Unmanned vehicle dynamic path planning method based on multi-source sensor fusion

Abstract

The invention relates to the technical field of unmanned vehicle path planning, and discloses an unmanned vehicle dynamic path planning method based on multi-source sensor fusion. The method comprises the following steps: an environment state dynamic coding step: generating an environment state coding vector through a dynamic coding model; a prediction path generation step of generating a prediction path sequence of the unmanned vehicle; a multi-dimensional path difference analysis step: performing multi-dimensional difference comparison on the predicted path sequence and real-time environment data to generate a path difference coefficient tensor; a risk area probability positioning step: generating a risk probability distribution diagram through space environment topology analysis based on the path difference coefficient tensor; a dynamic path cost calculation step of calculating a dynamic path cost vector in combination with the risk probability distribution diagram and the environment state coding vector; a path planning strategy optimization step: optimizing a path planning strategy through a strategy learning model; and a real-time dynamic path generation step: generating a real-time dynamic path based on the optimized strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned vehicle path planning, and in particular to a dynamic path planning method for an unmanned vehicle based on multi-source sensor fusion. Background Art

[0002] With the rapid development of intelligent transportation technology, autonomous vehicles, as a crucial component of the future transportation system, are facing a core research focus on the reliability and safety of their autonomous navigation capabilities. Dynamic path planning is one of the key technologies enabling autonomous navigation for autonomous vehicles. In complex and ever-changing driving environments, autonomous vehicles must perceive their surroundings in real time, accurately assessing key factors such as obstacle locations, pedestrian movements, and traffic light status, and accordingly plan safe and efficient routes.

[0003] Currently, autonomous vehicle path planning technology relies heavily on single-sensor data for environmental perception. While visual sensors can provide rich environmental texture information, they are prone to data distortion or feature loss in poor lighting conditions (such as at night, during heavy rain, or in fog) or when obstructions occur. LiDAR (LiDAR) can accurately obtain the three-dimensional position of obstacles using point cloud data, but its detection range is significantly affected by weather. In rainy and snowy conditions, the laser beam is easily scattered, resulting in reduced data accuracy. Millimeter-wave radar has strong penetration and is less affected by weather, but the data it generates is low-resolution, making it difficult to accurately identify small obstacles or distinguish between different types of targets. These limitations of single-sensor data make it difficult for autonomous vehicles to fully and accurately perceive their surroundings in complex environments, creating safety risks for subsequent path planning.

[0004] Traditional methods often use static or semi-static environmental modeling to process environmental states, making it difficult to capture dynamic changes in the environment in real time. For example, when a sudden traffic accident causes lane congestion, pedestrians jaywalking, or other vehicles suddenly change lanes, traditional models often fail to quickly update environmental states, resulting in a planned path that is out of sync with actual environmental requirements. Furthermore, the path sequences generated by existing path prediction methods lack effective comparison with real-time environmental data, making it difficult to quantify the differences between the path and the actual environment, resulting in a lack of clear basis for path adjustments.

[0005] When locating risk areas, existing technologies often rely on simple distance thresholds or rule-based judgments, failing to fully incorporate the spatial topology of the environment. This results in insufficient accuracy in locating risk areas, making them prone to missed or misjudgment. For example, misidentifying non-risk areas as risky, leading to excessive route detours, or failing to promptly identify potential risk areas, potentially causing accidents. Furthermore, path cost calculations often only consider single factors such as distance and time, ignoring multidimensional dynamic factors such as risk probability, road roughness, and traffic density. This results in an incomplete cost assessment and compromises the rationality of path planning.

[0006] At the level of strategy optimization, traditional path planning strategies mostly rely on preset rules or static algorithms, lacking adaptive learning ability for dynamic environment. When unknown changes occur in the driving environment, the strategy cannot be quickly adjusted, resulting in insufficient flexibility and adaptability of path planning, which is difficult to meet the driving needs of unmanned vehicles in complex urban roads, rural roads and other diversified scenarios. SUMMARY

