Dynamic path planning method for unmanned vehicle based on multi-source sensor fusion

Abstract

This invention relates to the field of autonomous vehicle path planning technology and discloses a dynamic path planning method for autonomous vehicles based on multi-source sensor fusion. The method includes: an environmental state dynamic encoding step, which generates an environmental state encoding vector through a dynamic encoding model; a predicted path generation step, which generates a predicted path sequence for the autonomous vehicle; a multi-dimensional path difference analysis step, which compares the predicted path sequence with real-time environmental data to generate a path difference coefficient tensor; a risk area probabilistic localization step, which generates a risk probability distribution map based on the path difference coefficient tensor through spatial environmental topology analysis; a dynamic path cost calculation step, which calculates the dynamic path cost vector by combining the risk probability distribution map and the environmental state encoding vector; a path planning strategy optimization step, which optimizes the path planning strategy through a strategy learning model; and a real-time dynamic path generation step, which generates a real-time dynamic path based on the optimized strategy.

Description

Technical Field

[0001] This invention relates to the field of unmanned vehicle path planning technology, specifically to a dynamic path planning method for unmanned vehicles based on multi-source sensor fusion. Background Technology

[0002] With the rapid development of intelligent transportation technology, autonomous vehicles, as an important component of the future transportation system, have made the reliability and safety of their autonomous navigation capabilities a core research direction. Dynamic path planning is one of the key technologies for autonomous vehicles to achieve autonomous navigation. In complex and ever-changing driving environments, autonomous vehicles need to perceive surrounding environmental information in real time, accurately determine key elements such as obstacle positions, pedestrian dynamics, and traffic light status, and plan safe and efficient driving paths accordingly.

[0003] Currently, autonomous vehicle path planning technology relies heavily on data from a single sensor for environmental perception. While visual sensors can provide rich environmental texture information, they are prone to data distortion or feature loss under poor lighting conditions (such as at night, in heavy rain, or in dense fog) or when obstructed. LiDAR can accurately acquire the 3D position information of obstacles through point cloud data, but its detection range is greatly affected by weather; in rainy or snowy weather, the laser beam is easily scattered, leading to a decrease in data accuracy. Millimeter-wave radar has strong penetration capabilities and is less affected by weather, but the data it generates has low resolution, making it difficult to accurately identify small obstacles or distinguish different types of targets. These limitations of single-sensor data make it difficult for autonomous vehicles to comprehensively and accurately perceive their surroundings in complex environments, creating potential safety hazards for subsequent path planning.

[0004] In terms of environmental condition processing, traditional methods often employ static or semi-static environmental modeling approaches, making it difficult to capture dynamic changes in the environment in real time. For example, when a sudden traffic accident causes lane congestion, pedestrians cross the road, or other vehicles suddenly change lanes, traditional models often fail to update the environmental condition quickly, resulting in a disconnect between the planned path and the actual environmental requirements. Furthermore, existing path prediction methods lack effective comparison between the generated path sequences and real-time environmental data, making it difficult to quantify the differences between the path and the actual environment, thus leaving path adjustments without a clear basis.

[0005] In the risk area location phase, existing technologies mostly rely on simple distance thresholds or rule-based judgments, failing to fully consider the spatial topology of the environment. This results in insufficient accuracy in risk area location, easily leading to missed or incorrect identifications. For example, misclassifying non-risk areas as risk areas may cause excessive route detours, or failing to identify potential risk areas in a timely manner may lead to safety accidents. Furthermore, route cost calculations often only consider single factors such as distance and time, ignoring multi-dimensional dynamic factors such as risk probability, road smoothness, and traffic flow density, resulting in an incomplete cost assessment and affecting the rationality of route planning.

[0006] At the strategy optimization level, traditional path planning strategies mostly rely on preset rules or static algorithms, lacking the ability to adaptively learn from dynamic environments. When the driving environment undergoes unknown changes, the strategy cannot be adjusted quickly, resulting in insufficient flexibility and adaptability in path planning, making it difficult to meet the driving needs of autonomous vehicles in diverse scenarios such as complex urban roads and rural trails. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic path planning method for unmanned vehicles based on multi-source sensor fusion, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a dynamic path planning method for unmanned vehicles based on multi-source sensor fusion, the method comprising:

[0009] The process includes the following steps: Real-time multi-source sensor data acquisition: Based on the autonomous vehicle's driving environment, multi-source sensor data is acquired in real time, including visual image data, LiDAR point cloud data, and millimeter-wave radar reflection data. Dynamic environmental state encoding: Based on the multi-source sensor data, an environmental state encoding vector is generated using a dynamic encoding model. Predictive path generation: Based on the environmental state encoding vector, a predicted path sequence for the autonomous vehicle is generated. Multi-dimensional path difference analysis: The predicted path sequence is compared with real-time environmental data in multiple dimensions to generate a path difference coefficient tensor. Risk area probabilistic localization: Based on the path difference coefficient tensor, a risk probability distribution map is generated through spatial environmental topology analysis. Dynamic path cost calculation: Based on the risk probability distribution map and the environmental state encoding vector, a dynamic path cost vector is calculated. Path planning strategy optimization: Based on the dynamic path cost vector, the path planning strategy is optimized using a strategy learning model. Real-time dynamic path generation: Based on the path planning strategy, a real-time dynamic path for the autonomous vehicle is generated.

[0010] Preferably, the real-time acquisition step of multi-source sensor data specifically includes:

[0011] The visual image data is subjected to illumination adaptive correction processing to generate corrected visual data; the lidar point cloud data is subjected to noise filtering processing to generate filtered point cloud data; the millimeter-wave radar reflection data is subjected to Doppler effect compensation processing to generate compensated reflection data; the corrected visual data, the filtered point cloud data, and the compensated reflection data are time-aligned and fused to generate fused sensor data.

[0012] Preferably, the dynamic encoding step of the environmental state specifically includes:

[0013] The fused sensor data is input into a feature extraction network to extract multi-scale environmental features; the multi-scale environmental features are then subjected to dimensionality reduction encoding through a dynamic encoding model to generate an environmental state encoding vector; the environmental state encoding vector includes spatial obstacle distribution features and dynamic target motion features.

[0014] Preferably, the predicted path generation step specifically includes:

[0015] Based on the environmental state encoding vector, an initial path sequence is generated through a path prediction model; the initial path sequence is then smoothed and optimized to generate a predicted path sequence; the predicted path sequence includes a sequence of path point coordinates in the time dimension.

