Non-steady-state traffic flow intersection machine non-fusion perception obstacle avoidance method

Abstract

This invention relates to the field of intelligent transportation and autonomous driving technology, and discloses a non-motorized vehicle-non-motorized vehicle fusion perception obstacle avoidance method for non-steady-state traffic flow intersections. The method includes the following steps: acquiring full-area video of the intersection using high-altitude video equipment; extracting continuous driving trajectories of motor vehicles and non-motorized vehicles using target detection and tracking algorithms; and constructing a standardized motor vehicle-non-motorized vehicle trajectory database containing multiple scenarios after optimization processing. Then, a simulation environment is constructed using this database as the data source, and an obstacle avoidance decision model is trained using a deep deterministic strategy gradient algorithm. The model takes the vehicle state, surrounding obstacle states, motor vehicle-non-motorized vehicle trajectory characteristics, traffic signals, and road characteristics as inputs, and outputs continuous speed and heading angle adjustments. The model is deployed on the vehicle controller and the vehicle is controlled to execute the algorithm. This invention provides a non-steady-state traffic flow intersection non-motorized vehicle-non-motorized vehicle fusion perception obstacle avoidance method, which achieves safe and smooth obstacle avoidance control with adaptability and decision-making accuracy in complex traffic scenarios.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and autonomous driving technology, and in particular to a non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections. Background Technology

[0002] Signalized intersections, as key nodes in urban traffic flow, are characterized by the mixing of motorized and non-motorized vehicles (hereinafter referred to as "motorized and non-motorized vehicles") and dynamic changes in traffic flow over time. Especially during peak hours, inclement weather, or sudden traffic incidents, non-steady-state traffic flow is prone to occur. Non-steady-state traffic flow is characterized by large fluctuations in flow rate, random driving trajectories, and frequent interactions between motorized and non-motorized vehicles, which greatly increases the difficulty of vehicle obstacle avoidance decisions, making traditional obstacle avoidance methods insufficient to meet the requirements of safe driving.

[0003] Existing vehicle safety obstacle avoidance technologies are mainly divided into two categories: one is sensor-based real-time perception obstacle avoidance, which uses onboard radar, cameras, and other devices to acquire information about the surrounding environment and generates obstacle avoidance commands using preset rules or simple algorithms. However, this type of method has shortcomings such as limited perception range, susceptibility to severe weather interference, and insufficient prediction of sudden conflicts between motorized and non-motorized vehicles under non-steady-state traffic flow. For example, in scenarios where non-motorized vehicles illegally cross intersections, obstacle avoidance decisions are often delayed due to perception delays. The other category is obstacle avoidance methods based on reinforcement learning, which achieve dynamic obstacle avoidance by building intelligent decision-making models. However, existing models are mostly trained using simulated traffic flow data and do not fully incorporate the trajectory characteristics of motorized and non-motorized vehicles at real intersections, resulting in poor adaptability and decision-making accuracy of the models in real non-steady-state traffic scenarios.

[0004] Meanwhile, existing technologies for extracting vehicle and non-motorized vehicle trajectories largely rely on fixed monitoring equipment, resulting in limited data collection range and difficulty in fully covering the non-steady-state traffic flow across the entire intersection. Furthermore, the lack of a standardized vehicle and non-motorized vehicle trajectory database hinders the training of deep reinforcement learning models from insufficient and realistic data sources. Therefore, accurately collecting vehicle and non-motorized vehicle trajectories at signal-controlled intersections with non-steady-state traffic flow, constructing a dedicated database, and training deep reinforcement learning models based on real trajectory data to achieve efficient and safe vehicle obstacle avoidance has become a pressing technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections, which achieves safe and smooth obstacle avoidance control with improved adaptability and decision-making accuracy in complex traffic scenarios.

[0006] This invention provides a non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections, comprising the following steps: acquiring full-area video of the signal control intersection through a high-altitude video acquisition device, extracting continuous driving trajectory data of motor vehicles and non-motor vehicles in the video using a target detection and tracking algorithm, and optimizing the trajectory data and calculating motion parameters; The extracted trajectory data and related metadata are stored in the database to form a standardized machine-non-vehicle trajectory database containing multiple scenarios and types of machine-non-vehicle trajectories; Based on the real vehicle-non-vehicle interaction trajectory data stored in the standardized vehicle-non-vehicle trajectory database, a simulation environment for simulating unsteady traffic flow is constructed, and a deep deterministic policy gradient algorithm is used to train an obstacle avoidance decision model in this simulation environment. The model takes the vehicle state, the state of surrounding obstacles, vehicle-non-vehicle trajectory features, traffic signal state, and intersection road features as inputs, and outputs continuous vehicle speed adjustment and heading angle adjustment. The vehicle-non-vehicle trajectory features are obtained by extracting the trajectory curvature of surrounding obstacles within a preset time period, and are used to characterize the uncertainty of obstacle driving trends under unsteady traffic flow. The trained model is deployed on the vehicle controller, and combined with real-time multi-sensor perception fusion data, obstacle avoidance action commands are generated and the vehicle is controlled to execute them.

[0007] Furthermore, the collected video is processed and trajectory data is extracted, specifically including: Establish the mapping relationship between video pixel coordinates and world coordinates, and calibrate the coordinate transformation coefficients through on-site measurements; Target features are extracted using a target detection algorithm, and a detection confidence threshold is set. The DeepSORT tracking algorithm is used to continuously track the target. When the target occlusion time does not exceed a preset threshold, the Kalman filter algorithm is used to predict the target position. Outlier removal and interpolation smoothing are performed on the original trajectory data to obtain continuous non-motor trajectory data; Based on the optimized trajectory data, the instantaneous velocity, instantaneous acceleration, and heading angle motion parameters of each target are calculated.

[0008] Furthermore, the database is built using MySQL and contains at least a basic information table, a trajectory data table, and a traffic signal table. The basic information table stores the data collection task identifier, collection time, and traffic scene information. The trajectory data table stores the target identifier, target type, timestamp, world coordinates, and speed information. The traffic signal table stores the approach direction, signal phase, and phase duration information. The tables in the database are linked through the data collection identifier.

[0009] Furthermore, the obstacle avoidance decision model constructed by the deep deterministic policy gradient algorithm includes a policy network and a value network, both of which are fully connected neural networks. The policy network takes the environment state vector as input and outputs a continuous obstacle avoidance action vector. The value network takes the environment state vector and action vector as input and outputs the action value.

[0010] Furthermore, the environment state vector is 38-dimensional, specifically including: The vehicle's status in 4 dimensions includes its position, speed, acceleration, and heading angle. 25-dimensional surrounding obstacle status: Select the k nearest obstacles within a preset range around the vehicle. Each obstacle includes its position, speed, acceleration, and type information, with the type distinguishing between motor vehicles and non-motor vehicles. 5D Trajectory Features of Motor Vehicles and Non-motor Vehicles: Extracting the trajectory curvature of surrounding obstacles within a preset time period; Traffic signal status in 2D: the current signal phase and remaining phase time of the approach lane where the vehicle is located; Two-dimensional characteristics of intersection roads: the width of the lane where the vehicle is located and the distance from the center of the intersection to the vehicle.

