positional deviations, improving the positioning accuracy to the millimeter level.

[0113] 3. Visual SLAM Heading Calibration: The vehicle-mounted camera acquires environmental images at 30fps, runs a visual SLAM algorithm to extract ORB feature points, constructs a local dense map, and performs tight coupling and registration with the LiDAR point cloud to calculate the heading angle deviation. When the deviation exceeds 0.5°, the IMU (Inertial Measurement Unit) angular velocity data is fused, and the calibrated heading angle is output through an extended Kalman filter (EKF).

[0114] The communication unit of this invention operates as follows:

[0115] 1. Data sharing: Each vehicle broadcasts its own position, speed, and predicted reversing trajectory data for the next 3 seconds via C-V2X (generated based on a polynomial fitting algorithm).

[0116] 2. Potential Conflict Detection: After receiving all data, the central decision node calculates the minimum future distance between vehicles and between vehicles and targets. If the distance is less than the dynamic threshold D (D=1.3~1.5·max(W,L), where W and L are the length and width of the target), it is marked as a potential conflict.

[0117] 3. Cooperative Collision Avoidance Path Generation: A distributed optimization algorithm (such as Distributed Model Predictive Control, DMPC) is used to generate a global collision avoidance path based on vehicle position, speed, and collision markers. Algorithm constraints include path curvature ≥ vehicle minimum turning radius and path spacing ≥ threshold D.

[0118] 4. Command Issuance and Execution: A dynamic priority algorithm calculates the collision risk value (Risk) of each vehicle in real time. The vehicle with the highest risk value preempts the dedicated communication channel and issues deceleration, heading adjustment, or braking commands to associated vehicles. Commands are triggered via hardware-level interrupts, ensuring a response within 20-25ms.

[0119] Structural details of the invention:

[0120] Communication module hardware: It adopts a C-V2X chipset (such as Qualcomm 9150), supports direct communication via PC5 interface, and has a transmission latency of ≤10ms.

[0121] Path optimization computing node: Deployed on the vehicle's local ECU, using a multi-core processor (such as ARM Cortex-A72) to calculate collision avoidance paths in parallel.

[0122] The hierarchical control logic for deceleration, heading adjustment, and braking commands in this invention is as follows:

[0123] 1. Coordinated deceleration command: When a potential conflict is detected, the central decision-making node broadcasts a deceleration command via V2X, with a target speed of v. safe =0.5·v current (50% of current speed). After receiving the command, each vehicle's motor controller adjusts the reverse motor torque using a PID control algorithm, smoothly reducing the speed to v within 1 second. safe .

[0124] 2. Heading Angle Adjustment Command: If a conflict risk still exists after deceleration, a heading angle correction Δθ is generated based on the optimized path. For example, if the path requires the vehicle to deviate 0.5 meters to the right, then Δθ = arctan(0.5 / D) p ), D p This represents the current distance from the target object. The steering system executes Δθ via EPS (Electric Power Steering), with a stepper motor driving the steering rack; the angle error is ≤0.1°.

[0125] 3. Cascaded Braking Protocol: When path conflicts cannot be resolved, braking is triggered based on the vehicles' distance from the conflict point. The closest vehicle takes priority in emergency braking (deceleration -5m / s²), followed by subsequent vehicles decelerating at -3m / s², -1m / s², and so on. Braking commands are sent to the ESP (Electronic Stability Program) via the CAN bus, and the hydraulic unit builds up braking pressure within 50ms.

[0126] Structural implementation of the present invention:

[0127] Deceleration control module: Integrated into the vehicle VCU (vehicle controller), it includes a speed planning submodule (generating deceleration gradients based on polynomial curves) and a motor drive interface.

[0128] Steering actuator: The steering system is driven by a brushless DC motor (such as Maxon EC45), and the encoder provides real-time feedback on the steering angle.

[0129] Braking hydraulic unit: Bosch iBooster electro-hydraulic braking system, supports brake-by-wire, pressure response time ≤20ms.

[0130] The constraints of the cooperative collision avoidance path in this invention are achieved through the following steps:

[0131] 1. Minimum Spacing Constraint: During path planning, the minimum future spacing between multiple vehicle paths is calculated using a geometric collision detection algorithm (such as the GJK algorithm). If the spacing is less than D = 1.3·max(W,L), the path is replanned to ensure that the spacing is ≥ D.