[0016] Preferably, the multidimensional path difference analysis step specifically includes:

[0017] Extract the path point coordinate sequence from the predicted path sequence and the actual obstacle position sequence from the real-time environment data; calculate the spatial offset between the path point coordinate sequence and the actual obstacle position sequence; calculate the path execution delay and the matching degree of the dynamic target motion features in the time dimension; integrate the spatial offset, the path execution delay, and the matching degree of the dynamic target motion features into a normalized tensor to generate a path difference coefficient tensor.

[0018] Preferably, the risk area probabilistic location step specifically includes:

[0019] A spatial environment topology map of the autonomous vehicle's driving environment is constructed, the spatial environment topology map including road node connection relationships and obstacle density parameters; the path difference coefficient tensor is mapped to the corresponding node of the spatial environment topology map; the risk diffusion path is calculated through a graph structure propagation algorithm; a risk probability distribution map is generated based on the risk diffusion path, the risk probability distribution map marking the coordinates of high-risk probability areas.

[0020] Preferably, the dynamic path cost calculation step specifically includes:

[0021] Extract the coordinates of high-risk probability areas from the risk probability distribution map; calculate the path safety cost component based on the spatial obstacle distribution characteristics in the environmental state encoding vector; calculate the path efficiency cost component based on the dynamic target motion characteristics in the environmental state encoding vector; calculate the path risk cost component based on the coordinates of the high-risk probability areas; and weight and fuse the path safety cost component, the path efficiency cost component, and the path risk cost component to generate a dynamic path cost vector.

[0022] Preferably, the path planning strategy optimization step specifically includes:

[0023] The dynamic path cost vector is input into the strategy learning model; the path planning strategy is updated by the cumulative cost minimization algorithm to generate an optimized path planning strategy; the optimized path planning strategy includes path point generation rules and speed control parameters.

[0024] Preferably, the real-time dynamic path generation step specifically includes:

[0025] Based on the optimized path planning strategy, a sequence of candidate path points is generated; the candidate path point sequence is subjected to dynamic feasibility verification processing to generate a real-time dynamic path; the real-time dynamic path includes a path instruction sequence marked with timestamps.

[0026] Preferably, the method further includes:

[0027] Adaptive path adjustment step: Based on the execution feedback of the real-time dynamic path, dynamically adjust the parameters of the dynamic encoding model and the policy learning model; the adaptive path adjustment step specifically includes:

[0028] Collect the execution feedback data of the real-time dynamic path, calculate the model parameter offset based on the execution feedback data, perform incremental parameter update processing on the dynamic encoding model and the policy learning model, and output the updated dynamic encoding model and the updated policy learning model.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This dynamic path planning method for autonomous vehicles based on multi-source sensor fusion effectively overcomes the limitations of single sensors in environmental perception through the collaborative application of multi-source sensor data. Real-time acquisition and fusion of visual image data, LiDAR point cloud data, and millimeter-wave radar reflection data enable the autonomous vehicle to acquire environmental information from multiple dimensions. The texture features provided by visual data, the 3D positioning accuracy of LiDAR, and the anti-interference capability of millimeter-wave radar complement each other, maintaining stable perception of key environmental elements such as obstacles, pedestrians, and road boundaries even under conditions of changing lighting, weather interference, or occlusion, thus improving the comprehensiveness and reliability of environmental perception.

[0031] The dynamic environmental state encoding step uses a dynamic encoding model to deeply process multi-source sensor data, transforming complex environmental information into structured environmental state encoding vectors. This encoding method not only preserves static features in the environment (such as road shape and fixed obstacle positions) but also captures dynamic features in real time (such as pedestrian movement speed and vehicle direction), enabling autonomous vehicles to understand environmental states more accurately and promptly, and providing high-quality foundational data for subsequent path planning.

[0032] The combination of predicted path generation and multidimensional path difference analysis enables a quantitative assessment of the deviation between the predicted path and the actual environment. The path difference coefficient tensor comprehensively reflects the differences between the path and real-time environmental data in multiple dimensions such as spatial location, driving speed, and obstacle avoidance distance, avoiding the subjectivity and ambiguity of traditional methods in judging the rationality of the path. Through this quantitative analysis, autonomous vehicles can clearly identify potentially unreasonable parts in the predicted path, providing a clear direction for path adjustment.

[0033] The risk area probabilistic localization step generates a risk probability distribution map based on path difference coefficient tensors and spatial environment topology analysis, fully considering the spatial correlation of various elements in the environment. For example, by combining information such as the topology of road intersections and pedestrian crossing habits, high-risk areas (such as blind spots and densely populated pedestrian areas) can be located more accurately, rather than simply relying on distance thresholds. This makes the risk identification of autonomous vehicles more objective and comprehensive, providing a scientific basis for path safety assessment.

[0034] Dynamic path cost calculation combines a risk probability distribution map with an environmental state encoding vector, comprehensively considering factors such as path length, travel time, risk probability, and road conditions. The generated dynamic path cost vector can comprehensively reflect the combined cost of different paths, avoiding the one-sidedness of path planning caused by considering only one factor in traditional cost calculation. Based on this comprehensive cost assessment, autonomous vehicles can achieve a better balance between safety and efficiency when selecting paths.

[0035] The path planning strategy optimization step learns and optimizes the dynamic path cost vector through a policy learning model, enabling the path planning strategy to adapt to dynamic environments. The policy learning model continuously adjusts decision parameters from historical data and real-time feedback. When encountering new environmental changes (such as sudden obstacles or road construction), it can quickly generate a planning strategy adapted to the current environment, rather than relying on fixed rules. This dynamic optimization capability significantly improves the flexibility and adaptability of path planning, enabling autonomous vehicles to maintain good navigation performance in complex and ever-changing driving scenarios.

[0036] The real-time dynamic path generation step, based on an optimized path planning strategy, can quickly respond to environmental changes and generate real-time paths that meet the needs of the current environment. Whether it's changes in traffic flow on urban roads or sudden obstacles on rural roads, the autonomous vehicle can adjust its driving path in a timely manner, ensuring the safety and efficiency of the driving process and improving the reliability and practicality of the autonomous vehicle in real-world applications. Attached Figure Description

[0037] Figure 1 This is a schematic diagram illustrating the working principle of the unmanned vehicle dynamic path planning method based on multi-source sensor fusion described in this invention.

[0038] Figure 2 A schematic diagram illustrating the working principle of dynamic coding of environmental states;

[0039] Figure 3 A diagram illustrating the working principle of path prediction;

[0040] Figure 4 A diagram illustrating the working principle of probabilistic location of risk areas. Detailed Implementation

[0041] 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.