[0011] Furthermore, the continuous obstacle avoidance action vector is 2-dimensional, including: speed adjustment amount, with a value range of [-5,2] m / s; and heading angle adjustment amount, with a value range of [-15°,15°].

[0012] Furthermore, the reward function used to train the obstacle avoidance decision model is a weighted sum of safety rewards, efficiency rewards, and rule rewards; Safety rewards are calculated based on the collision risk between the vehicle and obstacles. Positive rewards are given when there is no collision risk, negative rewards are given when there is a collision risk, and maximum negative rewards are given when a collision occurs. Efficiency rewards are calculated based on the deviation between the vehicle's speed and the preset expected speed; The rule-based rewards are calculated based on the vehicle's compliance with traffic signals and lane keeping.

[0013] Furthermore, the on-board controller further includes: Real-time perception fusion: The vehicle collects driving status and surrounding obstacle status through multiple on-board sensors, obtains real-time traffic signal status through the vehicle-road cooperative module, and uses filtering algorithms to fuse multi-source data; State vector construction: Real-time state vectors are constructed based on the fused data and standardized using normalization parameters consistent with those used in model training; Optimal action generation: Input the standardized state vector into the deployed DDPG model to generate the optimal obstacle avoidance action; Execution control and feedback adjustment: Convert action commands into control signals for the braking and steering systems, and collect the actual state of the vehicle after execution in real time. When the deviation from the model's predicted state exceeds a preset threshold, the control signal is corrected through a PID adjustment algorithm.

[0014] Furthermore, the high-altitude video acquisition equipment is a drone equipped with a high-definition camera with a resolution of 3840×2160 and a frame rate of 30fps.

[0015] Furthermore, it also includes a step of multi-dimensional evaluation of the trained obstacle avoidance decision model: using four dimensions—safety evaluation, efficiency evaluation, comfort evaluation, and rule compliance evaluation—assigning preset weights to calculate a comprehensive evaluation score to quantitatively assess the obstacle avoidance effect.

[0016] Compared with the prior art, the technical solution provided by this invention has the following advantages: This invention acquires full-area video of signal-controlled intersections using high-altitude video acquisition equipment, and extracts continuous driving trajectories of motor vehicles and non-motor vehicles using target detection and tracking algorithms. After optimization, a standardized database containing multiple scenarios and types of motor vehicle and non-motor vehicle trajectories is constructed, overcoming the problems of limited sensing range, susceptibility to interference, and poor adaptability of reinforcement learning models due to the lack of real trajectory data in traditional sensors. Furthermore, a simulation environment is constructed using this real trajectory database as the data source, and a deep deterministic policy gradient algorithm is used to train an obstacle avoidance decision model. This model considers the vehicle's state, the state of surrounding obstacles, and the trajectories of motor vehicles and non-motor vehicles. Using features, traffic signal status, and intersection road features as inputs, the system outputs continuous vehicle speed and heading angle adjustments. This addresses the technical bottlenecks of existing technologies, such as insufficient prediction of sudden vehicle-non-vehicle conflicts due to perception delays and low decision-making accuracy of simulated data-trained models in real non-steady-state scenarios. Finally, the trained model is deployed on the vehicle controller, and obstacle avoidance commands are generated by combining real-time multi-sensor perception fusion data. This achieves full-chain technology integration from accurate perception across the entire domain and real-trajectory data-driven model training to real-time obstacle avoidance control, significantly improving the vehicle's ability to predict complex vehicle-non-vehicle interaction scenarios and the accuracy and real-time performance of obstacle avoidance decisions in non-steady-state traffic flow intersections. Attached Figure Description

[0017] Figure 1 A simplified flowchart of a non-fusion perception obstacle avoidance method for non-steady-state traffic flow at intersections provided in an embodiment of the present invention; Figure 2 This invention provides a data acquisition and processing flowchart for a non-fusion perception obstacle avoidance method for non-steady-state traffic flow at intersections, as shown in the embodiment of the invention. Figure 3 This invention provides a technical roadmap for a non-fusion perception obstacle avoidance method for non-steady-state traffic flow at intersections, as provided in an embodiment of the invention. Figure 4 A flowchart illustrating a non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections, provided in an embodiment of the present invention; Figure 5 The above figure shows the speed-time trend of motorized and non-motorized vehicles during the morning rush hour, and the below figure shows the speed-time trend of motorized and non-motorized vehicles during the evening rush hour. Figure 6 The longitudinal acceleration of motor vehicles and non-motor vehicles provided in the embodiments of the present invention; Figure 7 The lateral acceleration of motor vehicles and non-motor vehicles provided in the embodiments of the present invention; Figure 8 The obstacle avoidance success rate under non-steady-state scenarios provided in the embodiments of the present invention; Figure 9 Obstacle avoidance rate under non-steady-state scenario comparison provided by embodiments of the present invention; Figure 10 This refers to the average traffic speed under parallel comparison of unsteady-state scenarios provided in the embodiments of the present invention. Figure 11 This is a comprehensive evaluation score provided by the embodiments of the present invention under the comparison of unsteady-state scenarios; Figure 12 The DDPG model training convergence curve of a non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections provided in this embodiment of the invention; Figure 13 This is a multi-dimensional comparison chart of obstacle avoidance performance in non-steady-state traffic flow scenarios provided by embodiments of the present invention; Figure 14 A comparison chart of the obstacle avoidance success rates of the obstacle avoidance methods provided in the embodiments of the present invention; Figure 15 This is a comparison chart of collision rates of obstacle avoidance methods provided in embodiments of the present invention; Figure 16 A comparison chart of comprehensive evaluation scores for obstacle avoidance methods provided in embodiments of the present invention; Figure 17 This is a diagram illustrating the obstacle avoidance effect of vehicles at an intersection with unsteady traffic flow, provided in an embodiment of the present invention. Figure 18 This is a comparison diagram of obstacle avoidance trajectories using multiple methods at signal-controlled intersections, provided in an embodiment of the present invention. Detailed Implementation

[0018] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0021] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for obstacle avoidance at non-steady-state traffic flow intersections using fusion perception of motorized and non-motorized vehicles, comprising the following steps: acquiring full-area video of the signal-controlled intersection through a high-altitude video acquisition device; extracting continuous driving trajectory data of motorized and non-motorized vehicles from the video using a target detection and tracking algorithm; optimizing the trajectory data and calculating motion parameters; storing the extracted trajectory data, acquisition scene information, traffic signal timing parameters, and equipment parameters into a database to form a standardized motorized and non-motorized vehicle trajectory database containing multiple scenarios and types of motorized and non-motorized vehicle trajectories; constructing a simulation environment using the constructed standardized motorized and non-motorized vehicle trajectory database as the data source, and training an obstacle avoidance decision model in the simulation environment using a deep deterministic strategy gradient algorithm; the model takes the vehicle state, surrounding obstacle state, motorized and non-motorized vehicle trajectory characteristics, traffic signal state, and intersection road characteristics as inputs, and outputs continuous vehicle speed adjustment and heading angle adjustment; deploying the trained model on the vehicle controller, combining real-time multi-sensor perception fusion data, generating obstacle avoidance action commands, and controlling the vehicle to execute them.