[0132] 2. Curvature radius constraint: Based on the vehicle's minimum turning radius R min (e.g. R) min =3.5 meters), during the path generation stage, a Clothoid curve is used for smooth transition to ensure that the path curvature 1 / R ≤ 1 / R min The curvature k is calculated using the derivative formula in the Frenet coordinate system: k = (d²y / ds²) / (1 + (dy / ds)). 2 ) (3 / 2) .

[0133] The path planning process of this invention:

[0134] Input: vehicle's current position (x, y), heading angle θ, and velocity v.

[0135] Output: A sequence of path points that satisfy the constraints {(x i ,y i )}.

[0136] Algorithm: A hybrid A* algorithm is used to search for feasible paths, and Dijkstra's algorithm is combined to optimize the global cost (path length and curvature).

[0137] Hardware support for this invention:

[0138] Computing unit: NVIDIA Xavier chip, providing 30 TOPS of computing power for real-time solution of path optimization problems.

[0139] Storage module: Pre-stores vehicle dynamic parameters (such as wheelbase and steering ratio) for curvature constraint calculation.

[0140] The hardware structure of the multi-source fusion positioning submodule of this invention includes a GPS receiver (U-blox ZED-F9P), a 16-line LiDAR (Hesai Pandar16), and a binocular camera (Basler acA2440). Its workflow is as follows:

[0141] 1. GPS positioning: ZED-F9P outputs RTK (real-time dynamic positioning) data at a frequency of 10Hz, with a horizontal accuracy of ±1cm+1ppm.

[0142] 2. LiDAR point cloud matching: Pandar16 scans the environment to generate point clouds, which are then matched with high-precision maps using the NDT (Normal Distribution Transform) algorithm to output position corrections (Δx, Δy) with an accuracy of ±2cm.

[0143] 3. Visual SLAM Mapping: Images are acquired using a binocular camera, and the ORB-SLAM3 algorithm is run to extract feature points and construct a local map. The visual coordinate system is transformed to the vehicle coordinate system using a hand-eye calibration matrix, and the heading angle θ is output with an accuracy of ±0.3°.

[0144] The data fusion logic of this invention:

[0145] Loosely coupled fusion: Location data from GPS, LiDAR, and visual SLAM are fused using a Kalman filter, with the state vector being [x,y,θ,v]. m The observation equation is Z = HX + v m Where Z is the observation vector of the Kalman filter observation equation, H is the observation matrix, X is the state vector, and v m This is the observed noise vector.

[0146] Tightly coupled calibration: The poses of the LiDAR and visual SLAM are registered using the ICP algorithm, and the final localization result is output after minimizing the residual.

[0147] The sensor layout of this invention is as follows: the lidar is mounted in the center of the roof, the GPS antenna is located on the top of the trunk, and the binocular camera is fixed to the inside of the windshield. Heat dissipation design: the positioning unit housing has built-in heat dissipation fins and a fan to ensure stable operation of the sensor in environments ranging from -20℃ to 60℃.

[0148] The dynamic anti-interference submodule of the positioning unit consists of cooperative positioning logic, a visual inertial odometry (VIO) switching module, and a real-time self-calibration unit, and is implemented as follows:

[0149] Cooperative positioning mode:

[0150] 1. GPS signal loss detection: When the GPS module fails to output valid data for 5 consecutive seconds (data packet loss rate > 90%), the cooperative positioning mode is triggered.

[0151] 2. Acquisition of neighboring vehicle data: Receive positioning data (including position, heading angle and confidence level) of neighboring vehicles within a 50-meter radius via C-V2X communication.

[0152] 3. Relative position calculation: The least squares method is used to calculate the self-position. Let the position of the adjacent vehicle be {(x... i ,y i )}, its relative position (x,y) satisfies the system of equations: (x i -x) 2 +(y i -y) 2 =d i 2 , (d i (The distance is V2X measured), and by iteratively optimizing the solution (x,y), the accuracy can reach ±5cm.

[0153] VIO mode switching:

[0154] 1. LiDAR failure detection: If the LiDAR point cloud matching fails for 3 consecutive frames (150ms interval) (residual > 10cm), switch to VIO mode.

