Automatic driving safety control system based on accident analysis and predefined coping strategy

Abstract

The invention relates to an automatic driving safety control system based on accident analysis and predefined coping strategies, and the system comprises a multi-source data collection module which is used for fusing real-time vehicle-mounted sensor data, vehicle state data, high-precision map data and V2X communication data, historical accident data and accident prevention success case data are obtained from the cloud historical database; the accident mode analysis module is connected with the multi-source data acquisition module and is configured to analyze the historical data based on a machine learning model so as to construct an accident risk probability model and generate a dynamic local risk thermodynamic diagram in combination with the real-time data; and the strategy dynamic generation module is connected with the accident mode analysis module and is configured to be used for analyzing the local risk thermodynamic diagram and the current vehicle driving state. Through real-time risk assessment and strategy execution, the active safety of the vehicle in a complex traffic scene is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an autonomous driving safety control system based on accident analysis and predefined response strategies. Background Technology

[0002] Current autonomous driving systems primarily rely on sensors to perceive the environment in real time and make decisions based on rules or machine learning models. However, such systems have the following shortcomings when dealing with unexpected dangers (such as other vehicles illegally changing lanes or pedestrians suddenly crossing the road):

[0003] 1. Lack of historical accident data-driven approach: Existing technologies do not fully explore the potential patterns in historical accident data and historical successful accident prevention data, making it impossible to optimize response strategies for high-frequency hazardous scenarios.

[0004] 2. Static strategies: Predefined strategies are mostly fixed rules, which are difficult to dynamically adapt to the complexity and uncertainty of different traffic scenarios.

[0005] 3. Insufficient real-time performance: Traditional methods rely on online computation, which may lead to delays in critical decisions and increase the risk of accidents. Summary of the Invention

[0006] The purpose of this invention is to provide an autonomous driving safety control system based on accident analysis and predefined response strategies, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows.

[0008] An autonomous driving safety control system based on accident analysis and predefined response strategies includes:

[0009] The multi-source data acquisition module is configured to integrate real-time vehicle sensor data, vehicle status data, high-precision map data, V2X communication data, as well as historical accident data and accident prevention success case data obtained from the cloud historical database.

[0010] The accident mode analysis module is connected to the multi-source data acquisition module and is configured to analyze the historical data based on a machine learning model to construct an accident risk probability model, and generate a dynamic local risk heat map in combination with the real-time data.

[0011] The strategy dynamic generation module is connected to the accident mode analysis module and is configured to dynamically match and generate the optimal safety response strategy from a predefined hierarchical strategy library based on the local risk heat map and the current vehicle driving status.

[0012] The strategy execution and feedback module is connected to the strategy dynamic generation module. It is configured to execute the generated security response strategies and collect strategy execution effect data to feed back to the cloud for continuous optimization of the accident risk probability model and hierarchical strategy library.

[0013] Furthermore, the training and optimization of the accident risk probability model are completed in the cloud, and the trained model is then distributed to the vehicle.

[0014] The real-time generation of the local risk heatmap, the dynamic matching and execution of strategies are all completed on the vehicle side.

[0015] Furthermore, the hierarchical strategy library includes:

[0016] The Level 1 Emergency Avoidance Strategy is used to trigger emergency braking and active steering to avoid obstacles when the risk level exceeds the first threshold.

[0017] The Level 2 risk mitigation strategy is used to trigger following distance adjustment, speed limit control, and lane centering enhancement when the risk level is within the second threshold range.

[0018] The three-level early warning strategy is used to prompt the driver to take over when the risk level is below the second threshold but a warning is still required, through HUD or audible and visual alarms.

[0019] Furthermore, when matching strategies, the strategy dynamic generation module also takes into account the prediction results of the behavioral intentions of surrounding vehicles in order to optimize strategy selection.