[0022] The modules are sequentially connected: real data acquisition of video from unmanned aerial vehicles (UAVs) of unsteady traffic flow at signalized intersections; extraction of vehicle and non-vehicle trajectories at signalized intersections based on Datafromsky; construction of a database of vehicle and non-vehicle trajectories at signalized intersections; and vehicle safety obstacle avoidance methods based on deep reinforcement learning and vehicle and non-vehicle trajectories. This module uses Datafromsky software to extract the trajectories of motor vehicles and non-motor vehicles from the pre-processed intersection video, achieving accurate capture of the driving trajectories of motor vehicles and non-motor vehicles. The specific steps are as follows: In a specific embodiment of this invention, Datafromsky 2.5 Professional Edition is used to adapt to the high-altitude video acquisition and analysis requirements of UAVs. The hardware requirements are: CPU ≥ Intel Core i7-12700K, GPU ≥ NVIDIA RTX 3060 (6GB VRAM), Memory ≥ 16GB, support for 64-bit Windows 10 / 11 systems, maximum video resolution of 7680×4320, frame rate of 10-60fps, timestamp synchronization accuracy ≤1ms, and pixel coordinate resolution error ≤0.5 pixels. Furthermore, for UAVs operating at 150-200m altitudes and in mixed-traffic scenarios, core parameters are customized: the detection module is set to small target enhancement mode, confidence threshold is set to 0.85, and NMS is adjusted accordingly. The IOU threshold is set to 0.45, limiting the entire intersection area to the detection zone. The tracking module uses an enhanced appearance feature matching mode, with a tracking IOU threshold of 0.5 and a matching threshold of 0.6. The target occlusion time threshold is set to 3 seconds. The Kalman filter position and velocity prediction errors are ≤0.3m and ≤0.2m / s, respectively. A binary feature library for motorized and non-motorized vehicles is established to achieve accurate subclass classification. The trajectory calculation module uses a custom world coordinate mapping mode, retaining coordinate transformation coefficient accuracy to 0.001m / pixel. The sliding window size is set to 5 frames for trajectory smoothing. Velocity is calculated using the finite difference method at 1 / 30s time intervals, and acceleration is calculated using the second-order finite difference method. The heading angle calculation accuracy is 0.1°. Synchronous output of structured data such as CSV / JSON / SQL is supported. The target detection and classification models built into the Datafromsky software are pre-trained general-purpose models. In a specific implementation, a transfer learning approach was used for lightweight fine-tuning training: 10% of the 20,000 frames of images from multi-scene videos captured by drones were manually labeled in VOC format and divided into training and validation sets in a 7:3 ratio. The first 10 convolutional layers of the pre-trained model were frozen, and only the last 5 layers and fully connected layers were trained. A learning rate of 1×10⁻⁻⁻⁴ was set. 4 The batch size is 16, and the training iterations are 50 epochs. The cross-entropy loss function is used. Training is stopped when the accuracy of the validation set is ≥98% and there is no improvement for 5 consecutive epochs. The fine-tuned model parameters are imported into the software to replace the default model, so as to achieve accurate detection and classification of non-targets at intersections.

[0023] Lightweight secondary development was performed based on the software Python API interface (SDK V2.0). A trajectory data preprocessing module was added, embedding the 3σ criterion and cubic spline interpolation algorithm to achieve automatic outlier removal and missing data interpolation supplementation; the default data output fields were modified to add metadata fields such as collection ID, intersection ID, and traffic scene type, which correspond one-to-one with subsequent database fields, and the output file can be directly imported into a MySQL database.

[0024] Launch Datafromsky software, import the pre-processed video file, and set the video parsing parameters: pixel size consistent with the original video (3840×2160), timestamp synchronized with video frame rate (30fps); establish a mapping relationship between pixel coordinates and world coordinates based on the actual geographical coordinates of the intersection, and set the coordinate transformation coefficients. (unit: ),in The calibration was performed by measuring the width of the intersection approach lanes on-site, and the calibration error was ≤0.1m.

[0025] In the software, select the target area of ​​the intersection (including all approach lanes, exit lanes, and the center area of ​​the intersection), and define the target type as "motor vehicles" and "non-motor vehicles". Motor vehicles include cars, trucks, buses, etc., and non-motor vehicles include electric bicycles, bicycles, tricycles, etc.

[0026] Furthermore, the collected video of the entire signal-controlled intersection is processed and trajectory data is extracted. Specifically, this includes: establishing a mapping relationship between video pixel coordinates and world coordinates, and calibrating the coordinate transformation coefficient through on-site measurement; extracting target features using a target detection algorithm and setting a detection confidence threshold; continuously tracking the target using the DeepSORT tracking algorithm, and predicting the target position using a Kalman filter algorithm when the target occlusion time does not exceed a preset threshold; removing outliers and performing interpolation smoothing on the original trajectory data to obtain continuous vehicle and non-vehicle trajectory data; and calculating the instantaneous velocity, instantaneous acceleration, and heading angle motion parameters of each target based on the optimized trajectory data.

[0027] The software's built-in target detection algorithm is used to extract target features (such as contour, size, and color), and the detection confidence threshold is set to 0.85 to ensure the accuracy of target recognition. The Deep SORT tracking algorithm is used to continuously track the detected targets, and the tracking IOU threshold is set to 0.5. When the target occlusion time is ≤3s, the Kalman filter algorithm is used to predict the target position to ensure the continuity of tracking.

[0028] In a specific embodiment, the Deep SORT tracking algorithm executes as follows: For each target detected by YOLOv3, appearance features are extracted using a convolutional neural network (CNN) to extract a high-dimensional feature vector. The feature dimension is 128, achieving a unique representation of the target's appearance and reducing the probability of ID switching caused by occlusion or similar targets. Based on the target's motion state in the previous frame, the position, velocity, and other motion states of the target in the current frame are predicted using a Kalman filter algorithm, providing prior information for target matching. A cascaded matching strategy is adopted, first calculating the IOU similarity between the detected bounding box and the predicted bounding box in the current frame, then calculating the cosine similarity of the appearance feature vectors, fusing the two methods. Similarity is used to obtain a comprehensive matching score. A matching threshold of 0.6 is set. If the score is higher than the threshold, the target ID matching is completed. A continuous trajectory file is created for each successfully matched target, recording the target's historical motion state and appearance features. Unmatched detection boxes are identified as new targets and assigned a new tracking ID. Unmatched prediction boxes are identified as targets that are temporarily occluded. The trajectory file is retained and prediction continues. When the occluded target reappears and a match is completed, the motion state and appearance features of the trajectory file are updated based on the detection results of the current frame to achieve continuous tracking. When the target exceeds the detection area or the occlusion time exceeds 3 seconds, the corresponding trajectory file is deleted.

[0029] Furthermore, both the state equation and the observation equation of the Kalman filter are designed based on this state vector, as follows: State vector definition: ,in Let x and y be the x and y coordinates of the target in the world coordinate system at time k. Let k be the target's velocities in the horizontal and vertical directions; State equation: , used to predict the target state at time k based on the state at time k-1; State transition matrix: , Where T is the inter-frame time interval, which is synchronized with the video frame rate and has a value of 1 / 30s; The process noise covariance matrix Q is a diagonal matrix, and its values ​​are... It characterizes the random disturbances during the target's motion process and the non-uniform motion characteristics of the non-target in the adapter; Observation equation: It is used to match the predicted state vector with the actual detected target location, and the observation matrix is ​​used to... Since actual detection can only obtain the target's location information, the location dimension is extracted from the state vector through the observation matrix; Observation noise covariance matrix This characterizes the positional error during target detection, and the detection accuracy is calibrated based on Datafromsky software to ensure the accuracy of the filtered prediction.