[0155] 2. Data Fusion: The IMU (model: ADIS16470) outputs angular velocity and acceleration, which are fused with the binocular vision feature tracking results (ORB feature points) through tight-coupled filtering. Visual-Inertial Odometry (VINS-Fusion) calculates pose in real time, with an output frequency of 100Hz and a position drift rate of <0.1% / m.

[0156] Real-time self-calibration logic:

[0157] 1. Multi-source data verification: Compare positioning data from GPS, LiDAR, and visual SLAM every 30 seconds. If the horizontal deviation between GPS and LiDAR is >2cm, trigger global relocation.

[0158] 2. Relocation process: The vehicle stops moving, the LiDAR scans the environmental point cloud, matches it with the preset map, and re-initializes the positioning coordinates, which takes ≤2 seconds.

[0159] Cooperative positioning module: integrated into the vehicle communication gateway (model: NXP S32G), with a built-in ARM Cortex-A53 processor, running the least squares solution algorithm.

[0160] VIO processing unit: Employs Xilinx Zynq UltraScale+MPSoC to achieve hardware-level synchronous acquisition and filtering of IMU and visual data.

[0161] Self-calibration trigger circuit: Timeout detection logic is designed based on FPGA (model: Intel Cyclone V) to monitor the validity of data from each sensor.

[0162] The dynamic feature weight allocation and tight coupling calibration of the visual SLAM algorithm are achieved through the following steps:

[0163] Dynamic feature weight allocation:

[0164] 1. Light intensity detection: The camera has a built-in ambient light sensor (model: TI OPT3001) to monitor the light intensity Lux value in real time.

[0165] When Lux > 10,000 (strong light) or Lux < 50 (low light), the edge feature extraction mode is enabled, and the Canny operator is used to detect high-contrast edges, with the weight increased to 0.8 and the texture feature weight reduced to 0.2.

[0166] Under normal lighting conditions (50≤Lux≤10,000), the weights of edge and texture features are evenly distributed (0.5 each).

[0167] 2. Target density determination: Through lidar point cloud density analysis, if the number of point clouds per cubic meter is >1000, it is determined to be a dense target area, and the semantic segmentation network (based on MobileNetV3) is activated.

[0168] The segmentation network outputs a passable area mask, retaining only ground planar feature points (such as lane lines and landmarks) and filtering out obstacle features with a height > 0.5 meters.

[0169] Multi-sensor tight-coupled calibration:

[0170] 1. Point cloud and visual map registration: Align the local dense map (point cloud format) generated by visual SLAM with the single-frame point cloud of LiDAR using the ICP algorithm, and calculate the rotation and translation matrix [T].

[0171] 2. Heading angle deviation correction: If the heading angle deviation Δθ = |θ visual -θ lidar |>0.5°, activate the Extended Kalman Filter (EKF), with the state vector [θ,ω] (ω being the IMU angular velocity), and the observation equation is θ measure =θ+ν k By predicting and updating the calibrated θ through a loop, the accuracy is ±0.1°, where θ visual θ represents the heading angle acquired by the visual SLAM system. lidar θ is the heading angle measured by the lidar. measure The heading angle is obtained by measuring the equation, ν k It is observation noise.

[0172] Dynamic weight allocation module: Deployed on GPU (model: NVIDIA Jetson AGX Xavier), running OpenCV library to achieve real-time feature extraction and weight adjustment.

[0173] Semantic segmentation accelerator: Employs Google Coral Edge TPU, with inference speed ≥30fps.

[0174] ICP acceleration unit: Parallel point cloud matching is implemented based on FPGA, with a single frame processing time of ≤10ms.

[0175] The workflow of the visual SLAM abnormal state self-healing mechanism is as follows:

[0176] Anomaly detection:

[0177] 1. Continuous tracking failure determination: If visual SLAM fails to output a valid pose for 5 consecutive frames (approximately 166ms) (feature point matching number <10), a self-repair process is triggered.

[0178] Cooperative positioning mode switching:

[0179] 1. Weighted average of neighboring vehicle heading angles: Obtain the heading angles {θ1, θ2, θ3} of the three nearest neighboring vehicles within a 30-meter radius via V2X, and calculate the weighted average θ. avg =w1θ1+ w2θ2+ w3θ3 (weight w i =1 / D i 2 D i (distance).