[0042] Please see Figure 1 This invention provides a dynamic path planning method for unmanned vehicles based on multi-source sensor fusion. This method achieves continuous dynamic optimization of path planning through real-time processing and fusion of multi-source perception data, combined with dynamic environmental state modeling and risk probability analysis. The specific implementation process is as follows:

[0043] During the autonomous vehicle's operation, visual image data, LiDAR point cloud data, and millimeter-wave radar reflection data are simultaneously collected, forming a multi-source sensor data input. After preprocessing, this raw data is input into a dynamic coding model, which extracts and compresses key environmental information, outputting an environmental state coding vector containing spatial obstacle distribution characteristics and dynamic target motion characteristics. Based on this coding vector, the path prediction module generates a predicted path sequence for a future time window.

[0044] The predicted path sequence is compared and analyzed with real-time acquired environmental data from multiple dimensions. Difference indicators such as spatial offset, time delay, and motion feature matching degree are calculated and integrated to form a path difference coefficient tensor. Using this tensor, combined with a pre-constructed spatial environment topology map, a graph structure propagation algorithm is used to identify and quantify potential risk areas in the environment, generating a risk probability distribution map. This distribution map, along with the environmental state encoding vector, is input into the dynamic path cost calculation module to calculate the path safety cost component, path efficiency cost component, and path risk cost component. These components are then weighted and fused to obtain the dynamic path cost vector. This vector drives the strategy learning model for optimization, updating the waypoint generation rules and speed control parameters to form an optimized path planning strategy.

[0045] Based on this strategy, a sequence of candidate path points is generated. After dynamic feasibility verification, a sequence of path instructions with timestamps is output as the basis for the real-time dynamic path execution of the autonomous vehicle. The entire process forms a closed loop, and the execution feedback data can be used for dynamic adjustment of model parameters.

[0046] Example 1: See Figure 2 The implementation of the real-time acquisition of multi-source sensor data begins with the parallel processing of the raw sensor data. Visual image data is input to the illumination adaptive correction module, which analyzes the pixel brightness distribution characteristics of the input image and identifies areas of uneven illumination. For identified overly dark areas, an algorithm based on regional contrast enhancement is applied to dynamically increase local brightness; for overexposed areas, a nonlinear compression method is used to restore details. The processing considers the overall lighting conditions of the scene, avoids introducing artificial artifacts, and outputs corrected visual data with color balance and clear details. The processing of LiDAR point cloud data focuses on noise suppression and data simplification. Point cloud data is first input to the statistical outlier removal module, which calculates the average Euclidean distance between each point and a specified number of neighboring points. Based on the statistical distribution characteristics of the average distance of all points, a standard deviation threshold is set to remove discrete noise points whose distances significantly deviate from the group mean. Subsequently, optionally, spatial mesh generation technology is applied to divide the three-dimensional space into uniform voxel units. Within each voxel cell, the geometric center coordinates of all contained points are calculated, and this center point replaces all original points within the cell, achieving downsampling of the point cloud data. This significantly reduces the data volume while preserving the environmental structure outline, outputting filtered point cloud data. The core of millimeter-wave radar reflection data processing lies in the accurate correction of velocity information. The Doppler effect compensation module receives millimeter-wave radar reflection data and the real-time velocity vector information of the unmanned vehicle. The module calculates the angle between the radar beam emission direction and the unmanned vehicle's velocity vector. Based on the Doppler frequency shift principle, and combining this angle with the unmanned vehicle's velocity magnitude, the target radial velocity measurement deviation caused by the unmanned vehicle's own motion is calculated. Reverse compensation correction is performed on the radial velocity value of each reflection point detected by the radar, eliminating the measurement error introduced by the vehicle's motion, and outputting compensated reflection data that more accurately reflects the target's true relative motion.

[0047] After completing the independent preprocessing described above, the time-series alignment and fusion stage begins. The goal of this stage is to unify data streams from different sensors, which may have slight timestamp differences, to the same time reference. Alignment is achieved using methods based on hardware synchronization or software interpolation. If the sensors support hardware timestamp synchronization, a precise global clock signal is used to mark the timestamp of each data packet, and data with the same timestamp are directly selected for fusion. If hardware synchronization is unavailable or inaccurate, a software interpolation method is used. Based on the acquisition timestamps of each sensor data stream, linear interpolation or spline interpolation is performed on data at adjacent time points to generate estimated sensor data values ​​at a specified target time point. The aligned, corrected visual data, filtered point cloud data, and compensated reflection data are input into the fusion module. Fusion strategies can employ early fusion or mid-term fusion. An example of early fusion is projecting the filtered point cloud data onto the image plane corresponding to the corrected visual data to generate a pixel-level depth information map, which is then stitched with the original image data to form RGB-D data. Mid-term fusion extracts features from the three data sources respectively. For example, edge detection or preliminary convolution operations are applied to the corrected visual data to extract image features; distance- or normal-vector-based feature extraction methods are applied to the filtered point cloud data; and intensity- or Doppler velocity-based feature extraction methods are applied to the compensated reflection data. These extracted heterogeneous features are then stitched or connected along the feature dimension to form fused sensor data. This data is a multi-dimensional feature set that integrates visual appearance, spatial geometry, and motion velocity information.

[0048] The dynamic encoding step of the environmental state takes fused sensor data as input. This data is first fed into a feature extraction network. This network is typically designed with a multi-branch structure to efficiently handle different types of data. One branch processes visually relevant data (such as RGB or RGB-D images), containing a multi-layer convolutional neural network structure that progressively extracts image features from local to global through convolution and pooling operations, such as edge texture information, object contours, and semantic segmentation results. Another branch processes point cloud data, possibly using a point-operation-based point cloud network structure to directly process 3D point coordinates and their attributes, extracting spatial distribution features of the point cloud, such as the shape, size, and distance information of obstacles, as well as the surface geometry of the scene. A third branch processes millimeter-wave radar data, possibly containing recurrent neural networks or temporal convolutional layers, processing sequential data of reflection points to extract target motion trajectory features, velocity distribution patterns, and stability information of dynamic targets. Each branch works independently, outputting its extracted multi-scale environmental features. These features may contain information at different spatial resolutions or temporal granularities.

[0049] The extracted multi-scale environmental features are then input into a dynamic coding model. The core task of this model is to fuse heterogeneous features from different sensors and capture the dynamic changes and dependencies of the environmental state in the spatiotemporal dimensions. The model structure can be either an attention-based sequence model or a graph neural network. If an attention-based sequence model (such as the Transformer encoder) is used, the features are serialized before being input into the model. The model calculates the association weights between different feature elements through a self-attention mechanism, highlighting important features and suppressing secondary features, while using positional encoding to preserve spatial information. If a graph neural network is used, the environmental features are represented as node features in a graph structure, and the connections (edges) between nodes can be defined based on spatial proximity or semantic relevance. The graph neural network updates the feature representation of each node by aggregating information from neighboring nodes through a message-passing mechanism. Regardless of the structure used, the dynamic coding model fuses, compresses, and abstracts high-dimensional, redundant multi-scale environmental features through multiple layers of nonlinear transformations. The model outputs a low-dimensional, dense environmental state encoding vector. This vector is a compact representation of the environmental state, with a dimension much lower than the original input features. This structured vector encodes key environmental information, which necessarily includes feature components representing the distribution of spatial obstacles, such as their location, size, and density; it also necessarily includes components representing the motion characteristics of dynamic targets, such as their instantaneous velocity, direction of motion, acceleration trend, and trajectory prediction information. This environmental state encoding vector serves as the core input for subsequent path prediction and analysis modules.