[0030] Furthermore, the built-in software automatically outputs the raw trajectory data for each target, including the target ID and timestamp. World coordinates ;Optimize the original trajectory data and remove outliers (using...) The criteria are to remove points that deviate from the trajectory mean by more than three standard deviations. Then, the missing data points are supplemented by cubic spline interpolation algorithm to finally obtain smooth and continuous machine and non-machine trajectory data.

[0031] Based on the optimized trajectory data, the motion parameters of each target are calculated, including instantaneous velocity. Instantaneous acceleration Heading angle The calculation formula is as follows: , , , in, The time interval is 1 / 30 of a second.

[0032] Furthermore, the database is built using MySQL and contains at least a basic information table, a trajectory data table, and a traffic signal table. The basic information table stores the data collection task identifier, collection time, and traffic scene information. The trajectory data table stores the target identifier, target type, timestamp, world coordinates, and speed information. The traffic signal table stores the approach direction, signal phase, and phase duration information. The tables in the database are linked through the data collection identifier.

[0033] A MySQL database is used, named "Signalized Intersection_Trajectory_DB", containing three core data tables: a basic information table, a trajectory data table, and a traffic signal table. The structure of each table is as follows: The BasicInfo table includes: Acquisition ID: a unique identifier for each acquisition task; Intersection ID: a unique identifier for the target intersection; Acquisition Time: DATETIME, the specific time of video acquisition; Traffic Scenario: morning rush hour / evening rush hour / off-peak / emergency event; Video Path: the storage path for the pre-processed video.

[0034] The trajectory data table includes: Trajectory ID: uniquely identifies each motorized / non-motorized trajectory; Collection ID: associated with the basic information table; Target type: motorized / non-motorized vehicle; Timestamp t: the time corresponding to the trajectory point; Coordinate x: world coordinates. Value; coordinate y: world coordinates value Velocity v: instantaneous velocity .

[0035] The traffic signal table includes: Acquisition ID (associated with the basic information table); Approach direction: East / South / West / North; Signal phase: Red / Yellow / Green; Phase start time: DATETIME (the start time of the current phase); Phase duration: the duration of the current phase. .

[0036] The extracted motor vehicle and non-motor vehicle trajectory data, acquisition unit data, and traffic signal data are entered into the database according to the above table structure. Data integrity verification (ensuring that each trajectory data is associated with the corresponding acquisition ID and traffic signal data) and data consistency verification (ensuring that the coordinate data is within a reasonable range at the intersection) are used to mark unqualified data and re-extract or supplement it. Finally, a motor vehicle and non-motor vehicle trajectory database containing multiple scenarios and types is formed.

[0037] Statistical analysis was performed on the data in the database to generate the following descriptive statistical table, which intuitively reflects the data distribution.

[0038] Furthermore, the obstacle avoidance decision model constructed by the deep deterministic policy gradient algorithm includes a policy network and a value network, both of which are fully connected neural networks. The policy network takes the environment state vector as input and outputs a continuous obstacle avoidance action vector. The value network takes the environment state vector and action vector as input and outputs the action value.

[0039] A Deep Deterministic Policy Gradient (DDPG) algorithm is used to construct an obstacle avoidance decision model, comprising four core parts: an agent (the vehicle to be avoided), an environment, a policy network, and a value network. The functions of each part are as follows: Agent: The vehicle to be avoided, whose core task is to perceive the environmental state and execute obstacle avoidance actions to achieve safe driving. Environment: A simulation environment of a non-steady-state traffic flow signal-controlled intersection built based on a database of motorized and non-motorized vehicle trajectories. The environmental state changes dynamically over time and includes dynamic elements such as motorized vehicles, non-motorized vehicles, and traffic signals. Policy Network: A fully connected neural network. The input is an environmental state vector, and the output is a continuous obstacle avoidance action vector. The network structure is: Input layer (24 neurons) → Hidden layer 1 (64 neurons, ReLU activation function) → Hidden layer 2 (32 neurons, ReLU activation function) → Output layer (2 neurons, tanh activation function). Value Network: A fully connected neural network. The input is an environmental state vector and an action vector, and the output is the action value. The network structure is consistent with the policy network.

[0040] Furthermore, the environmental state vector is 38-dimensional, specifically including: 4-dimensional vehicle state: vehicle position, speed, acceleration, and heading angle; 25-dimensional surrounding obstacle state: selecting the k nearest obstacles within a preset range around the vehicle, each obstacle including position (x, y), speed, acceleration, and type information, where the type distinguishes between motor vehicles and non-motor vehicles; 5-dimensional motor vehicle and non-motor vehicle trajectory features: extracting the trajectory curvature of surrounding obstacles within a preset time period; 2-dimensional traffic signal state: the current signal phase and remaining phase time of the approach lane where the vehicle is located; and 2-dimensional intersection road features: the width of the lane where the vehicle is located and the distance from the center of the intersection to the vehicle.

[0041] State space, action space, and reward function design: State space Define the state vector The total dimensions of the state space are 4 + 25 + 5 + 2 + 2 = 38.

[0042] The specific clustering and mapping details are as follows: Vehicle status, including vehicle location. ,speed acceleration Heading angle ; : Surrounding obstacle status, select the k nearest obstacles within a 50m radius around the vehicle. Each obstacle state includes its position. ,speed acceleration ,type ( For motor vehicles, (For non-motorized vehicles), a total of 5×5=25 dimensions; for the k nearest obstacles in the collected full-domain video of the signal-controlled intersection, k is set to 5. This value is based on the statistical law of real non-motorized vehicle trajectory data of non-steady-state traffic flow signal intersection, the engineering requirements of real-time obstacle avoidance decision, quantitative verification by multiple sets of simulation comparison experiments, and the determination of intersection road geometric features. It is the optimal value to cover core collision risks and balance model accuracy and inference efficiency. It can cover more than 95% of obstacle avoidance requirements in non-steady-state scenarios and meet the real-time requirement of vehicle controller inference delay ≤50ms. The system extracts the trajectory curvature of surrounding obstacles within the last 10 seconds (reflecting the obstacle's driving trend) across five dimensions. This extraction and processing is only performed on the top k=5 nearest obstacles. The trajectory data is derived from real-time perception and tracking results of the vehicle's multi-sensor system (LiDAR + HD camera + millimeter-wave radar), not directly from a historical obstacle trajectory database. The historical database serves only as a scene-level and feature-level completion reference during brief sensor occlusion (target-level matching without specific obstacles). Real-time perception calculation resumes immediately after completion. The average trajectory curvature of each of the top 5 nearest obstacles is calculated independently over the last 10 seconds. These five average curvature values ​​are then directly integrated into a 5-dimensional obstacle trajectory feature vector. The trajectory curvature is updated synchronously with the vehicle's sensor perception frequency at 100ms / time to ensure the features accurately reflect the real-time driving trend of the obstacles. Traffic signal status, including the current signal phase of the approach lane where the vehicle is located. ( For red light, Yellow light (Green light) and remaining phase time ; Intersection road characteristics, including the width of the lane where the vehicle is located. Distance from the center of the intersection to your vehicle .