[0020] Furthermore, the multi-source data acquisition module includes an onboard LiDAR, a camera, a millimeter-wave radar, an inertial navigation unit, and a vehicle-to-everything (V2X) communication unit.

[0021] Furthermore, the strategy execution and feedback module is electrically connected to the vehicle's brake-by-wire system, steering-by-wire system, and electronic stability system, and uses a reinforcement learning mechanism to optimize the timing and control parameters of strategy execution.

[0022] Furthermore, the accident mode analysis module includes:

[0023] Scene clustering units identify high-frequency accident scene patterns from historical accident data through machine learning;

[0024] The risk prediction unit trains the accident risk probability model based on the high-frequency accident scenario patterns and related environmental and behavioral characteristics.

[0025] A computer-readable storage medium storing program code for executing the aforementioned control system.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0027] 1. Response speed: Traditional methods rely on online computing, which often results in a long delay from perception to decision-making, potentially missing the best response opportunity. This invention, however, leverages edge-cloud collaborative computing, enabling local modules to achieve low-latency decision-making and quickly respond and execute strategies within 100ms, effectively reducing the risk of incidents.

[0028] 2. Strategy Adaptability: Traditional methods often predefine strategies based on fixed rules, making it difficult to dynamically adapt to the complexity and uncertainty of different traffic scenarios. This invention uses historical accident data to drive dynamic strategy generation and combines real-time feedback to achieve self-evolution. It can intelligently match the optimal strategy level according to different scenarios (such as missing an exit on a highway, a neighboring vehicle suddenly turning, etc.), balancing safety and traffic efficiency.

[0029] 3. Decision-making accuracy: Traditional methods lack sufficient mining of historical accident data, resulting in limited accuracy in responding to sudden dangers. This invention uses machine learning to extract potential patterns from historical accident data and successful accident prevention data, constructing an accident risk probability model. Combined with real-time environmental data, it dynamically generates local risk heatmaps, making decision-making more scientific and accurate. For example, in a scenario where a car on a highway is about to miss an exit and brakes abruptly, preparing to turn right to exit, it can make precise decisions based on multi-dimensional data.

[0030] 4. System Scalability: Traditional methods are usually limited to a single computing mode and have poor scalability. This invention adopts edge-cloud collaborative computing, with local modules responsible for low-latency decision-making, and the cloud completing big data analysis and model training, which facilitates subsequent system upgrades and expansions, and allows for continuous optimization of the incident model and strategy library;

[0031] 5. Continuous Optimization Capability: Traditional methods struggle to continuously optimize strategies once they are predefined. This invention records the effects of strategy execution and feeds them back to the cloud, continuously iterating the incident model and strategy library to achieve data-strategy closed-loop optimization. As data accumulates and feedback is received, system performance continuously improves. Attached Figure Description

[0032] Figure 1 This is a diagram of the overall architecture of the present invention;

[0033] Figure 2 This is a flowchart of the accident mode analysis of the present invention;

[0034] Figure 3 The logic diagram for the strategy of this invention is dynamically generated. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] This invention provides an autonomous driving safety control system based on accident analysis and predefined response strategies, such as... Figures 1-3 As shown, it includes:

[0037] The multi-source data acquisition module is configured to integrate real-time vehicle sensor data, vehicle status data, high-precision map data, V2X communication data, as well as historical accident data and accident prevention success case data obtained from the cloud historical database.

[0038] The accident mode analysis module is connected to the multi-source data acquisition module and is configured to analyze the historical data based on a machine learning model to construct an accident risk probability model, and generate a dynamic local risk heat map in combination with the real-time data.

[0039] The strategy dynamic generation module is connected to the accident mode analysis module and is configured to dynamically match and generate the optimal safety response strategy from a predefined hierarchical strategy library based on the local risk heat map and the current vehicle driving status.

[0040] The strategy execution and feedback module is connected to the strategy dynamic generation module. It is configured to execute the generated security response strategies and collect strategy execution effect data to feed back to the cloud for continuous optimization of the accident risk probability model and hierarchical strategy library.