[0050] Example 2: See Figure 3 The path prediction generation step uses an environmental state encoding vector as its core input. This vector is input to the path prediction model, which is responsible for generating predictions of the driving trajectory over a future period. The path prediction model typically employs a sequence generation architecture, such as a recurrent neural network based on gated recurrent units or a Transformer model based on a self-attention mechanism. The model receives sequences of current and historical environmental state encoding vectors, capturing the temporal evolution of environmental states. The model's internal structure includes a hidden state propagation mechanism, enabling it to remember historical information and infer future trends. The model output is a sequence of predicted path point coordinates for consecutive future time steps, forming the initial path sequence. Each element in this sequence contains a three-dimensional spatial coordinate value and a corresponding timestamp, representing the expected location of the autonomous vehicle at a specific time. The initial path sequence may exhibit a discrete point distribution or contain local non-smoothness.

[0051] The initial path sequence is then processed by a smoothing optimization module. The goal of this module is to generate a continuous trajectory that conforms to vehicle dynamics constraints and provides a comfortable ride. The smoothing optimization process can employ parametric curve fitting or numerical optimization methods. If parametric curve fitting, such as B-spline curves or Bézier curves, is used, discrete initial path points are used as control points to fit a smooth, continuous curve. The curve parameters must satisfy the minimum turning radius constraint of the autonomous vehicle, avoiding sharp turns that exceed the vehicle's physical limits. If numerical optimization methods are used, such as gradient descent or quadratic programming-based optimizers, an objective function is defined. The objective function typically includes smoothness terms (such as the sum of squares of path curvature or the rate of change of direction between path points) and a fit term to the initial path (such as the sum of squares of distances between path points and initial predicted points). During optimization, constraints are simultaneously applied, including the maximum permissible curvature of path points, the minimum and maximum distances between adjacent path points (reflecting the speed range), and the minimum safe distance between path points and known static obstacles. The optimization solver searches for the path point sequence that minimizes the objective function value while satisfying all constraints. After smoothing optimization, the final predicted path sequence is output. The sequence also contains a series of path point coordinates in the time dimension, but its spatial distribution is more continuous and smooth, and it satisfies the kinematic feasibility of the vehicle.

[0052] The implementation of the multi-dimensional path difference analysis step requires the simultaneous acquisition of predicted path sequences and real-time environmental data. Real-time environmental data refers to the latest environmental perception information collected in real-time by multi-source sensors within the time window corresponding to the predicted path sequence, and then preprocessed and fused. The analysis process begins with data extraction. The predicted path sequence is extracted, containing a sequence of path point coordinates, with each coordinate point associated with its predicted timestamp. From the real-time environmental data, target detection and tracking algorithms (such as Kalman filtering based on fusion perception or multi-target tracking algorithms) are used to extract the actual obstacle location sequence. This sequence contains the location coordinates and identification information of each tracked obstacle at different timestamps.

[0053] The calculation process revolves around three core difference dimensions. The first dimension is the calculation of spatial offset. For each path point in the predicted path sequence, its corresponding timestamp is determined. At this timestamp, all existing obstacle positions are located from the actual obstacle position sequence. The Euclidean distance between the coordinates of this path point and the coordinates of each actual obstacle position is calculated. The minimum of these distances is taken as the spatial offset of the path point at the current moment. This process is repeated for all path points in the predicted path sequence, calculating the spatial offset for each point to form a spatial offset sequence. The second dimension is the calculation of path execution delay. This delay reflects the time lag introduced by the planning system processing, control command issuance, and vehicle response. The planned arrival timestamp (i.e., the predicted time) of each path point in the predicted path sequence is recorded. In actual operation, the actual timestamp of the autonomous vehicle arriving at a location near the path point (e.g., within a set tolerance range) is recorded. The difference between the actual timestamp and the planned timestamp is calculated as the path execution delay for that path point. Similarly, the path execution delay sequence is obtained by traversing the sequence. The third dimension is the calculation of the matching degree of dynamic target motion features. The generation of the predicted path sequence implicitly involves an expectation of the future motion state of dynamic targets in the environment (such as other vehicles and pedestrians). This expected information originates from the dynamic target motion features in the environmental state encoding vector. Real-time environmental data also includes dynamic target motion features updated based on the latest perception, such as instantaneous velocity vectors, acceleration, and direction of motion. For key dynamic targets (e.g., targets nearby or potentially affecting the vehicle), at the corresponding time point, the predicted motion features (e.g., predicted velocity direction) are compared with the actually perceived motion features (e.g., actual velocity direction). Matching degree calculation can employ vector similarity metrics, such as calculating the cosine similarity (dot product divided by modulus) between the predicted and actual velocity direction vectors; a value closer to 1 indicates a higher matching degree. Alternatively, the Mahalanobis distance between the predicted and actual positions can be calculated (considering motion uncertainty). Similarly, a matching degree sequence reflecting the accuracy of the motion state prediction can be obtained.

[0054] After obtaining the spatial offset sequence, path execution delay sequence, and dynamic target motion feature matching degree sequence, normalization is performed before data integration. Normalization aims to eliminate differences in units and numerical ranges between different dimensions of data, making them comparable and facilitating subsequent processing. For the spatial offset sequence, a reasonable maximum expected offset value is set based on sensor accuracy and the scene. All offsets are normalized by dividing by this maximum value, ensuring the result falls within the [0,1] interval (or adjusted according to actual conditions). For the path execution delay sequence, a maximum tolerable delay threshold is set, and the delay is normalized by dividing by this threshold. For the dynamic target motion feature matching degree sequence, if cosine similarity (range [-1,1]) is used, it can be mapped to the [0,1] interval (e.g., (similarity value + 1) / 2); if distance metric is used, a maximum expected distance is also set for normalization. The three normalized sequences—spatial offset sequence, path execution delay sequence, and dynamic target motion feature matching degree sequence—are aligned according to the same time step or path point index. These are combined into a three-dimensional data structure, namely the path difference coefficient tensor. The first dimension of this tensor corresponds to the spatial offset, the second dimension to the path execution delay, and the third dimension to the dynamic target motion feature matching degree. Each element in the tensor is a three-dimensional vector containing three normalized difference coefficients at a specific path point or time step. This tensor comprehensively quantifies the inconsistencies between the predicted path and the actual environmental perception results across multiple dimensions, providing input for subsequent risk assessment.