[0007] The purpose of the present application is to provide an unmanned vehicle dynamic path planning method based on multi-source sensor fusion to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides an unmanned vehicle dynamic path planning method based on multi-source sensor fusion, which comprises: Multi-source sensor data real-time acquisition step: based on the driving environment of the unmanned vehicle, real-time acquisition of multi-source sensor data, the multi-source sensor data including visual image data, laser radar point cloud data and millimeter wave radar reflection data; environment state dynamic coding step: based on the multi-source sensor data, generating an environment state coding vector through a dynamic coding model; predicted path generation step: based on the environment state coding vector, generating a predicted path sequence of the unmanned vehicle; multi-dimensional path difference analysis step: comparing the predicted path sequence with real-time environment data in multi-dimensional difference, generating a path difference coefficient tensor; risk area probability positioning step: based on the path difference coefficient tensor, generating a risk probability distribution map through spatial environment topology analysis; dynamic path cost calculation step: based on the risk probability distribution map and the environment state coding vector, calculating a dynamic path cost vector; path planning strategy optimization step: based on the dynamic path cost vector, optimizing the path planning strategy through a strategy learning model; real-time dynamic path generation step: based on the path planning strategy, generating a real-time dynamic path of the unmanned vehicle.

[0009] Preferably, the multi-source sensor data real-time acquisition step specifically comprises: The visual image data is subjected to illumination adaptive correction processing to generate corrected visual data; the laser radar 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 subjected to time sequence alignment fusion to generate fused sensor data.

[0010] Preferably, the environment state dynamic coding step specifically comprises: The fused sensor data is input into a feature extraction network to extract multi-scale environmental features; the multi-scale environmental features are subjected to dimensionality reduction coding processing through 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.

[0011] Preferably, the predicted path generating 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 smoothed and optimized to generate a predicted path sequence; the predicted path sequence includes a path point coordinate sequence in a time dimension.

[0012] Preferably, the multidimensional path difference analysis step specifically includes: Extracting a path point coordinate sequence from the predicted path sequence and an actual obstacle position sequence from the real-time environment data; calculating a spatial offset between the path point coordinate sequence and the actual obstacle position sequence; calculating a matching degree between a path execution delay amount and a dynamic target motion feature in a time dimension; and integrating the spatial offset, the path execution delay amount, and the matching degree of the dynamic target motion feature into a normalized tensor to generate a path difference coefficient tensor.

[0013] Preferably, the risk area probabilistic positioning step specifically includes: Construct a spatial environment topology map of the unmanned vehicle's driving environment, wherein the spatial environment topology map includes road node connection relationships and obstacle density parameters; map the path difference coefficient tensor to the corresponding nodes of the spatial environment topology map; calculate the risk diffusion path through a graph structure propagation algorithm; and generate a risk probability distribution map based on the risk diffusion path, wherein the risk probability distribution map marks the coordinates of high-risk probability areas.

[0014] Preferably, the dynamic path cost calculation step specifically includes: Extract the coordinates of the high-risk probability area in the risk probability distribution map; calculate the path safety cost component based on the spatial obstacle distribution characteristics in the environmental state coding vector; calculate the path efficiency cost component based on the dynamic target motion characteristics in the environmental state coding vector; calculate the path risk cost component based on the coordinates of the high-risk probability area; and perform weighted fusion of the path safety cost component, the path efficiency cost component, and the path risk cost component to generate a dynamic path cost vector.

[0015] Preferably, the path planning strategy optimization step specifically includes: input the dynamic path cost vector into the strategy learning model; update the path planning strategy through a cumulative cost minimization algorithm to generate an optimized path planning strategy; the optimized path planning strategy comprises a path point generation rule and a speed control parameter.

[0016] Preferably, the real-time dynamic path generation step specifically comprises: based on the optimized path planning strategy, generate a candidate path point sequence; perform dynamic feasibility verification processing on the candidate path point sequence to generate a real-time dynamic path; the real-time dynamic path comprises a time-stamped path instruction sequence.

[0017] Preferably, the method further comprises: an 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 strategy learning model; the adaptive path adjustment step specifically comprises: collect execution feedback data of the real-time dynamic path, calculate model parameter offsets based on the execution feedback data; perform incremental parameter update processing on the dynamic encoding model and the strategy learning model, and output an updated dynamic encoding model and an updated strategy learning model.