[0180] 2. Reverse Pose Estimation: Based on the outer contour feature points of the target object (provided by LiDAR), the PnP (Perspective-n-Point) algorithm is used to solve its own pose. Let the corner coordinates of the target object be {P}. i}, by minimizing the reprojection error ∑||u i - K[R T |t]P i || 2 Solve for the rotation matrix R T Translation vector t, where u i Let K be the actual projection coordinates of the i-th corner point on the image plane, and K be the intrinsic parameter matrix of the camera.

[0181] SLAM parameter reset:

[0182] 1. Initialization parameter overload: θ avg The ORB-SLAM3 tracker is reset using the initial heading angle and the initial position t.

[0183] 2. Local map reconstruction: An initial map is generated based on the LiDAR point cloud for subsequent visual SLAM tracking, with a recovery time of ≤100ms.

[0184] Anomaly detection circuit: Integrated into the vision processing unit, it monitors the validity of SLAM output through a counter.

[0185] PnP solution module: adopts Intel Movidius Myriad X VPU to accelerate matrix operations, with a single solution time of ≤5ms.

[0186] Parameter reset interface: Communicates with the SLAM master controller via the SPI bus, supporting fast writing of initialization parameters.

[0187] The dynamic channel optimization and multi-hop relay transmission of the C-V2X communication unit are implemented as follows:

[0188] Dynamic channel optimization:

[0189] 1. Channel Status Monitoring: A spectrum analysis module (model: Analog Devices AD9361) is used to monitor channel load (RSSI) and interference intensity (SINR) in real time.

[0190] 2. Reinforcement Learning Resource Allocation: Based on the DQN (Deep Q-Network) algorithm, the state space is [channel load, interference intensity, vehicle risk level], and the action space is a time-frequency resource block allocation scheme. During the training phase, a high-density scenario is simulated (≥20 vehicles / 100 meters), and the optimization objective is to minimize the transmission delay of high-risk vehicles (Risk≥3).

[0191] 3. High-risk vehicles take priority: When a vehicle has a Risk ≥ 3, it is allocated a dedicated resource block (10MHz bandwidth), and the remaining vehicles share the remaining resources.

[0192] Multi-hop relay transmission:

[0193] 1. Relay Node Selection: The central decision-making node (model: Huawei MH5000 module) periodically broadcasts a neighbor list. Relay candidate nodes must meet the following requirements:

[0194] Signal strength ≥ -90dBm (actual measurement distance ≤ 50 meters);

[0195] Remaining battery power > 30% (to prevent relay nodes from disconnecting midway).

[0196] 2. Segmented Transmission Protocol: Data packets are split into multiple segments, each of which is forwarded via relay nodes. Relay nodes add timestamps to ensure a delay of ≤10ms per hop. For example, vehicle data 250 meters from the central node, transmitted via 2 hops (125 meters apart), has a total delay of ≤25ms.

[0197] Data encryption mechanism:

[0198] 1. Quantum Key Distribution (QKD): Collision warning data uses the BB84 protocol to generate keys, which are transmitted through fiber optic channels, with a key update frequency of 1 time / second.

[0199] 2. Lightweight national cryptographic algorithm: The location and speed data use the SM4 algorithm (128-bit block length), and the encryption engine is integrated into the communication module with a throughput of ≥100Mbps.

[0200] Reinforcement learning accelerator: Employs Google Edge TPU, runs TensorFlow Lite models, and has an inference latency of ≤2ms.

[0201] Relay control module: Based on STM32H7 microcontroller, it realizes relay logic and power consumption management.

[0202] Encryption chip: Integrated with the SJ-A100 chip, which is certified by the national cryptographic standard and supports SM4 hardware acceleration.

[0203] The implementation steps of the dynamic priority algorithm and communication preemption mechanism are as follows:

[0204] Risk value calculation:

[0205] 1. Parameter acquisition: Real-time acquisition of vehicle TTC (Time to Collision), speed v, heading angle deviation θ, and path curvature radius R.

[0206] 2. Adaptive weight adjustment: Environmental coefficients α, β, and γ are dynamically adjusted according to the scene. For example, during nighttime testing (illuminance < 50 Lux), the TTC weight α is increased by 0.6; when targets are dense, the heading angle weight γ is increased by 0.3.