[0055] Example 3: See Figure 4 The risk area probabilistic localization step uses the path difference coefficient tensor as its core input. This step first constructs a spatial environment topology map describing the autonomous vehicle's driving environment. This topology map is represented using a graph data structure, where nodes represent key location points or center points of dividing areas in the environment, such as the center of a road intersection, the center point of a lane, or the center of a grid in a raster map. Edges connect nodes, representing spatial accessibility or road connectivity between them; for example, adjacent lane nodes are connected, and nodes in the same lane are connected. Each node is associated with a set of attribute data, including its spatial coordinates (such as geodetic coordinates or coordinates in a local coordinate system), obstacle density parameters calculated based on historical statistics or real-time perception (such as the frequency or average number of obstacles per unit area), and optional additional attributes such as road type (main road, ramp, etc.) and speed limits. The topology map can be constructed based on high-precision map information or dynamically generated and updated through real-time perception data when no prior map is available.

[0056] The path difference coefficient tensor needs to be mapped to the corresponding node in the spatial environment topology graph. The mapping process is based on the spatial location where the path differences occur. For each element in the path difference coefficient tensor (corresponding to a specific path point or time step), the associated spatial location coordinates (i.e., the coordinates of the predicted path point or the location of the difference calculation point) are extracted. In the spatial environment topology graph, the topology graph node with the nearest Euclidean distance to these spatial location coordinates is found. Once the nearest neighbor node is found, the three normalized difference values ​​(spatial offset, path execution delay, and dynamic target motion feature matching degree) contained in the path difference coefficient tensor element are associated with that topology graph node as a set of additional attributes. If multiple path difference points are mapped to the same topology node, the average, maximum, or other aggregation functions can be used to calculate the comprehensive difference coefficient of that node.

[0057] Based on a spatial environment topology map using mapped path difference coefficients, a graph structure propagation algorithm is used to calculate the risk diffusion path and final risk probability distribution within the environment. This algorithm simulates the process of risk propagation from high-difference areas to surrounding areas. A feasible implementation is an iterative message-passing algorithm. Initially, each topology node has an initial risk value. The value is determined by the associated comprehensive path difference coefficient, for example, by taking the weighted sum of three difference coefficients. In each subsequent iteration... In, each node The risk value is based on its neighboring nodes. Risk value of the previous iteration ,node and Distance between (or the weights of the connecting edges, which can be designed as a decreasing function of the distance, such as...) ), and nodes Its own obstacle density parameters Update. A simplified example of an update rule is as follows:

[0058]

[0059] in: Represents a node In the Risk value after the next iteration; Represents a node The set of all neighboring nodes; Represents a node With nodes The Euclidean distance between them; Represents a node Obstacle density parameters; and These are weighting coefficients between 0 and 1, controlling the impact of neighbor risk propagation and local obstacle density on the current risk, respectively, and satisfying the following conditions: ; It is a node The risk value from the previous iteration is included to retain some historical risk information. This formula reflects the characteristics of risk propagation: nearby high-risk neighbor nodes significantly increase the risk of the current node; high local obstacle density also directly increases risk; and risk has a certain degree of persistence. The algorithm iterates until the risk value changes of all nodes are less than a set threshold or the maximum number of iterations is reached, ultimately obtaining a stable risk probability value for each node. Based on the final risk probability values ​​of all nodes, a risk probability distribution map is generated. This map can directly present the risk values ​​on the topology map using node color depth or numerical labels, or it can be converted into a raster map format, dividing the environment into grids. The risk value of each grid is obtained by interpolation or averaging the risk values ​​of the nodes it contains. The map clearly marks the coordinates of areas where the risk probability exceeds a set threshold, identifying them as high-risk probability areas.

[0060] The implementation of the dynamic path cost calculation step is based on a risk probability distribution map and an environmental state encoding vector. First, the coordinates of all areas identified as high-risk probability regions and their corresponding risk probability values ​​are extracted from the risk probability distribution map. This information quantifies the likelihood of different locations in the environment posing a threat to the autonomous vehicle. Cost calculation involves the calculation of three independent components. The path safety cost component relies on the spatial obstacle distribution characteristics in the environmental state encoding vector. This component assesses the safety margin between the planned path and all known obstacles in the environment (including predicted trajectories of static and dynamic obstacles). The calculation process traverses each point on the planned path, calculating the Euclidean distance from that point to the nearest obstacle (or obstacle trajectory). When this distance is less than a preset safety distance threshold, the cost component increases significantly, with the cost increment being larger the closer the distance. Simultaneously, the geometry of the path is also considered, calculating the local curvature of the path. Excessive curvature implies sharp turns, potentially increasing the risk of loss of control or passenger discomfort; therefore, high-curvature path points also increase the safety cost component. The path efficiency cost component relies on the dynamic target motion characteristics in the environmental state encoding vector and global path information. This component focuses on the path's driving efficiency. The total length of the path is calculated; a longer path typically means a longer travel time. The travel time is estimated by combining the planned speed curve (which may be provided by the strategy or estimated based on path curvature and speed limits). The consistency between the path direction and the global target direction (e.g., a straight line from the start to the end point) is evaluated, and the deviation angle between the path point direction and the target direction is calculated; a larger deviation results in a higher efficiency cost component. Furthermore, if the dynamic target motion characteristics indicate the presence of high-speed moving objects that may affect traffic efficiency (e.g., slower vehicles ahead), path points located in potentially obstructed areas will also increase efficiency costs. The path risk cost component is calculated directly based on the extracted high-risk probability area coordinates and their probability values. This component assesses the tendency of the planned path to approach or cross high-risk areas. The distance from each point on the path to the nearest high-risk area boundary is calculated. When the distance is less than a preset risk distance threshold, the risk cost component increases, and the closer the distance or the higher the risk probability value of the target area, the greater the cost increment. If a path segment directly crosses a high-risk area, the risk cost component increases significantly based on the length of the crossed area and the average risk probability value of that area. The calculated path safety cost component (denoted as...) is then... ), path efficiency cost component (denoted as ) and path risk cost component (denoted as The data will be merged. The merging process uses a weighted summation method:

[0061]

[0062] in: It is the final generated dynamic path cost vector, which may contain cost values ​​for different segments or different evaluation points on the path; , , These are the weighting coefficients for the path safety cost component, efficiency cost component, and risk cost component. These weighting coefficients are preset or configurable parameters, and their values ​​reflect the different priorities of safety, driving efficiency, and risk avoidance in path planning. For example, in a complex urban environment, the safety weight... and risk weight The weighting might be set too high; however, in highway scenarios, efficiency weighting is less important. It may be relatively higher. The dynamic path cost vector obtained after weighted fusion. This provides a quantitative basis for optimizing subsequent path planning strategies.