[0018] Compared with the prior art, the present application has the following advantages: The dynamic path planning method for unmanned vehicles based on multi-source sensor fusion effectively makes up for 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, laser radar point cloud data and millimeter wave radar reflection data enable the unmanned vehicle to obtain environmental information from multiple dimensions. The texture features provided by the visual data, the three-dimensional position accuracy of the laser radar and the anti-interference ability of the millimeter wave radar complement each other, and in the case of light changes, weather interference or the presence of obstructions, the stable perception of key environmental elements such as obstacles, pedestrians and road boundaries can still be maintained, improving the comprehensiveness and reliability of environmental perception. The environmental state dynamic encoding step can convert complex environmental information into structured environmental state encoding vectors through deep processing of multi-source sensor data by the dynamic encoding model. This encoding method not only retains static features in the environment (such as road shape and fixed obstacle position), but also captures dynamic features (such as pedestrian movement speed and vehicle driving direction) in real time, making the unmanned vehicle's understanding of the environmental state more accurate and timely, and providing high-quality basic data for subsequent path planning. The combination of the predicted path generation and the multi-dimensional path difference analysis realizes the quantitative evaluation of the deviation between the predicted path and the actual environment. The path difference coefficient tensor can comprehensively reflect the difference between the path and the real-time environment data in multiple dimensions such as spatial position, driving speed and obstacle avoidance distance, avoiding the subjectivity and ambiguity of the traditional method in judging the rationality of the path. Through this quantitative analysis, the unmanned vehicle can clearly identify the unreasonable part in the predicted path, providing a clear direction for path adjustment. The risk area probability positioning step generates a risk probability distribution map based on the path difference coefficient tensor and spatial environment topology analysis, fully considering the spatial correlation of various elements in the environment. For example, combined with the topological structure of road intersections, the habitual path of pedestrians crossing the street and other information, it can more accurately locate high-risk areas (such as blind spots and pedestrian dense areas) rather than simply relying on distance threshold for judgment. This makes the identification of risks by the unmanned vehicle more objective and comprehensive, providing a scientific basis for path safety evaluation. The dynamic path cost calculation combines the risk probability distribution map with the environment state encoding vector, considering multiple factors such as path length, driving time, risk probability and road traffic conditions. The generated dynamic path cost vector can comprehensively reflect the comprehensive cost of different paths, avoiding the one-sidedness of path planning in traditional cost calculation due to considering only a single factor. Based on this comprehensive cost evaluation, the unmanned vehicle can achieve a better balance between safety and efficiency when selecting a path. The path planning strategy optimization step learns and optimizes the dynamic path cost vector through a strategy learning model, enabling the path planning strategy to adapt to dynamic environments. The strategy learning model can continuously adjust decision parameters from historical data and real-time feedback, quickly generating a planning strategy that adapts to the current environment when encountering new environmental changes (such as sudden obstacles or road construction), rather than relying on fixed rules. This dynamic optimization capability significantly improves the flexibility and adaptability of path planning, enabling the unmanned vehicle to maintain good navigation performance in complex and changing driving scenarios. The real-time dynamic path generation step generates a real-time path that meets the current environmental requirements based on the optimized path planning strategy, quickly responding to environmental changes. Whether it is traffic flow changes on urban roads or sudden obstacles on rural roads, the unmanned vehicle can adjust the driving path in a timely manner to ensure safety and efficiency during driving, improving the reliability and practicality of the unmanned vehicle in actual applications. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The working principle diagram of the unmanned vehicle dynamic path planning method based on multi-source sensor fusion described in the present application; Figure 2 The working principle diagram of the dynamic encoding of the environment state; Figure 3 Schematic diagram of the working principle for the predicted path generation; Figure 4 Diagram showing the working principle of probabilistic positioning of risk areas. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] See also Figure 1 The present 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 sensor data, combined with dynamic modeling of environmental states and risk probability analysis. The specific implementation process is as follows: During driving, the autonomous vehicle simultaneously collects visual image data, LiDAR point cloud data, and millimeter-wave radar reflection data, forming multi-source sensor data input. After preprocessing, this raw data is fed into a dynamic encoding model. This model extracts and compresses key environmental information, outputting an environmental state encoding vector that contains spatial obstacle distribution characteristics and dynamic target motion characteristics. Based on this encoding vector, the path prediction module generates a predicted path sequence for a future time window.