[0207] 3. Formula solution: Risk = 0.6 / TTC + 0.2·v / D + 0.2·θ / R, and the calculation result is normalized to the range of 0-10.

[0208] Channel preemption classification:

[0209] 1. Dedicated Emergency Channel: When Risk≥8, 10MHz bandwidth is allocated, emergency flags are added to data packets, routing priority is the highest, and transmission delay is ≤15ms.

[0210] 2. High-priority shared channel: When 5≤Risk<8, a 5MHz bandwidth is allocated, and the CSMA / CA mechanism is used to avoid low-priority data.

[0211] Strong synchronization mechanism for instructions:

[0212] 1. Hardware interrupt trigger: After the instruction receiving module (model: Xilinx AXI Interrupt Controller) detects an emergency instruction, it immediately sends a hardware interrupt signal (priority IRQ0) to the reversing motor controller (model: TI DRV8305) and the steering ECU (model: InfineonAurix).

[0213] 2. Parameter preloading: Cooperative parameters (such as target deceleration gradient -3m / s², heading correction Δθ=2°) are written into the controller's circular buffer before the command is issued, and are directly loaded after the interrupt is triggered, reducing software parsing time.

[0214] 3. Timing monitoring: The hardware timer (model: Microchip MCP7940N) records the delay from instruction reception to execution completion. When the timeout (>25ms) occurs, an alarm is triggered and redundancy control is started.

[0215] Interrupt controller: Integrated into the vehicle's central gateway, supporting 16 interrupt priority levels.

[0216] Parameter buffer: A dual-port RAM (model: Cypress CY7C024) is used to achieve high-speed read and write (access time ≤10ns).

[0217] The specific implementation of the hardware-level interrupt and preload coordination parameters is as follows:

[0218] Hardware interrupt module:

[0219] 1. Interrupt Signal Generation: After receiving the instruction packet from the communication module, the programmable logic device (model: Lattice iCE40UP5K) analyzes the risk level. If Risk ≥ 8, a priority interrupt signal (high level 3.3V) is generated and directly connected to the interrupt input ports of the motor controller and steering system via the GPIO pin.

[0220] 2. Interrupt response: The motor controller interrupt service routine (ISR) immediately suspends the current control loop, reads the coordination parameters from the preload buffer, and the response time is ≤1ms.

[0221] Coordination parameter preloading:

[0222] 1. Parameter writing: Parameters generated by the dynamic priority algorithm (such as target speed and heading angle correction) are transferred to the controller's preload area (address 0x8000-0x8FFF) via DMA (direct memory access) to avoid CPU intervention delay.

[0223] 2. Parameter verification: The CRC verification module (which generates the checksum using polynomial 0x04C11DB7) verifies data integrity and requests retransmission if the verification fails.

[0224] Timing and error control:

[0225] 1. Clock synchronization: The IEEE 1588 Precision Time Protocol (PTP) is adopted, and the clock deviation between vehicles is ≤1μs, ensuring that the start time of instruction execution is aligned.

[0226] 2. Hardware timer monitoring: From the moment the instruction is received (T0) to the completion of execution (T1), the timer records Δt = T1 - T0. If Δt > 25ms, emergency braking lock is triggered (power output is cut off).

[0227] 3. Error Verification: The positioning unit feeds back the actual displacement Δs. If |Δs| < Δs, then the error is verified. target -Δs|>0.2 meters, the brake lock protocol forces the vehicle to stop and broadcasts a fault code via V2X, Δs target This indicates the target displacement expected by the positioning unit.

[0228] Interrupt interface circuit: Optical isolation (model: Sharp PC817) is used to prevent electrical interference.

[0229] DMA controller: integrated into the MCU (model: NXP S32K144), supporting multi-channel concurrent transmission.

[0230] Clock synchronization module: Based on Marvell 88X3310P PHY chip, it achieves nanosecond-level time synchronization.