[0041] The multi-source data acquisition module is responsible for collecting data from sensors such as LiDAR, cameras, and millimeter-wave radar, and acquiring traffic environment information through V2X communication. The accident mode analysis module performs machine learning analysis based on historical accident data provided by the cloud data center to construct a risk probability model. The strategy dynamic generation module selects appropriate response strategies from a hierarchical strategy library based on the risk analysis results. The strategy execution and feedback module sends control commands to vehicle actuators, including systems such as brake-by-wire and steering-by-wire, and simultaneously feeds back the execution effect data to the cloud data center, forming a complete closed-loop optimization system.

[0042] The beneficial effects of the above technical solution are as follows: By constructing a system architecture comprising four core modules, a complete technical chain of data-driven, dynamic evaluation, intelligent decision-making, and closed-loop optimization is achieved. Deep integration of historical experience and real-time perception enables the system to learn from the past, providing more comprehensive decision-making basis. The dynamic generation and matching mechanism of strategies overcomes the rigidity of traditional static rules, improving the system's adaptability to complex and ever-changing traffic scenarios. The final feedback loop ensures that the system can continuously learn and optimize from actual operation, achieving continuous performance improvement, marking a fundamental shift in autonomous driving safety control from passive response to proactive prevention.

[0043] The training and optimization of the accident risk probability model are completed in the cloud, and the trained model is then sent to the vehicle.

[0044] The real-time generation of the local risk heatmap, the dynamic matching and execution of strategies are all completed on the vehicle side.

[0045] The system's computing architecture adopts an edge-cloud collaborative model. The cloud data center possesses powerful storage and computing capabilities, enabling it to clean, label, and feature-engineer massive amounts of historical accident data and successful prevention case data from a wide range of regions. It also uses machine learning to identify high-frequency accident scenario characteristics and construct accident risk probability models. These models are periodically or on-demand distributed to the vehicle. The vehicle's computing unit utilizes the distributed risk probability model, inputs real-time collected multi-source sensor data, and rapidly performs forward inference to generate local heatmaps reflecting the current surrounding environmental risks (such as blind spots at intersections and construction zones). Subsequently, the strategy dynamic generation module performs millisecond-level strategy matching and execution based on this heatmap and vehicle status.

[0046] The beneficial effects of the above technical solution are as follows: Through a reasonable division of tasks between the cloud and the vehicle, optimal allocation of computing resources is achieved. The cloud focuses on time-consuming model training, ensuring the model's accuracy and generalization ability; the vehicle focuses on low-latency real-time inference and decision-making, ensuring system response speed. This architecture frees the vehicle system from a heavy computational burden, allowing it to concentrate valuable computing resources on perception and emergency decision-making, effectively avoiding accidents caused by computational latency. Simultaneously, continuous updates to the cloud model enable the entire fleet to share the latest safety knowledge, achieving synchronous improvement in collective intelligence.

[0047] The hierarchical strategy library includes:

[0048] The Level 1 Emergency Avoidance Strategy is used to trigger emergency braking and active steering to avoid obstacles when the risk level exceeds the first threshold.

[0049] The Level 2 risk mitigation strategy is used to trigger following distance adjustment, speed limit control, and lane centering enhancement when the risk level is within the second threshold range.

[0050] The three-level early warning strategy is used to prompt the driver to take over when the risk level is below the second threshold but a warning is still required, through HUD or audible and visual alarms.

[0051] When the risk level exceeds the first threshold, the system triggers a Level 1 emergency avoidance strategy, executing emergency braking and active steering to avoid obstacles. This is the system's last line of defense, prioritizing the protection of life. When the risk level is within the second threshold range, the system triggers a Level 2 risk mitigation strategy, adjusting following distance, implementing speed limits, and enhancing lane keeping to minimize interference with normal driving while ensuring safety. When the risk level is below the second threshold but still requires warning, the system triggers a Level 3 warning alert strategy, prompting the driver to take over via a HUD (Head-Up Display) and audible / visual alarms, achieving human-machine collaborative safety assurance.