[0022] This predicted path sequence is compared with real-time environmental data in multiple dimensions for analysis. Difference metrics such as spatial offset, time delay, and motion feature matching are calculated and integrated to form a path difference coefficient tensor. This tensor, combined with a pre-constructed spatial environmental topology map, is used via a graph structure propagation algorithm 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, which calculates the path safety cost component, the path efficiency cost component, and the path risk cost component. A weighted fusion of these components yields a dynamic path cost vector. This vector drives the policy learning model to optimize, update the path point generation rules and speed control parameters, and ultimately form an optimized path planning strategy.

[0023] Based on this strategy, a sequence of candidate path points is generated. After dynamic feasibility verification, a time-stamped sequence of path instructions is output, which serves as the basis for the autonomous vehicle's real-time dynamic path execution. The entire process forms a closed loop, and the execution feedback data can be used to dynamically adjust the model parameters.

[0024] Example 1: See Figure 2The implementation of the real-time multi-source sensor data acquisition process begins with a parallel processing flow for the raw sensory data. Visual image data is fed into the adaptive illumination 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, nonlinear compression methods are used to restore details. This processing considers the overall lighting conditions of the scene to avoid introducing artifacts, and outputs corrected visual data with balanced color and clear details. LiDAR point cloud data processing focuses on noise suppression and data reduction. The point cloud data is first fed into 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 across all points, a threshold of multiples of the standard deviation is set to remove discrete noise points whose distances significantly deviate from the group mean. Subsequently, a spatial gridding technique is optionally applied to divide the three-dimensional space into uniform voxel units. Within each voxel, the geometric center coordinates of all points are calculated and used to replace all original points within the cell. This downsampling of the point cloud data significantly reduces the data volume while preserving the outline of the environmental structure, and outputs filtered point cloud data. The core of millimeter-wave radar reflection data processing lies in accurately correcting velocity information. The Doppler effect compensation module receives millimeter-wave radar reflection data and the real-time velocity vector of the unmanned vehicle. The module calculates the angle between the radar beam direction and the unmanned vehicle's velocity vector. Based on the principle of Doppler shift, this angle and the vehicle's velocity are combined to calculate the target's radial velocity measurement error caused by the vehicle's own motion. The radial velocity value of each reflection point detected by the radar is inversely compensated to eliminate measurement errors introduced by the vehicle's motion, outputting compensated reflection data that more accurately reflects the target's true relative motion.

[0025] After completing the above independent preprocessing, the timing alignment and fusion phase begins. This phase aims to unify data streams from different sensors, which may have slight timestamp discrepancies, to a common time base. Alignment is achieved using either hardware synchronization or software interpolation. If the sensor supports hardware timestamp synchronization, a precise global clock signal is used to timestamp each data packet, and data with the same timestamp is directly selected for fusion. If hardware synchronization is unavailable or inaccurate, software interpolation is used. Based on the acquisition timestamps of each sensor data stream, linear or spline interpolation is performed on adjacent time points to generate an estimate of the sensor data at the specified target time point. The aligned rectified visual data, filtered point cloud data, and compensated reflectance data are input into the fusion module. Fusion strategies can be either early fusion or mid-term fusion. For example, early fusion projects the filtered point cloud data onto the image plane corresponding to the rectified visual data to generate a pixel-level depth map, which is then concatenated with the original image data to form RGB-D data. Mid-term fusion extracts features from each of the three data sources. For example, edge detection or preliminary convolution operations are applied to the corrected visual data to extract image features; distance- or normal-based feature extraction methods are applied to the filtered point cloud data; and intensity- or Doppler-based feature extraction methods are applied to the compensated reflection data. These heterogeneous features are then concatenated or joined across the feature dimensions to form fused sensor data. This data is a multidimensional feature set that integrates visual appearance, spatial geometry, and motion velocity information.

[0026] The dynamic encoding 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 process different types of data. One branch processes visual data (such as RGB or RGB-D images). This branch comprises a multi-layer convolutional neural network (CNN) that, through convolution and pooling operations, progressively extracts image features from local to global scales, such as edge texture information, object contours, and semantic segmentation results. Another branch processes point cloud data. This branch may employ a point cloud network architecture based on point operations to directly process 3D point coordinates and their attributes, extracting spatial distribution features of the point cloud, such as the shape, size, and distance of obstacles, as well as the surface geometry of the scene. The third branch processes millimeter-wave radar data. This branch may comprise a recurrent neural network or a temporal convolutional layer, processing the sequential data of reflection points to extract target trajectory features, velocity distribution patterns, and stability information of dynamic targets. Each branch operates independently, outputting its own extracted multi-scale environmental features. These features may contain information at different spatial resolutions or temporal granularities.