[0043] Specifically, following the principles of not compressing core features, retaining high-risk features in detail, fusing low-risk features, and reasonably expanding auxiliary features, the original 38-dimensional state vector is accurately mapped to a 24-dimensional strategy network input: The vehicle's 4-dimensional core features (position, speed, acceleration, heading angle) are retained one-to-one; the 25-dimensional features of surrounding obstacles are fused into 12 dimensions based on collision risk weighting, with the first three high-risk obstacles retaining their complete 15-dimensional features, and the last two low-risk obstacles fused into a 2-dimensional comprehensive risk feature; the 5-dimensional curvature features of motorized and non-motorized vehicle trajectories are reorganized into 3 dimensions (lateral, longitudinal, and comprehensive rate of change) based on motion trends; the 2-dimensional features of traffic signals (phase and remaining time) are accurately mapped; and the 2-dimensional features of intersection roads are linearly expanded into 3 dimensions through traffic margin. This ultimately achieves a 24-dimensional mapping, ensuring no loss of key information.

[0044] Furthermore, the continuous obstacle avoidance action vector is 2-dimensional, including: speed adjustment, with a value range of [-5, 2] and a unit of m / s; and heading angle adjustment, with a value range of [-15°, 15°].

[0045] Specifically, define continuous action vectors ,in: The vehicle speed adjustment range is [-5, 2] m / s, with a negative sign indicating deceleration and a positive sign indicating acceleration. : Adjustment of the vehicle's heading angle, with a range of [-15°, 15°]. A negative sign indicates a left turn, and a positive sign indicates a right turn.

[0046] Furthermore, the reward function used to train the obstacle avoidance decision model is a weighted average of safety rewards, efficiency rewards, and rule rewards. Safety rewards are calculated based on the collision risk between the vehicle and obstacles. Positive rewards are given when there is no collision risk, negative rewards are given when there is a collision risk, and maximum negative rewards are given when a collision occurs. Efficiency rewards are calculated based on the deviation between the vehicle speed and the preset expected speed. Rule rewards are calculated based on the vehicle's compliance with traffic signals and lane keeping.

[0047] reward function Taking into account obstacle avoidance safety, driving efficiency, and traffic rule compliance, the reward function is designed as follows: , in, For the weighting coefficients, satisfying The value was optimized through experiments. , , .

[0048] Safety Rewards Based on the speed obstacle theory, the collision risk value r between the vehicle and the obstacle is calculated. When there is no collision risk, When there is a risk of collision, When a collision occurs, .

[0049] , in The minimum safe distance between the vehicle and the obstacle, when hour, No risk of collision; when hour, There is a risk of collision.

[0050] Efficiency Rewards Encourage vehicles to maintain a reasonable speed. ( (This refers to the current speed of the vehicle, in m / s). When exceeding the intersection speed limit (e.g., 50km / h = 13.9m / s), .

[0051] Rules and Rewards Encourage drivers to obey traffic rules. When the light is green at the entrance where your vehicle is located, When the light is red or yellow, When a car crosses the lane line, .

[0052] Furthermore, the deployment on the vehicle-mounted controller includes: real-time perception fusion: collecting driving status and surrounding obstacle status through multiple vehicle sensors, and obtaining real-time traffic signal status through the vehicle-road cooperative module, and fusing multi-source data using a filtering algorithm; state vector construction: constructing real-time state vectors based on the fused data, and standardizing them using normalization parameters consistent with model training; optimal action generation: inputting the standardized state vectors into the deployed DDPG model to generate the optimal obstacle avoidance action; execution control and feedback adjustment: converting action commands into control signals for the braking and steering systems, and collecting the actual state of the vehicle after execution in real time. When the deviation from the model's predicted state exceeds a preset threshold, the control signal is corrected using a PID adjustment algorithm.

[0053] Furthermore, the high-altitude video acquisition equipment is a drone equipped with a high-definition camera with a resolution of 3840×2160 and a frame rate of 30fps.

[0054] Specifically, a drone equipped with 4K high-definition video recording capabilities and adjustable frame rate was selected. Camera parameters were set as follows: resolution 3840×2160, frame rate 30fps, exposure time 1 / 50s, and automatic ISO sensitivity adjustment. GPS positioning and attitude stabilization were also enabled on the drone to ensure smooth video capture with a positioning accuracy error ≤0.5m. The drone's circular flight radius was half the diagonal length of the intersection, and the flight altitude was set at 150-200m. This altitude range ensures both high resolution and full coverage. When hovering, the hovering point was directly above the center of the intersection, also at an altitude of 150-200m.

[0055] Furthermore, it also includes a step of multi-dimensional evaluation of the trained obstacle avoidance decision model: using four dimensions—safety evaluation, efficiency evaluation, comfort evaluation, and rule compliance evaluation—assigning preset weights to calculate a comprehensive evaluation score to quantitatively assess the obstacle avoidance effect.

[0056] This module constructs a standardized database of vehicle and non-vehicle trajectories at signalized intersections based on extracted vehicle and non-vehicle trajectory data, providing data support for deep reinforcement learning model training. The specific steps are as follows: Data preprocessing: 70% of the data from the non-motorized vehicle trajectory database was randomly selected as the training set, and 30% as the test set; the state vectors were standardized by using min-max normalization to map all state parameters to the [0,1] interval, as shown in the following formula: , in These are the original state parameters. These are the minimum and maximum values ​​of the parameter in the database, respectively.

[0057] Model initialization and parameter configuration: Initialize the parameters of the policy network and value network. Initialize the weights of the fully connected layers using the initialization method, and initialize the bias terms to 0. Set supplementary network structure parameters: Add a batch normalization layer (BN layer) between hidden layer 1 and hidden layer 2, and set the momentum parameter to 0.9 to prevent overfitting. The core training parameter configuration is as follows:

[0058] Experience replay pool capacity: 100 empirical samples; Learning rate: policy network learning rate Value network learning rate All use the Adam optimizer; discount factor (Weight representing future rewards); Soft update coefficient (Control the rate at which the target network updates to the online network); Exploratory noise: Time-varying Gaussian noise is used, with an initial intensity of... The noise level decreases linearly with training iterations, with a decay coefficient of 0.9995, resulting in the lowest noise intensity. Batch sampling size: 64 samples / batch; Target network update frequency: Target network parameters are updated once per iteration. The initial parameters of the target policy network and target value network are completely consistent with those of the online network.

[0059] Complete training steps: Simulation environment initialization: Based on the non-steady-state traffic flow simulation environment constructed from the vehicle and non-motorized vehicle trajectory database, load the geographical parameters of the target intersection (lane width, intersection size) and traffic signal timing rules, randomly initialize the initial position of the vehicle (within 50m after the stop line of the approach lane), initial speed (0-5m / s), and the initial state of surrounding vehicle and non-motorized vehicle targets (generated based on the trajectory statistical distribution of the corresponding scene in the database), and obtain the initial environmental state. .