[0052] The beneficial effects of the above technical solution are as follows: By introducing a hierarchical strategy mechanism, refined management of safety control is achieved. This layered design ensures that the system no longer resorts to a crude, binary intervention, but rather takes the most appropriate measures based on the severity of the situation. It effectively balances safety and traffic efficiency, avoids traffic flow disruptions caused by overly conservative strategies, and ensures that the most effective risk avoidance measures can be taken in genuine emergencies, greatly improving system availability and user acceptance.

[0053] When matching strategies, the strategy dynamic generation module also takes into account the prediction results of the behavioral intentions of surrounding vehicles in order to optimize strategy selection.

[0054] The strategy dynamic generation module integrates behavior prediction functionality. Based on real-time perception data (such as the position, speed, acceleration, and turn signal status of surrounding vehicles) and historical trajectories, this function uses machine learning models to predict short-term behavioral intentions, such as lane changes, lane cuts, and deceleration. When performing strategy matching, the module not only relies on the current static risk heatmap but also uses these prediction results as important inputs.

[0055] The strategy dynamic generation module of this system employs multiple algorithms working in tandem, specifically including:

[0056]

[0057]

[0058] A key application of this system lies in its use of decision trees for rapid matching of typical high-frequency scenarios. The decision rules are based on the analysis and summarization of historical accident data and successful prevention cases. Below are examples of decision tree datasets for two typical scenarios:

[0059] Decision Tree Dataset 1: A car on a highway is about to miss its exit and brakes hard, preparing to turn right to exit.

[0060]

[0061]

[0062] Decision Tree Dataset 2: The vehicle is traveling in the right lane, and a vehicle on the left, which is aligned with it, suddenly turns towards the vehicle.

[0063] When the scene features perceived by the system match a row in the aforementioned dataset, the corresponding predefined strategy can be quickly invoked to respond, achieving low-latency and high-reliability safety control. This rapid matching mechanism based on historical experience effectively complements the main risk assessment model and is particularly suitable for emergency situations with extremely high response speed requirements.

[0064] The beneficial effects of the above technical solutions are as follows: strategy selection is not only based on static risk assessment but also incorporates dynamic predictions of future traffic participant behavior, making decision-making more forward-looking. The collaborative work of multiple algorithms ensures optimal decision-making results in different scenarios. The application of decision trees provides a rapid response channel for high-frequency typical scenarios, significantly improving the system's real-time performance. This proactive decision-making logic can identify potential conflicts earlier and take smoother, less invasive intervention measures, significantly improving driving smoothness and safety, avoiding a large number of unnecessary emergency brakings, and also improving the overall efficiency of traffic flow.

[0065] The multi-source data acquisition module includes an onboard LiDAR, a camera, a millimeter-wave radar, an inertial navigation unit, and a vehicle-to-everything (V2X) communication unit.

[0066] LiDAR provides precise 3D point cloud distance information, cameras provide rich texture and semantic information (such as traffic signs, traffic light status, and vehicle type), millimeter-wave radar is extremely sensitive to the speed detection of moving targets and is unaffected by adverse weather conditions (rain, fog, and dust), ultrasonic radar is suitable for low-speed, close-range parking and obstacle detection, inertial navigation units provide short-term, continuous position and attitude estimation when GPS signals are lost or interfered with, and vehicle-to-everything (V2X) communication units enable information interaction between vehicles (V2V), between vehicles and the road (V2I), and between vehicles and the cloud (V2C), acquiring beyond-line-of-sight traffic situation and risk information.