[0027] The extracted multi-scale environmental features are then fed into a dynamic encoding model. This model's core task is to fuse heterogeneous features from different sensors and capture the dynamic changes and dependencies of environmental states across time and space. The model architecture 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 and then fed into the model. The model uses a self-attention mechanism to calculate the association weights between different feature elements, highlighting important features and suppressing minor ones, while also preserving spatial information through positional encoding. If a graph neural network is used, environmental features are represented as node features in a graph structure. The connections (edges) between nodes can be defined based on spatial proximity or semantic relevance. The graph neural network uses a message passing mechanism to aggregate information from neighboring nodes to update the feature representation of each node. Regardless of the architecture, the dynamic encoding 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 encoding vector of the environmental state. This vector is a compact representation of the environmental state, with a dimensionality significantly lower than that of the original input features. This vector structurally encodes key information about the environment, including components that characterize the spatial distribution of obstacles, such as their location, size, and density. It also includes components that characterize the motion of dynamic targets, such as their instantaneous velocity, direction, acceleration trend, and trajectory prediction. This environmental state encoding vector serves as the core input for subsequent path prediction and analysis modules.

[0028] Example 2: See Figure 3 The implementation of the predicted path generation step uses the environment state encoding vector as the core input. This vector is input into the path prediction model, which is responsible for generating driving trajectory predictions for a period of time in the future. The path prediction model usually adopts a sequence generation architecture, such as a recurrent neural network based on a gated recurrent unit or a Transformer model based on a self-attention mechanism. The model receives a sequence of current and historical environment state encoding vectors to capture the temporal evolution of the environment state. The internal structure of the model includes a hidden state transfer mechanism that can memorize historical information and infer future trends. The model output is a sequence of predicted path point coordinates at future consecutive time steps, which constitutes the initial path sequence. Each element in the sequence contains a three-dimensional spatial coordinate value and a corresponding timestamp, indicating the location where the unmanned vehicle is expected to arrive at a specific moment. The initial path sequence may have a discrete point distribution or local non-smoothness.

[0029] The initial path sequence is then input into a smoothing optimization module for processing. The goal of this module is to generate a continuous trajectory that is compliant with vehicle dynamics constraints and is comfortable to drive. The smoothing optimization process can employ either a parametric curve fitting method or a numerical optimization method. If a parametric curve fitting method is employed, such as a B-spline curve or a Bezier curve, the discrete initial path points are taken as control points and a smooth continuous curve is fitted. The choice of curve parameters needs to satisfy the minimum turning radius constraint of the ego vehicle to avoid generating sharp turns that exceed the physical limits of the vehicle. If a numerical optimization method is employed, such as a gradient descent or quadratic programming based optimizer, an objective function is defined. The objective function typically includes a smoothness term (such as the sum of squares of path curvature or the rate of change of direction between path points) and a fitting term to the initial path (such as the sum of squares of distances between path points and the initial predicted points). During the optimization process, constraints are applied simultaneously, including the maximum allowed curvature of path points, the minimum and maximum distances between adjacent path points (reflecting the speed range), and the minimum safety distance between path points and known static obstacles. The optimization solver finds the path point sequence that minimizes the objective function value while satisfying all the constraints. After the smoothing optimization process, the final predicted path sequence is output. This 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.

[0030] The implementation of the multi-dimensional path difference analysis step requires the simultaneous acquisition of the predicted path sequence and real-time environment data. Real-time environment data refers to the latest environment perception information collected in real time by multiple sources of sensors and pre-processed and fused within the time window corresponding to the predicted path sequence. The analysis process first performs data extraction. From the predicted path sequence, the path point coordinate sequence it contains is extracted, and each coordinate point is associated with its predicted timestamp. From the real-time environment data, the actual existing obstacle position sequence is extracted through target detection and tracking algorithms (such as Kalman filtering or multi-target tracking algorithms based on fused perception). This sequence contains the position coordinates of each tracked obstacle at different timestamps and its identification information.