[0060] Exploration and experience collection: The policy network based on the current state Generate initial actions Time-varying Gaussian noise is superimposed to enhance exploration capabilities; the agent executes actions. Subsequently, the environment calculates the changes in the motion state of the vehicle and surrounding targets according to the machine-non-machine interaction rules, and outputs an instantaneous reward value. and the state at the next moment ; Determine if the termination condition is triggered (vehicle reaches the end of the intersection exit lane, a collision occurs, the vehicle exceeds the intersection boundary, or the duration of a single interaction is ≥60s). If triggered, mark the end of this training round; otherwise, repeat this step; Experience playback and storage: Store the experience samples (S) of each step. t A t R t ,St+1 The data (,done) is stored in the experience replay pool, where done is a termination status marker (done=True indicates termination, done=False indicates continuation); Batch training is triggered when the number of samples in the experience replay pool is ≥1000, batch training is started, and 64 samples are randomly selected using the Priority Experience Replay (PER) strategy. The sample priority is determined by the absolute value of the TD error, and the priority weight parameter is... Importance sampling weights are normalized. Value Network Update: Calculating Target Value ,in Output to the target value network. The optimal action for the next time step is generated for the target policy network; the error between the predicted value and the target value is calculated using the MSE loss function, and the value network parameters are updated through backpropagation. The loss function formula is as follows: , Policy network update: The policy gradient ascent method is used to maximize the expected value function, calculate the policy gradient, and update the policy network parameters. The gradient calculation formula is as follows: , in The action output by the policy network; Target network soft update: A soft update strategy is used to synchronize the parameters of the target network and the online network. The update formula is as follows: , , Training termination criteria: Repeat the exploration and experience collection steps – soft update of the target network – until one of the following termination conditions is met: the number of training iterations reaches 1×10. 5 The average loss value of each training session and 100 consecutive training sessions is ≤0.01, and the average reward value of 100 consecutive training sessions is stable in the range of [8-12] (the standard for judging the stability of the reward value is: standard deviation ≤0.5).

[0061] Simulation Testing Process: Test Environment Setup: A test environment replicating the real intersection at a 1:1 scale was built using the CARLA simulation platform. High-resolution maps, lane markings, and traffic signal facility models of the target intersection were imported. Based on real data from the motorized and non-motorized vehicle trajectory database, four typical non-steady-state test scenarios were configured: Scenario 1 - Morning Rush Hour (Sunny Day) (Motorized vehicle traffic 1200 pcu / h, Non-motorized vehicle traffic 800 pcu / h), Scenario 2 - Evening Rush Hour (Rainy Day) (Motorized vehicle traffic 1100 pcu / h, Non-motorized vehicle traffic 850 pcu / h), Scenario 3 - Off-Peak Accident (Motorized vehicle traffic 600 pcu / h, Non-motorized vehicle traffic 400 pcu / h, One truck scrapes and occupies the lane at the approach), Scenario 4 - Snowy Off-Peak (Motorized vehicle traffic 500 pcu / h, Non-motorized vehicle traffic 300 pcu / h, Road surface adhesion coefficient reduced to 0.3). Test parameter settings: Turn off exploration noise during training ( To ensure the model outputs the optimal decision; each test scenario is repeated 50 times, with each test lasting ≤60 seconds; test equipment configuration: CPU is Intel Core i9-13900K, GPU is NVIDIA RTX 4090, memory is 64GB, to ensure simulation real-time performance (frame rate ≥25fps). Test Execution: Scene Initialization: Randomly generate the initial state of the vehicle and the initial positions and velocities of surrounding motorized and non-motorized targets (consistent with the trajectory statistical characteristics of the corresponding scene); Model Deployment: Import the trained DDPG obstacle avoidance model into the simulation platform and establish a real-time data interaction interface between the model and the simulation environment; Obstacle Avoidance Execution: The vehicle obtains the real-time state vector through the sensor simulation module. Input model to generate obstacle avoidance actions Control the vehicle to complete speed and heading angle adjustments; Data recording: Real-time recording of the obstacle avoidance sequence, vehicle and obstacle motion parameters, reward value changes and test results (successful obstacle avoidance / collision) for each test. Comparative Experiment Design: Three sets of comparative experiments were set up to verify the model performance: Control Group 1 - traditional rule-based obstacle avoidance method (preset safe distance threshold, emergency braking after triggering the threshold), Control Group 2 - reinforcement learning obstacle avoidance method based on DQN algorithm (discrete action space), Experimental Group - obstacle avoidance method based on DDPG of this invention (continuous action space). The three sets of experiments used the same test scenario and initial conditions.

[0062] This invention constructs a multi-dimensional safety obstacle avoidance evaluation system to quantitatively assess obstacle avoidance effectiveness, and clarifies the execution process of actual obstacle avoidance decisions, as follows: Multi-dimensional evaluation system for obstacle avoidance safety: Safety evaluation (weight 0.5): Core indicators include ① Obstacle avoidance success rate = number of successful obstacle avoidances / total number of tests × 100%, the criterion for successful obstacle avoidance is that the vehicle passes through the intersection smoothly without colliding with any obstacles; ② Minimum safe distance = the shortest distance between the vehicle and the obstacle during the test, the passing standard is ≥2m; ③ Collision risk time = the cumulative time when the distance between the vehicle and the obstacle is <5m, the passing standard is ≤2s; ④ Collision rate = number of collisions / total number of tests × 100%; Efficiency evaluation (weight 0.2): Core indicators include ① Average traffic speed = Total distance traveled by vehicle through the intersection / Total travel time; ② Traffic delay = Actual travel time in the test scenario - Travel time in free flow state, with free flow speed calculated based on an intersection speed limit of 50km / h. Comfort evaluation (weight 0.15): The core indicators are ① maximum absolute value of longitudinal acceleration ≤ 3m / s² (avoiding sudden acceleration / sudden braking); ② rate of change of acceleration ≤ 1.5m / s³ (characterizing smoothness of movement); ③ rate of change of maximum heading angle ≤ 5° / s (avoiding sharp turns); Traffic rule compliance evaluation (weight 0.15): Core indicators include ① number of traffic rule violations (cumulative number of violations such as running red lights, crossing lane lines, and driving against traffic); ② signal phase matching rate = number of times passing through green light phase / total number of times passing through × 100%; Evaluation results calculation: The weighted summation method is used to calculate the comprehensive evaluation score. The score = 0.5 × safety score + 0.2 × efficiency score + 0.15 × comfort score + 0.15 × rule compliance score. The scores of each dimension are linearly normalized to the range of [0,100]. A comprehensive score ≥85 is excellent, 70-84 is good, and <70 is unqualified.