[0063] Example 4: The path planning strategy optimization step uses a dynamic path cost vector as the core driving input. This vector quantifies the comprehensive cost of different path choices or path points across different dimensions (safety, efficiency, risk). This vector is input into the strategy learning model. This model is typically implemented as a parameterized function approximator, such as a deep neural network, whose input includes the current environment state encoding vector and the dynamic path cost vector. The model's goal is to output a path planning strategy that defines how to generate a specific driving path under the current environment state and cost constraints. Internally, the model is trained or optimized online using a cumulative cost minimization algorithm. This algorithm evaluates the expected cumulative cost of the candidate strategies proposed by the model over multiple time steps or up to the task objective. The calculation of the expected cumulative cost considers a discount factor, i.e., the impact of future costs on the current decision gradually decreases. The algorithm uses optimization methods (such as stochastic gradient descent, evolutionary strategies, or sampling-based optimization) to adjust the internal parameters of the strategy learning model, aiming to find a set of parameters that minimizes the expected cumulative cost corresponding to the model's output strategy. After parameter updates, the model outputs the optimized path planning strategy. This strategy specifically includes two core components: path point generation rules and speed control parameters. The path point generation rules define the decision-making logic for generating path point sequences. For example, when a static obstacle is detected on the left, the generated path point sequence should shift to the right by a specific distance; when a dynamic target is detected approaching from the front, the path point sequence should plan an avoidance trajectory in advance, which may include a combination of rules for lateral offset and longitudinal deceleration points; when approaching an intersection and needing to turn, the path point sequence should follow specific rules for turning radius and entry angle. Speed ​​control parameters define the recommended speed or acceleration curves to be used in different sections of the path. For example, in curved sections, the speed parameter should be reduced to below the set value; when a path point approaches a high-risk probability area, the speed parameter should be set to a cautious passage speed; in straight sections without obstacles, the speed parameter can be increased to an efficiency-priority speed. These rules and parameters together constitute the optimized driving decisions.

[0064] The real-time dynamic path generation step is implemented based on the optimized path planning strategy described above. This strategy drives the path generation module. The path generation module receives the latest environmental state encoding vector. According to the path point generation rules, the module generates a set of candidate path point sequences. For example, in a specific scenario: the current environmental state encoding vector indicates that there is a stationary obstacle occupying part of the lane 100 meters ahead, a vehicle is traveling in the same direction at a high speed in the adjacent lane on the left, and there is a curb on the right. According to the path point generation rule "when encountering an obstacle traveling in the same direction and there is a vehicle on the left, generate a path point sequence that slightly adjusts to the right and prepares to avoid it," the module may generate the following sequence: the first 20 path points (corresponding to the next 2 seconds) slightly adjust to the right in the current lane, maintaining a safe distance from the right curb; the next 20 path points (corresponding to the next 2 seconds) generate a path point sequence that cuts to the left into the center line of the adjacent lane after confirming that the vehicle on the left has passed; the last 20 path points (corresponding to the next 2 seconds) generate a path point sequence that returns to the original lane center line. Meanwhile, the speed control parameters are set according to the rules: maintain the current speed during the initial fine-tuning phase; slightly decelerate when merging into an adjacent lane; and resume the original speed after returning to the original lane. Refer to Table 1, which shows some candidate path point sequences and their associated speed control parameters in this scenario.

[0065] Table 1: Candidate Path Point Sequence and Velocity Control Parameters

[0066] Path point sequence index Path point coordinates (X, Y) (meters) Timestamp (seconds) Recommended speed (m / s) Association rule description 1 (10.0,3.5) 0.0 8.0 Starting point, maintain current speed 5 (50.0,3.8) 1.0 8.0 Slightly adjust to the right, away from the obstacle. 10 (100.0,4.0) 2.0 8.0 Approaching an obstacle, maintain a safe distance to the right. 15 (150.0,3.0) 3.0 7.5 Cut into the adjacent lane on the left and begin to decelerate. 20 (200.0,2.5) 4.0 7.5 Drive in the center of the adjacent lane 25 (250.0,2.8) 5.0 7.5 Preparing to return to the original lane 30 (300.0,3.5) 6.0 8.0 Drive back to the center of the original lane and resume speed.

[0067] After the candidate path point sequence is generated, it is input into the dynamic feasibility verification module for rigorous validation. This module performs multi-level checks. First, dynamic constraint verification is performed: the distance difference between adjacent path points is calculated and divided by the time difference to obtain the required average speed, and it is checked whether this speed is within the speed range allowed by the autonomous vehicle; the local curvature of the path point sequence is calculated (estimated by the angle formed by three consecutive points), and it is checked whether the curvature is less than the maximum allowable curvature of the autonomous vehicle to avoid sharp turns exceeding the vehicle's physical limits; the rate of change of direction between adjacent path points is checked whether it is less than the maximum allowable value. Second, static obstacle collision detection is performed: the trajectory formed by the path point sequence (usually represented by line segments or simple geometric shapes such as capsules) is geometrically intersected with the static obstacle map of the environment built based on real-time perception. If an intersection is detected, the path point or segment is determined to have a collision risk. Finally, dynamic obstacle spatiotemporal conflict detection is performed: the predicted trajectories of dynamic targets in the environment (such as vehicles and pedestrians) are obtained (usually also time-stamped path point sequences). The candidate path point sequence of the vehicle is compared with the predicted trajectory sequence of each dynamic target in the same time coordinate system. For each timestamp, the module calculates whether the area occupied by the vehicle (considering the vehicle's outline) overlaps with the area occupied by the dynamic target. If an overlap is detected at a future time point, a spatiotemporal conflict is determined. The dynamic feasibility verification module traverses the entire candidate path point sequence to perform the above checks. For path points that violate dynamic constraints, the module may make local adjustments, such as smoothing curvature or limiting the rate of change of speed. For path points that collide with static obstacles, the module attempts to search for safe alternatives nearby or fine-tune the path point positions. For path point sequence segments that have spatiotemporal conflicts with dynamic targets, the module may attempt to adjust the timestamps (i.e., speed curves) of the path points in that segment to stagger the conflict time, or adjust the spatial positions of the path points to avoid the conflict. If the adjustment fails or the conflict is severe, the candidate sequence segment may be discarded and replanning or emergency measures may be triggered. Path point sequences that pass all verifications or satisfy all constraints after adjustment are assembled into the final real-time dynamic path. The path is an ordered list, where each element is a path instruction containing precise target location coordinates (X, Y), a timestamp (T) indicating the expected arrival time at that location, and a recommended speed (V) or more specific control instruction (such as desired acceleration) for that point or segment. This sequence of timestamped path instructions is output to the autonomous vehicle's underlying motion control system, which tracks the sequence and generates specific throttle, braking, and steering control signals to drive the vehicle along the planned real-time dynamic path.