[0031] The calculation process revolves around three core dimensions of differentiation. The first dimension is the calculation of spatial offset. For each pathpoint in the predicted path sequence, its corresponding timestamp is determined. At this timestamp, all existing obstacle locations are searched from the actual obstacle location sequence. The Euclidean distance between the pathpoint's coordinates and the coordinates of each actual obstacle location is calculated. The minimum of these distances is taken as the spatial offset of the pathpoint at the current moment. All pathpoints in the predicted path sequence are traversed, and the spatial offset corresponding to each point is calculated, forming a sequence of spatial offsets. The second dimension is the calculation of path execution delay. This delay reflects the time lag introduced by planning system processing, control command issuance, and vehicle response. The planned arrival timestamp (i.e., predicted time) of each pathpoint in the predicted path sequence is recorded. In actual operation, the actual timestamp of the unmanned vehicle's arrival near the pathpoint (e.g., within the set tolerance range) is recorded. The difference between the actual timestamp and the planned timestamp is calculated to represent the path execution delay for that pathpoint. Similarly, the sequence is traversed to obtain a sequence of path execution delays. The third dimension is the calculation of the matching degree of dynamic target motion features. The generation of a predicted path sequence implicitly anticipates the future motion states of dynamic objects in the environment (such as other vehicles and pedestrians). This anticipated information is derived from the dynamic object motion features encoded in the environmental state vector. Real-time environmental data also includes dynamic object motion features based on the latest perception updates, such as instantaneous velocity vector, acceleration, and motion direction. For key dynamic objects (such as those adjacent to or potentially impacting the vehicle), the predicted motion features (such as predicted velocity and direction) are compared with the actual perceived motion features (such as actual velocity and direction) at the corresponding time point. The degree of matching can be calculated using vector similarity metrics, such as the cosine similarity (dot product divided by the product of the modulo) between the predicted and actual velocity and direction vectors, with values ​​closer to 1 indicating a higher degree of match. Alternatively, the Mahalanobis distance between the predicted and actual positions (accounting for motion uncertainty) can be calculated. Similarly, a matching degree sequence reflecting the accuracy of the motion state prediction is generated.

[0032] After obtaining the spatial offset sequence, path execution delay sequence, and dynamic target motion feature matching degree sequence, normalization processing is performed before data integration. The purpose of normalization is to eliminate the differences in dimension and numerical range of different dimensional data, making them comparable and facilitating subsequent processing. For the spatial offset sequence, a reasonable maximum offset expected value is set according to the sensor accuracy and the scene, and all offsets are divided by the maximum value for normalization, so that 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 divided by the threshold for normalization. For the dynamic target motion feature matching degree sequence, if cosine similarity is used (range [-1, 1]), it can be mapped to the [0, 1] interval (such as (similarity value + 1) / 2); if distance measurement is used, a maximum expected distance is also set for normalization. After normalization, the three 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. They are combined into a three-dimensional data structure, i.e., a path difference coefficient tensor. The first dimension of the tensor corresponds to the spatial offset, the second dimension corresponds to the path execution delay, and the third dimension corresponds 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 multi-dimensional inconsistency between the predicted path and the actual environmental perception results, providing input for subsequent risk assessment.

[0033] Embodiment 3: Refer to Figure 4 The risk area probability positioning step is implemented with the path difference coefficient tensor as the core input. This step first constructs a spatial environment topology graph that describes the driving environment of the unmanned vehicle. The topology graph uses a graph data structure, where nodes represent key location points or center points of divided regions in the environment, such as road intersection centers, lane center points, or grid center points of a grid map. Edges are used to connect nodes, representing spatial accessibility or road connection relationships between nodes, such as connecting adjacent lane nodes or connecting nodes in the same lane. Each node is associated with a set of attribute data, including its spatial coordinate position (such as geodetic coordinates or coordinates in a local coordinate system), obstacle density parameters based on historical statistics or real-time perception calculations (such as the frequency or average number of obstacles per unit area), and optional additional attributes such as road type (main road, ramp, etc.), speed limit, etc. The construction of the topology graph can be based on high-precision map information, or dynamically generated and updated through real-time perception data when there is no prior map.