[0063] Actual obstacle avoidance decision execution process: Real-time perception fusion: Vehicle status (position, speed, acceleration, heading angle) and surrounding obstacle status (position, speed, type) are collected through onboard sensors (LiDAR + HD camera + millimeter-wave radar). Simultaneously, real-time traffic signal status (phase, remaining duration) is obtained through the vehicle-road cooperative module. A Kalman filter algorithm is used to fuse multi-sensor data, eliminating abnormal data and improving perception accuracy (position error ≤ 0.3m, speed error ≤ 0.2m / s). State vector construction and standardization: A 38-dimensional real-time state vector St is constructed based on the perception fusion data. The state vector is standardized using the min-max normalization parameters determined during training (based on the maximum and minimum values ​​of each parameter statistically obtained from the vehicle-non-vehicle trajectory database) to ensure the input format is consistent with model training. Optimal action generation: The standardized state vector is input into the DDPG model deployed in the onboard controller (MCU). The model calculates and outputs the optimal obstacle avoidance action through forward propagation. Output delay ≤50ms; Execution control and feedback adjustment: The on-board controller converts action commands into control signals for the braking system and steering system. Corresponding braking pressure / throttle opening, (corresponding to the steering motor rotation angle), control the vehicle to perform obstacle avoidance actions; at the same time, the controller collects the actual state of the vehicle after performing the action in real time and compares it with the model prediction state. If the deviation is >5%, the control signal is corrected by the PID adjustment algorithm to ensure that the obstacle avoidance action is executed accurately; dynamic update and cancellation judgment: during the obstacle avoidance process, steps (1)-(4) are repeated continuously to dynamically adjust the obstacle avoidance action; when the vehicle successfully passes through the intersection exit lane (reaches the preset end point) or triggers the emergency termination condition (collision risk time >3s), the obstacle avoidance decision process is terminated. If it is an emergency termination, the warning system is triggered to remind the driver.

[0064] Example 1: Video Acquisition at Intersections Based on Unmanned Aerial Vehicle (UAV) for Non-steady-state Traffic Flow Signal Control This embodiment uses a DJI Phantom 4 RTK drone equipped with a 4K high-definition camera. The target intersection is selected as the intersection of a main urban road and a secondary road (intersection ID: INT-001). The specific implementation steps are as follows: Data collection area planning: The diagonal length of the intersection is 80m. The drone is set to fly in a circular pattern with a radius of 40m, a flight altitude of 180m, and a flight speed of 5m / s to ensure coverage of all entrance lanes, exit lanes, and the central area of ​​the intersection. Device parameter configuration: camera resolution set to 3840×2160, frame rate 30fps, exposure time 1 / 50s, GPS positioning and attitude stabilization function enabled, positioning accuracy error controlled within 0.5m; Scene data collection: Two non-steady-state scenarios were collected: morning rush hour (7:30-8:00) in sunny weather and evening rush hour (17:30-18:00) in rainy weather. Each scenario was collected for 30 minutes, and metadata such as traffic signal timing (40s green light, 30s red light, 3s yellow light) and weather conditions were recorded simultaneously. Video preprocessing: The original video was preprocessed using the OpenCV open-source library, converted to MP4 format, and denoised using 3×3 Gaussian filtering. Image stabilization correction was performed based on the SIFT feature point matching algorithm. The video clarity after preprocessing met the requirements for trajectory extraction.

[0065] Example 2: Extraction of Non-Trajectory of Cross-Traffic Machine Based on Datafromsky Signal Control This embodiment is based on the video acquired in Embodiment 1, and uses Datafromsky version 2.5 software to extract vehicle and non-vehicle trajectories. The specific implementation steps are as follows: Video import and parameter settings: Import the pre-processed MP4 video, set the pixel size to 3840×2160, and the timestamp to 30fps; the actual measured width of the intersection approach lane is 12m, which corresponds to a video pixel width of 240 pixels. Calculate the coordinate transformation coefficient k=12 / 240=0.05m / pixel; Target area and type definition: Select the entire intersection area as the target area, and define the target type as "motor vehicles" and "non-motor vehicles". Motor vehicles include cars and buses, and non-motor vehicles include electric bicycles and bicycles. Feature extraction and target tracking: The YOLOv3 detection confidence threshold is set to 0.85 and the Deep SORT tracking IOU threshold is set to 0.5. Non-motorized targets in the video are detected and tracked. When the target occlusion time is ≤3s, the target position is predicted by Kalman filtering. Trajectory Extraction and Optimization: The software outputs raw trajectory data and uses... The criteria remove outliers (such as coordinate abrupt changes caused by occlusion), supplement missing data points through cubic spline interpolation, calculate parameters such as velocity, acceleration, and heading angle of each target, and finally obtain smooth non-vehicle trajectory data with a trajectory extraction accuracy of 98.5%.

[0066] Table 1. Descriptive statistics of continuous variable data Table 2 Descriptive statistics of categorical variable data Example 3: Construction of Non-Trajectory Database for Signal Control Intersection Machines This embodiment uses a MySQL 8.0 database and constructs the database based on the trajectory data extracted in Embodiment 2. The specific implementation steps are as follows: Database architecture design: Create a database "SignalizedIntersection_Trajectory_DB" containing basic information tables, trajectory data tables, and traffic signal tables; Data entry: Enter the data acquisition metadata of Example 1, the trajectory data and traffic signal data of Example 2 into the corresponding data table, and establish the association of acquisition ID; Quality control: SQL statements were used to verify data integrity and consistency, 3 incomplete trajectory data were removed, and 2 abnormal trajectory data were marked; Data statistics: Descriptive statistics were performed on the database data. The descriptive statistics of the key variables are shown in Tables 1 and 2.

[0067] Example 4: A Vehicle Safety Obstacle Avoidance Method Based on Deep Reinforcement Learning and Motor Vehicle Trajectories This embodiment is based on the database built in Embodiment 3, and uses the Python programming language and the TensorFlow framework to implement the DDPG algorithm. The specific implementation steps are as follows: Data preprocessing: 70% of the data in the database is selected as the training set and 30% as the test set. Min-max normalization is used to map the state vector to the [0,1] interval. Model initialization: Initialize the parameters of the policy network and value network; the experience replay pool capacity is 1×10. 6 Learning rate: 1×10⁻ for policy network 4 Value network 1×10⁻³, discount factor 0.99, soft update coefficient 1×10⁻³; Model training: Iterative training 1×10 5 Each iteration has a batch size of 64, and the loss function gradually converges to 0.008 during training. Model Testing: Four types of unsteady-state scenarios were tested in the CARLA simulation testing environment, with each scenario repeated 50 times. The model performance was verified through comparative experiments, and the test results are shown in the table below: Table 3 Model Test Results Note: The calculation of core indicators is as follows: 1. Obstacle avoidance success rate (%) = Number of successful obstacle avoidances / Total number of tests × 100%, where the criterion for "successful obstacle avoidance" is that the vehicle passes through the intersection smoothly without colliding with any obstacles; 2. Collision rate (%) = Number of collisions / Total number of tests × 100%; 3. Average travel delay (s) = Average actual travel time in each test scenario - Travel time in free flow state, where the free flow speed is calculated based on the intersection speed limit of 50km / h (i.e., 13.9m / s); 4. Comprehensive evaluation score = 0.5 × Safety score + 0.2 × Efficiency score + 0.15 × Comfort score + 0.15 × Rule compliance score, and the scores of each dimension are linearly normalized to the range [0,100].