[0067] The beneficial effects of the above technical solution are as follows: By deeply fusing heterogeneous sensors, a redundant and reliable perception system is constructed. These sensors each have their own characteristics and complementary advantages. Through data fusion algorithms, a comprehensive, accurate, and real-time environmental situation map can be generated. This redundant design greatly reduces the possibility of system misjudgment or failure due to the failure, performance limitation, or existence of perception blind spots of a single sensor, providing a solid and reliable data foundation for subsequent risk assessment and decision-making, and is a fundamental guarantee for the safety of high-level autonomous driving.

[0068] The strategy execution and feedback module is electrically connected to the vehicle's brake-by-wire system, steering-by-wire system, and electronic stability system, and uses a reinforcement learning mechanism to optimize the timing and control parameters of strategy execution.

[0069] The strategy execution and feedback module is deeply integrated with the underlying drive-by-wire actuators via the vehicle bus (such as CAN FD, Ethernet) to ensure that control commands (such as target deceleration and steering angular velocity) can be executed quickly and accurately. More importantly, this module uses a reinforcement learning mechanism to optimize the strategy execution process. It treats each strategy execution as a trial-and-error process, constructing a reward function based on the execution results collected by the effect evaluation unit (such as obstacle avoidance success rate, trajectory smoothness, and passenger comfort indicators). Then, it uses reinforcement learning algorithms (such as DDPG, PPO) to optimize the timing of strategy execution (such as intervention timing and duration) and control parameters (such as the slope of the braking curve and steering angular velocity) online or offline, so that the same strategy command can achieve the optimal execution effect under different operating conditions.

[0070] The beneficial effects of the above technical solution are as follows: Deep integration with the vehicle's underlying drive-by-wire system ensures direct and rapid control. The introduction of a reinforcement learning mechanism enables the system to learn and optimize strategy execution timing, ensuring low-latency response (<100ms), recording the strategy execution effect and feeding it back to the cloud for continuous iteration of the accident model and strategy library.

[0071] The accident mode analysis module includes:

[0072] Scene clustering units identify high-frequency accident scene patterns from historical accident data through machine learning;

[0073] The risk prediction unit trains the accident risk probability model based on the high-frequency accident scenario patterns and related environmental and behavioral characteristics.

[0074] The scene clustering unit employs unsupervised learning algorithms (such as DBSCAN and K-Means) to perform clustering analysis on multidimensional features (such as weather, lighting, road type, traffic density, participant type and behavior) from massive accident reports, automatically summarizing typical accident scene patterns such as high-speed rear-end collisions, cross-traffic collisions at intersections, and vehicles illegally cutting in. The risk prediction unit, for each identified scene pattern, uses supervised learning algorithms (such as random forest and gradient boosting tree GBDT) to train a classification or regression model, taking scene features, environmental features, and real-time behavioral features as input, and outputting the probability value of an accident or near-accident occurring in that scene.

[0075] The beneficial effects of the above technical solution are as follows: Through a two-step approach of scene clustering and risk prediction, chaotic historical data is transformed into structured and quantifiable knowledge. Scene clustering enables the system to understand and classify different types of hazards, rather than treating each situation as an isolated case. The risk prediction model then translates this understanding into specific, calculable probability values, providing a clear quantitative basis for decision-making. For example, the system not only knows that there is a car ahead, but can also calculate the probability of a rear-end collision under the current following distance, relative speed, and road adhesion conditions: X%. This probability-based decision-making is far more refined, scientific, and reliable than binary judgments (danger or safety) based on simple rules.

[0076] A computer-readable storage medium storing program code for executing any of the aforementioned control systems.

[0077] The storage medium can be any volatile or non-volatile storage device, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk. The program code stored on it includes, but is not limited to: instructions for controlling the multi-source data acquisition module to perform data fusion; algorithmic instructions for performing accident mode analysis, including scene clustering and risk prediction; logical instructions for implementing dynamic strategy matching and generation; and instructions for controlling vehicle actuators and completing data recording and feedback. When this program code is run by one or more computing devices (such as an onboard computer or a cloud server), it enables the computing device to perform the system functions of the control system as described above.