[0231] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0232] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A safety device for reducing the risk of reversing collisions during autonomous driving testing, characterized in that, include: Positioning units are deployed on all test vehicles and risk targets. The positioning units acquire millimeter-level position coordinates and heading angles of each vehicle in real time. The communication unit based on the C-V2X protocol constructs real-time data links between vehicles and between vehicles and targets based on the position coordinates and heading angles obtained by the positioning unit to share the following information: the real-time position, speed, and reversing trajectory prediction data of each vehicle, as well as the coordinates of the outer contour feature points and safety boundary parameters of the risky target. The communication unit receives the reversing trajectory prediction data uploaded by all vehicles, as well as the coordinates of the outer contour feature points and safety boundary parameters of the risk targets, and performs the following operations: Based on the reversing trajectory prediction data, the minimum future distance between each vehicle and between a vehicle and a risky target is calculated. If the minimum future distance is less than the dynamic threshold D, it is marked as a potential conflict. The dynamic threshold D ranges from 1.3 to 1.5 times the larger of the width and length of the risky target. Through a distributed optimization algorithm, a cooperative collision avoidance path is dynamically generated based on the position and speed data of each vehicle. At the same time, a dynamic priority algorithm is used to calculate the collision risk value of each vehicle in real time based on the dynamically generated cooperative collision avoidance path and potential conflict data. Communication channel resources are preferentially allocated to the vehicle with the highest collision risk value, and synchronization adjustment instructions are issued to its neighboring vehicles. The synchronization adjustment instructions include deceleration, adjustment of heading angle and braking, and the neighboring vehicles are forced to respond and execute the synchronization adjustment instructions within 20-25ms. The method for calculating the collision risk value of a vehicle and allocating communication channel resources is as follows: Collision risk value (Risk) is calculated based on dynamic weighting. Where θ is the vehicle heading angle deviation, R is the path curvature radius, α, β, γ are environmental adaptive weight coefficients with values ​​of α=8, β=2, γ=1, TTC is the collision time, v is the vehicle speed, and D is the dynamic threshold. Communication channel resources are allocated through a channel preemption hierarchical strategy, specifically as follows: Vehicles with a collision risk value of Risk ≥ 8 preempt the dedicated emergency channel, which has a bandwidth of 10MHz and a command transmission delay of ≤ 15ms; vehicles with a collision risk value of 5 ≤ Risk < 8 share the high-priority channel, which has a bandwidth of 5MHz and a delay of ≤ 25ms. When forcing adjacent vehicles to execute synchronization adjustment commands, a strong synchronization mechanism is used. The mechanism operates as follows: after receiving the command, the adjacent vehicle triggers the control signal through a hardware interrupt, preloads coordination parameters in the reversing motor controller and steering system, sets the trajectory correction or braking response time to within 20-25ms, and the error tolerance is ≤0.2 meters.

2. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 1, characterized in that, The specific response method for issuing synchronization adjustment instructions is as follows: When issuing a deceleration control command, a coordinated deceleration command is sent to all associated vehicles via V2X communication, so that the driving speed of each vehicle is reduced synchronously to the preset safety value. When issuing control commands to adjust the heading angle, the heading angle of each vehicle is dynamically adjusted based on the optimized path to ensure that its trajectory deviates from the safety boundary of the risky target. When a braking control command is issued, if the path conflict cannot be eliminated, a cascaded braking protocol is triggered, and emergency braking is performed sequentially according to the distance of the vehicles from the conflict point, from closest to furthest.

3. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 1, characterized in that, The constraints used when dynamically generating cooperative collision avoidance paths for each vehicle include: The minimum spacing between the reversing paths of each vehicle is greater than or equal to the dynamic threshold D; and The radius of curvature of the collision avoidance path is greater than or equal to the vehicle's minimum turning radius.

4. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 1, characterized in that, The positioning unit includes a multi-source fusion positioning submodule, which integrates GPS, LiDAR point cloud matching, and visual SLAM algorithms to achieve millimeter-level positioning through the following steps: The initial position coordinates are obtained through GPS, and the surrounding environmental feature point cloud is scanned by LiDAR. The data is then matched in real time with a high-precision preset site map to correct position deviations. Based on the image data collected by the vehicle-mounted camera, a visual SLAM algorithm is used to generate a local dense map to calibrate the heading angle.

5. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 4, characterized in that, The positioning unit also includes a dynamic anti-interference submodule, which performs the following operations when the signal is interfered with: When GPS signal obstruction is detected, a cooperative positioning mode based on vehicle-to-vehicle (V2X) communication is initiated to receive positioning data from neighboring vehicles and calculate its own relative position using the least squares method. If the lidar point cloud fails to match for at least 3 consecutive frames, it switches to pure visual inertial odometry (VIO) mode, fusing IMU data with visual feature tracking results to maintain positioning accuracy. Real-time self-calibration is performed, and a multi-source data consistency check is executed every 30 seconds. When the positioning deviation between GPS and LiDAR exceeds 2 cm, a global relocation process is triggered to correct the positioning deviation.

6. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 5, characterized in that, The steps involved in using visual SLAM algorithms are as follows: A dynamic feature weight allocation submodule is set up, which dynamically adjusts the extraction weights of visual feature points based on ambient light intensity and the density of risky targets. The specific operation mode is as follows: If the ambient light intensity is strong light or low light, high-contrast edge feature points are extracted first, and the extraction ratio and weight of texture features are reduced to reduce texture dependence; when the density of risky objects reaches the high density threshold, the semantic segmentation network is enabled to identify passable areas, and only ground plane feature points are retained for SLAM mapping. Configure a multi-sensor tightly coupled calibration submodule, which performs the following operations: The local dense map generated by visual SLAM is registered frame by frame with the LiDAR point cloud, and the heading angle deviation is calculated by the ICP algorithm. When the heading angle deviation exceeds 0.5°, the extended Kalman filter (EKF) is constructed by fusing IMU angular velocity data for coupling calibration, and the calibrated heading angle is output.

7. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 6, characterized in that, The visual SLAM algorithm also includes an abnormal state self-repair mechanism in its operating logic, which operates as follows: When visual SLAM tracking fails for 5 consecutive frames, it switches to cooperative localization mode based on V2X communication, receives heading angle data from neighboring vehicles and performs weighted averaging, reverse-calculates its own pose using the outer contour feature points of the risky target, and resets the SLAM initialization parameters.

8. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 1, characterized in that, The communication unit includes: The dynamic channel optimization submodule performs the following operations: real-time monitoring of the communication channel load status and interference intensity; dynamic allocation of time-frequency resource blocks based on reinforcement learning algorithms; prioritizing data transmission for vehicles with collision risk values ​​above the risk threshold based on real-time calculated vehicle collision risk values; and in scenarios where there are ≥20 test vehicles and / or ≥5 risk targets within a 100-meter radius, initiating a multi-hop relay transmission mode: setting a central decision node, and transmitting vehicle data more than 200 meters away from the central decision node in segments through relay nodes of adjacent vehicles, with a maximum delay of ≤10ms per hop; the relay node selection criteria are a signal strength ≥-90dBm and remaining battery power >30%. The data priority encryption submodule performs hierarchical encryption on transmitted data: collision warning data is encrypted using quantum key distribution (QKD), and position and velocity status data are encrypted using the lightweight national cryptographic algorithm SM4. The central decision-making node dynamically adjusts the number of iterations of the path optimization algorithm based on communication latency and packet loss rate; if the packet loss rate is greater than 5%, a local collision avoidance mode is triggered, allowing each vehicle to perform emergency braking solely based on local sensor data.

9. The safety device for reducing the risk of reversing collisions during autonomous driving testing according to claim 1, characterized in that, The specific operation modes of the instruction strong synchronization mechanism include: The hardware-level interrupt triggering module is directly connected to the underlying hardware interface between the reversing motor controller and the steering system. Priority interrupt signals are generated through programmable logic devices to forcibly interrupt the current control command and load the cooperative parameters. The cooperative parameters are preloaded in the following way: the cooperative parameters generated by the dynamic priority algorithm are written into the controller and steering system before the command is issued by the cooperative parameter preloading module. The cooperative parameters include the target deceleration gradient, heading angle correction and braking pressure threshold. The timing constraint module is set up to align the start time of instruction execution based on the inter-vehicle clock synchronization protocol, and monitors the entire delay from instruction reception to execution completion through a hardware timer, strictly controlling the delay within the range of 20-25ms. An error tolerance verification module is set up, which provides real-time feedback of the actual displacement through the positioning unit after trajectory correction or braking is completed. If the deviation from the target displacement exceeds 0.2 meters, the emergency braking lock protocol is triggered.

Images