[0068] Example 5: The adaptive path adjustment step constitutes the closed-loop optimization mechanism of the entire path planning system. Its core lies in utilizing real-time dynamic path execution feedback data to continuously adjust the internal parameters of the dynamic coding model and the policy learning model to adapt to environmental changes or correct model biases. This step begins with the systematic collection of execution feedback data. During the execution of the real-time dynamic path by the autonomous vehicle, multiple dimensions of actual operational data are recorded and summarized. This data includes: the coordinate sequence of the autonomous vehicle's actual driving trajectory points and their corresponding timestamps, used for point-by-point or segment-by-segment comparison with the planned path sequence; the sequence of actual speed values ​​fed back by the underlying control system, compared with the planned speed commands; the actual minimum distance between the vehicle and the nearest obstacle (static or dynamic), monitored and recorded in real-time by the perception system; emergency event flags triggered by the safety monitoring system, such as the time, type, and intensity of events like automatic emergency braking activation and forced avoidance operation triggering; and global task progress indicators, such as the length of the traveled path, the remaining path length, and the expected arrival time deviation. This feedback data constitutes a comprehensive record of the actual execution effect of the planned path.

[0069] Based on the collected execution feedback data, the parameter offsets of the dynamic coding model and the policy learning model are calculated. This offset reflects the systematic deviation between the model's predicted or decisional output and the actual observations under the current parameter settings. The calculation process involves defining loss functions for different models and different feedback types. For the dynamic coding model, a key loss term is the environmental state prediction error. For example, the predicted components of obstacle positions or dynamic target velocities in the model's output environmental state encoding vector are compared with the corresponding values ​​actually perceived during execution, calculating the difference between the predicted and actual values ​​(such as Euclidean distance or absolute difference). Another loss term may focus on the accuracy of the predicted path sequence, i.e., comparing the spatial deviation between the sequence of predicted path points generated by the model and the sequence of actual driving trajectory points at corresponding time points. For the policy learning model, the core loss term is the deviation between the expected cumulative cost and the actual cumulative cost. The actual cumulative cost is calculated based on the execution feedback data; for example, actual driving time reflects efficiency costs, actual minimum safe distance reflects safety costs, and whether an emergency event occurs reflects risk costs. The actual cost calculated based on feedback is compared with the cumulative cost estimated by the policy model during decision-making. In addition, specific loss terms are defined for violations of safety margins (such as actual minimum safety distance falling below the threshold set during planning). These loss functions calculate their gradients with respect to their respective model parameters or directly calculate parameter adjustments. The calculations are typically performed on feedback data within a sliding time window to smooth out transient noise and obtain more stable offset estimates.

[0070] After calculating the parameter offsets, incremental parameter updates are performed on both the dynamic coding model and the policy learning model. The choice of update method depends on the model architecture, computational resource constraints, and real-time requirements. A common approach is incremental learning based on online gradient descent. For the dynamic coding model, the calculated environmental state prediction error loss gradient is used to update the model parameters with a small learning rate using gradient descent in one step, fine-tuning the model's perceptual feature extraction and state encoding capabilities. For the policy learning model, the cumulative cost bias loss gradient or policy gradient estimation is used to update the policy parameters using gradient descent, making them more inclined to produce decisions with lower actual costs. Another approach is fine-tuning, which keeps most of the model's parameters unchanged and only retrains the relevant sub-modules or specific layers on a small scale using new feedback data for specific biases identified (such as a systematic underestimation of the distance to a certain type of obstacle). A third approach is batch updates based on an experience replay buffer. Historical execution feedback data (including environmental state, adopted policies, actual results, and costs) is stored in a replay buffer. A batch of data is periodically sampled from the buffer, the loss is calculated, and the model parameters are updated. This method can utilize richer historical experience, but the update latency is relatively large. After the incremental update process is completed, the updated dynamic coding model and the updated policy learning model are output. These updated models are put into operation immediately or in the next planning cycle.

[0071] The updated dynamic coding model will be used in subsequent environmental state dynamic coding steps. This model leverages knowledge learned from new feedback, such as more accurately identifying obstacle outlines under specific lighting conditions or more precisely estimating the speed of dynamic targets in rainy or foggy weather, to generate more accurate environmental state coding vectors. The updated policy learning model will be used in subsequent path planning policy optimization steps. This model incorporates feedback on the actual effects of historical decisions. For example, in scenarios with slow-moving vehicles ahead, past aggressive overtaking strategies resulted in high costs (such as triggering emergency braking), while following strategies were less costly. The model will adjust its internal parameters to favor following strategies in similar future scenarios, thus outputting a more robust path planning strategy that better fits the constraints of the actual environment. Through this continuous adjustment of model parameters based on real-world operational data, the entire path planning system can adapt to specific operational environment characteristics (such as traffic flow patterns in specific cities or perception challenges in specific weather conditions) over long-term operation, gradually correcting its initial limitations or biases, and improving the overall performance and adaptability of the system in complex dynamic environments. This adaptive adjustment process continues throughout the entire operational lifecycle of the autonomous vehicle, forming a complete closed loop from planning, execution to feedback and learning.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic path planning method for unmanned vehicles based on multi-source sensor fusion, characterized in that, include: Real-time acquisition steps for multi-source sensor data: Based on the autonomous vehicle's driving environment, multi-source sensor data is acquired in real time, including visual image data, LiDAR point cloud data, and millimeter-wave radar reflection data; Dynamic environmental state encoding steps: Based on the multi-source sensor data, an environmental state encoding vector is generated through a dynamic encoding model; the dynamic encoding model is a Transformer based on an attention mechanism. The encoder takes multi-scale environmental features from fused sensor data processed by a feature extraction network as input and outputs a fixed-dimensional environmental state encoding vector. It captures spatial correlations and dynamic dependencies between multi-scale environmental features through a self-attention mechanism, preserves spatial location information of environmental features through position encoding, and achieves dimensionality reduction and fusion of multi-scale environmental features through multi-layer nonlinear transformation. The environmental state encoding vector includes spatial obstacle distribution features corresponding to the position, size, and density parameters of obstacles, and dynamic target motion features corresponding to the instantaneous velocity, direction of motion, and acceleration trend of dynamic targets. The predicted path generation step generates a predicted path sequence for the autonomous vehicle based on the environmental state encoding vector. The multi-dimensional path difference analysis step compares the predicted path sequence with real-time environmental data to generate a path difference coefficient tensor. The risk area probability localization step generates a risk probability distribution map through spatial environmental topology analysis based on the path difference coefficient tensor. The dynamic path cost calculation step calculates the dynamic path cost vector based on the risk probability distribution map and the environmental state encoding vector. Path planning strategy optimization steps: Based on the dynamic path cost vector, the path planning strategy is optimized through a policy learning model. The policy learning model is a reinforcement learning model based on the cumulative cost minimization objective. The input is the dynamic path cost vector, and the output is the optimized path planning strategy parameters. This model adopts the Proximal Policy Optimization (PPO) algorithm, using the weighted sum of the path safety cost component, path efficiency cost component, and path risk cost component in the dynamic path cost vector as the basis for calculating the cumulative cost. The policy network parameters are updated iteratively to satisfy the cumulative cost minimization constraint. The policy network contains three fully connected layers. The input layer dimension is consistent with the dynamic path cost vector dimension. The hidden layer uses the ReLU activation function, and the output layer uses the Softmax activation function to output the path point generation probability and the speed control parameter range. Real-time dynamic path generation steps: Based on the path planning strategy, a real-time dynamic path for the autonomous vehicle is generated; the path planning strategy includes path point generation rules and speed control parameters: the path point generation rules define the calculation logic of path point coordinates, the minimum safe distance between the path point and the obstacle in a static obstacle scenario is not less than 1.5 meters, and the lateral offset between the path point and the predicted trajectory of the dynamic target in a dynamic target scenario is not less than 0.8 meters; the speed control parameters include recommended speed ranges for different scenarios, 6-10 m / s for straight, obstacle-free scenarios, no more than 5 m / s for curved scenarios, and no more than 3 m / s for high-risk area scenarios.

2. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 1, characterized in that, The real-time data acquisition steps from the multi-source sensors specifically include: The visual image data is subjected to illumination adaptive correction processing to generate corrected visual data; the lidar point cloud data is subjected to noise filtering processing to generate filtered point cloud data; the millimeter-wave radar reflection data is subjected to Doppler effect compensation processing to generate compensated reflection data; the corrected visual data, the filtered point cloud data, and the compensated reflection data are time-aligned and fused to generate fused sensor data.

3. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 2, characterized in that, The dynamic encoding step for the environmental state specifically includes: The fused sensor data is input into a feature extraction network to extract multi-scale environmental features. This feature extraction network is a multi-branch heterogeneous network, including a visual feature branch, a point cloud feature branch, and a radar feature branch. The visual feature branch uses a 5-layer convolutional neural network (CNN), taking the corrected visual data as input. It extracts image edge texture, local contours, and global semantic features through 3×3 convolutional kernels, a stride of 1, and 2×2 max pooling, outputting a 256-dimensional visual feature vector. The point cloud feature branch uses a PointNet network structure, taking filtered point cloud data as input. It aligns point cloud coordinates using a T-Net module and extracts spatial distribution features of the point cloud through an MLP layer, outputting a 256-dimensional point cloud feature vector. The radar feature branch uses a 2-layer temporal convolutional network (TCN), taking the temporal sequence of compensated reflection data as input. It extracts multi-scale environmental features through 3×1... The convolutional kernel captures the temporal dependencies of radar data, outputting a 128-dimensional radar feature vector. This feature extraction network then merges the aforementioned visual feature vector, point cloud feature vector, and radar feature vector into a 640-dimensional multi-scale environmental feature vector through feature concatenation. The scale levels include... Meter-level detail features Meter-level local features, Meter-level global features; the multi-scale environmental features are dimensionality-reduced and encoded using a dynamic coding model to generate an environmental state coding vector; the environmental state coding vector includes spatial obstacle distribution features and dynamic target motion features.

4. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 3, characterized in that, The predicted path generation step specifically includes: Based on the environmental state encoding vector, an initial path sequence is generated through a path prediction model; the initial path sequence is then smoothed and optimized to generate a predicted path sequence; the predicted path sequence includes a sequence of path point coordinates in the time dimension.

5. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 4, characterized in that, The multidimensional path difference analysis steps specifically include: Extract the path point coordinate sequence from the predicted path sequence and the actual obstacle position sequence from the real-time environment data; calculate the spatial offset between the path point coordinate sequence and the actual obstacle position sequence; calculate the path execution delay and the matching degree of the dynamic target motion features in the time dimension; integrate the spatial offset, the path execution delay, and the matching degree of the dynamic target motion features into a normalized tensor to generate a path difference coefficient tensor.

6. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 5, characterized in that, The risk area probabilistic location step specifically includes: A spatial environment topology map of the autonomous vehicle's driving environment is constructed, the spatial environment topology map including road node connection relationships and obstacle density parameters; the path difference coefficient tensor is mapped to the corresponding node of the spatial environment topology map; the risk diffusion path is calculated through a graph structure propagation algorithm; a risk probability distribution map is generated based on the risk diffusion path, the risk probability distribution map marking the coordinates of high-risk probability areas.

7. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 6, characterized in that, The dynamic path cost calculation steps specifically include: Extract the coordinates of high-risk probability areas from the risk probability distribution map; calculate the path safety cost component based on the spatial obstacle distribution characteristics in the environmental state encoding vector; calculate the path efficiency cost component based on the dynamic target motion characteristics in the environmental state encoding vector; calculate the path risk cost component based on the coordinates of the high-risk probability areas; and weight and fuse the path safety cost component, the path efficiency cost component, and the path risk cost component to generate a dynamic path cost vector.

8. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 7, characterized in that, The path planning strategy optimization steps specifically include: The dynamic path cost vector is input into the strategy learning model; the path planning strategy is updated by the cumulative cost minimization algorithm to generate an optimized path planning strategy; the optimized path planning strategy includes path point generation rules and speed control parameters.

9. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 8, characterized in that, The real-time dynamic path generation step specifically includes: Based on the optimized path planning strategy, a sequence of candidate path points is generated; the candidate path point sequence is subjected to dynamic feasibility verification processing to generate a real-time dynamic path; the real-time dynamic path includes a path instruction sequence marked with timestamps.

10. The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion according to claim 9, characterized in that, Also includes: Adaptive path adjustment step: Based on the execution feedback of the real-time dynamic path, dynamically adjust the parameters of the dynamic encoding model and the policy learning model; The adaptive path adjustment step specifically includes: collecting the execution feedback data of the real-time dynamic path, and calculating the model parameter offset based on the execution feedback data; Incremental parameter update processing is performed on the dynamic coding model and the policy learning model to output the updated dynamic coding model and the updated policy learning model.

Images