[0078] The beneficial effects of the above technical solution are as follows: Protecting the invention through a computer-readable storage medium allows for convenient copying, distribution, and deployment of the technical solution in software form. This greatly promotes the commercial application and widespread adoption of this advanced autonomous driving safety control system, allowing different vehicle platforms or simulation testing environments to achieve consistent safety performance improvements by loading the same program code. Simultaneously, this also facilitates subsequent online upgrades and maintenance of the system.

[0079] The above is a detailed description of the present invention in conjunction with specific embodiments, and it should not be construed that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, any equivalent substitutions or obvious modifications made without departing from the concept of the present invention, and which have the same performance or use, should be considered to fall within the patent protection scope defined by the submitted claims.

Claims

1. An autonomous driving safety control system based on accident analysis and predefined response strategies, characterized in that, include: The multi-source data acquisition module is configured to integrate real-time vehicle sensor data, vehicle status data, high-precision map data, V2X communication data, as well as historical accident data and accident prevention success case data obtained from the cloud historical database. The accident mode analysis module is connected to the multi-source data acquisition module and is configured to analyze the historical data based on a machine learning model to construct an accident risk probability model, and generate a dynamic local risk heat map in combination with the real-time data. The strategy dynamic generation module is connected to the accident mode analysis module and is configured to dynamically match and generate the optimal safety response strategy from a predefined hierarchical strategy library based on the local risk heat map and the current vehicle driving status. The strategy execution and feedback module is connected to the strategy dynamic generation module. It is configured to execute the generated security response strategies and collect strategy execution effect data to feed back to the cloud for continuous optimization of the accident risk probability model and hierarchical strategy library.

2. The autonomous driving safety control system based on accident analysis and predefined response strategies according to claim 1, characterized in that: The training and optimization of the accident risk probability model are completed in the cloud, and the trained model is then sent to the vehicle. The real-time generation of the local risk heatmap, the dynamic matching and execution of strategies are all completed on the vehicle side.

3. The autonomous driving safety control system based on accident analysis and predefined response strategies according to claim 2, characterized in that: The hierarchical strategy library includes: The Level 1 Emergency Avoidance Strategy is used to trigger emergency braking and active steering to avoid obstacles when the risk level exceeds the first threshold. The Level 2 risk mitigation strategy is used to trigger following distance adjustment, speed limit control, and lane centering enhancement when the risk level is within the second threshold range. The three-level early warning strategy is used to prompt the driver to take over when the risk level is below the second threshold but a warning is still required, through HUD or audible and visual alarms.

4. The autonomous driving safety control system based on accident analysis and predefined response strategies according to claim 3, characterized in that: When matching strategies, the strategy dynamic generation module also takes into account the prediction results of the behavioral intentions of surrounding vehicles in order to optimize strategy selection.

5. The autonomous driving safety control system based on accident analysis and predefined response strategies according to claim 4, characterized in that: The multi-source data acquisition module includes an onboard LiDAR, a camera, a millimeter-wave radar, an inertial navigation unit, and a vehicle-to-everything (V2X) communication unit.

6. The autonomous driving safety control system based on accident analysis and predefined response strategies according to claim 5, characterized in that: The strategy execution and feedback module is electrically connected to the vehicle's brake-by-wire system, steering-by-wire system, and electronic stability system, and uses a reinforcement learning mechanism to optimize the timing and control parameters of strategy execution.

7. The autonomous driving safety control system based on accident analysis and predefined response strategies according to claim 5, characterized in that: The accident mode analysis module includes: Scene clustering units identify high-frequency accident scene patterns from historical accident data through machine learning; The risk prediction unit trains the accident risk probability model based on the high-frequency accident scenario patterns and related environmental and behavioral characteristics.

8. A computer-readable storage medium, characterized in that: It stores program code for executing the control system described in claims 1-7.

Images