[0068] Obstacle avoidance can be achieved through the attachment Figure 10 (Obstacle avoidance effect illustration) Visual demonstration: Figure 10The diagram clearly marks the initial position of the vehicle, the target trajectory, and the positions and trajectories of obstacles (motorized / non-motorized vehicles). The red dashed line represents the vehicle's trajectory without obstacle avoidance (posing a collision risk), the green solid line represents the obstacle avoidance trajectory using the method of this invention, and the yellow area represents the collision risk zone. As shown in the diagram, the method of this invention enables the vehicle to smoothly adjust its speed and heading angle, avoiding collision risk zones and accurately navigating intersections along a safe trajectory. Furthermore, based on the data in the table above, the collision rate of the model of this invention is reduced by more than 81% (compared to the traditional rule-based method), the obstacle avoidance success rate is improved by more than 5.7% (compared to the DQN algorithm), and the overall evaluation score reaches 89.6 points (excellent level), with all performance indicators superior to the comparative methods.

[0069] Actual Implementation: The model is deployed to the vehicle controller, and the actual obstacle avoidance execution process strictly follows the attached... Figure 5 (Diagram of vehicle obstacle avoidance execution process) The specific process is as follows: First, environmental data is collected and fused through onboard multi-sensor data to construct a standardized state vector input model; after the model outputs the optimal obstacle avoidance action, the controller converts the action command into braking and steering control signals, and simultaneously corrects the execution deviation through PID control; during the obstacle avoidance process, the environmental state is continuously and dynamically updated and the action is adjusted until the vehicle successfully passes through the intersection. In actual testing, the vehicle to be avoided was able to complete the obstacle avoidance action smoothly without obvious jerking, with good driving comfort, and no traffic rule violations throughout the process, achieving a consistency of over 97% with the simulation test results.

[0070] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections, characterized in that, Includes the following steps: The entire video of the signal control intersection is acquired by high-altitude video acquisition equipment. Target detection and tracking algorithms are used to extract continuous driving trajectory data of motor vehicles and non-motor vehicles in the video, and the trajectory data is optimized and motion parameters are calculated. The extracted trajectory data and related metadata are stored in the database to form a standardized machine-non-vehicle trajectory database containing multiple scenarios and types of machine-non-vehicle trajectories; Based on the real vehicle-non-vehicle interaction trajectory data stored in the standardized vehicle-non-vehicle trajectory database, a simulation environment for simulating unsteady traffic flow is constructed, and a deep deterministic policy gradient algorithm is used to train an obstacle avoidance decision model in this simulation environment. The model takes the vehicle state, the state of surrounding obstacles, vehicle-non-vehicle trajectory features, traffic signal state, and intersection road features as inputs, and outputs continuous vehicle speed adjustment and heading angle adjustment. The vehicle-non-vehicle trajectory features are obtained by extracting the trajectory curvature of surrounding obstacles within a preset time period, and are used to characterize the uncertainty of obstacle driving trends under unsteady traffic flow. The trained model is deployed on the vehicle controller, and combined with real-time multi-sensor perception fusion data, obstacle avoidance action commands are generated and the vehicle is controlled to execute them.

2. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 1, characterized in that, The process of processing the acquired video and extracting trajectory data specifically includes: Establish the mapping relationship between video pixel coordinates and world coordinates, and calibrate the coordinate transformation coefficients through on-site measurements; Target features are extracted using a target detection algorithm, and a detection confidence threshold is set. The DeepSORT tracking algorithm is used to continuously track the target. When the target occlusion time does not exceed a preset threshold, the Kalman filter algorithm is used to predict the target position. Outlier removal and interpolation smoothing are performed on the original trajectory data to obtain continuous non-motor trajectory data; Based on the optimized trajectory data, the instantaneous velocity, instantaneous acceleration, and heading angle motion parameters of each target are calculated.

3. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 1, characterized in that, The database is built using MySQL and includes a basic information table, a trajectory data table, and a traffic signal table. The basic information table stores the data collection task identifier, collection time, and traffic scene information. The trajectory data table stores the target identifier, target type, timestamp, world coordinates, and speed information. The traffic signal table stores the approach direction, signal phase, and phase duration information. The tables in the database are linked by the data collection identifier.

4. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 1, characterized in that, The obstacle avoidance decision model constructed by the deep deterministic policy gradient algorithm includes a policy network and a value network, both of which are fully connected neural networks. The policy network takes the environment state vector as input and outputs a continuous obstacle avoidance action vector. The value network takes the environment state vector and the action vector as input and outputs the action value.

5. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 4, characterized in that, The environment state vector is 38-dimensional, specifically including: The vehicle's status in 4 dimensions includes its position, speed, acceleration, and heading angle. 25-dimensional surrounding obstacle status: Select the k nearest obstacles within a preset range around the vehicle. Each obstacle includes its position, speed, acceleration, and type information, with the type distinguishing between motor vehicles and non-motor vehicles. 5D Trajectory Features of Motor Vehicles and Non-motor Vehicles: Extracting the trajectory curvature of surrounding obstacles within a preset time period; Traffic signal status in 2D: the current signal phase and remaining phase time of the approach lane where the vehicle is located; Two-dimensional characteristics of intersection roads: the width of the lane where the vehicle is located and the distance from the center of the intersection to the vehicle.

6. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 4, characterized in that, The continuous obstacle avoidance action vector is 2-dimensional and includes: a speed adjustment amount with a value range of [-5, 2] m / s; and a heading angle adjustment amount with a value range of [-15°, 15°].

7. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 4, characterized in that, The reward function used to train the obstacle avoidance decision model is a weighted sum of safety rewards, efficiency rewards, and rule rewards; The safety reward is calculated based on the collision risk between the vehicle and the obstacle. A positive reward is given when there is no collision risk, a negative reward is given when there is a collision risk, and a maximum negative reward is given when a collision occurs. The efficiency reward is calculated based on the deviation between the vehicle's speed and the preset expected speed; The rule-based rewards are calculated based on the vehicle's compliance with traffic signals and lane keeping.

8. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 1, characterized in that, The deployment on the vehicle-mounted controller further includes: Real-time perception fusion: The vehicle collects driving status and surrounding obstacle status through multiple on-board sensors, obtains real-time traffic signal status through the vehicle-road cooperative module, and uses filtering algorithms to fuse multi-source data; State vector construction: Real-time state vectors are constructed based on the fused data and standardized using normalization parameters consistent with those used in model training; Optimal action generation: Input the standardized state vector into the deployed DDPG model to generate the optimal obstacle avoidance action; Execution control and feedback adjustment: Convert action commands into control signals for the braking and steering systems, and collect the actual state of the vehicle after execution in real time. When the deviation from the model's predicted state exceeds a preset threshold, the control signal is corrected through a PID adjustment algorithm.

9. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in claim 1, characterized in that, The high-altitude video acquisition device is a drone equipped with a high-definition camera. The high-definition camera has a resolution of 3840×2160 and a frame rate of 30fps.

10. The non-fusion perception obstacle avoidance method for non-steady-state traffic flow intersections as described in any one of claims 1 to 9, characterized in that, It also includes a step of multi-dimensional evaluation of the trained obstacle avoidance decision model: using four dimensions—safety evaluation, efficiency evaluation, comfort evaluation, and rule compliance evaluation—assigning preset weights to calculate a comprehensive evaluation score to quantitatively assess the obstacle avoidance effect.

Images