[0034] The path difference coefficient tensor needs to be mapped to the corresponding node of the spatial environment topology map. The mapping process is performed based on the spatial location where the path difference occurs. For each element in the path difference coefficient tensor (corresponding to a specific path point or time step), extract the spatial position coordinates associated with the element (i.e., the predicted path point coordinates or the location of the difference calculation point). In the spatial environment topology map, find the topology map node with the closest Euclidean distance to the spatial position coordinates. Once the nearest neighbor node is found, the three normalized difference values ​​(spatial offset, path execution delay, dynamic target motion feature matching) contained in the path difference coefficient tensor element are associated with the topology map node as a set of additional attributes. If multiple path difference points are mapped to the same topology node, the comprehensive difference coefficient of the node can be calculated by taking the average, maximum value or other aggregation functions.

[0035] Based on the spatial environment topology of the mapped path difference coefficient, the graph structure propagation algorithm is used to calculate the risk diffusion path in the environment and the final risk probability distribution. This algorithm simulates the process of risk spreading from high-difference areas to surrounding areas. A feasible implementation is an iterative message passing algorithm. At initialization, the initial risk value of each topological node is It is determined by the associated comprehensive path difference coefficient, for example, the weighted sum of the three difference coefficients. In each node The risk value of the node is based on its neighbor nodes The risk value of the last iteration ,node and The distance between (or the weight of the connecting edge, the weight can be designed as a decreasing function of the distance, such as ), and nodes Its own obstacle density parameter To update. A simplified update rule example is:

[0036] in: Representation node In the The risk value after iterations; Representation node The set of all neighbor nodes of Representation node With node The Euclidean distance between Representation node Obstacle density parameter; and is a weight coefficient between 0 and 1, which controls the influence of neighbor risk propagation and local obstacle density on the current risk, and satisfies ; is a node The risk value of the previous iteration is introduced to retain some historical risk information. This formula reflects the characteristics of risk propagation: high-risk neighboring nodes in close proximity will significantly increase the risk of the current node; high local obstacle density also directly increases the risk; and risk has a certain degree of persistence. The algorithm iterates until the risk value change of all nodes is less than the set threshold or the maximum number of iterations is reached, and finally a stable risk probability value is obtained 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 ​​in the form of node color depth or numerical annotation on the topological map, or it can be converted into a grid map format, dividing the environment into grids. The risk value of each grid is obtained by interpolating or averaging the risk values ​​of the nodes it contains. The coordinates of the areas where the risk probability exceeds the set threshold are clearly marked in the map, identifying them as high-risk probability areas.

[0037] The dynamic path cost calculation step is based on the risk probability distribution map and the environment state encoding vector. First, the coordinates of all high-risk probability areas and their corresponding risk probability values ​​are extracted from the risk probability distribution map. This information quantifies the likelihood that different locations in the environment pose a threat to the autonomous vehicle. The cost calculation involves the calculation of three independent components. The path safety cost component relies on the spatial obstacle distribution characteristics in the environment state encoding vector. This component assesses the safety margin between the planned path and all known obstacles in the environment, including the predicted trajectories of static and dynamic obstacles. The calculation process iterates over each point on the planned path and calculates the Euclidean distance from that point to the nearest obstacle (or obstacle trajectory). When this distance falls below a preset safety distance threshold, the cost component increases significantly, with the cost increment increasing with closer distances. The path geometry is also considered, and the local curvature of the path is calculated. Excessive curvature indicates sharp turns, which may increase the risk of loss of control or passenger discomfort. Therefore, path points with high curvature also increase the safety cost component. The calculation of the path efficiency cost component relies on the dynamic object motion characteristics in the environment state encoding vector and global path information. This component focuses on the efficiency of the path. Calculate the total length of the path. A longer path generally means a longer travel time. Estimate the travel time of the path, which requires combining the planned speed profile (which may be given by the policy or estimated based on the path curvature and speed limit). Evaluate the consistency of the path direction with the global target direction (such as the straight line from the start point to the end point) and calculate the deviation angle between the path point direction and the target direction. The larger the deviation, the higher the efficiency cost component. In addition, if the dynamic target motion characteristics indicate that there are high-speed moving objects that may affect traffic efficiency (such as a slow vehicle ahead), if the path point is in an area that may be blocked, the efficiency cost will also increase. The calculation of the path risk cost component is directly based on the extracted high-risk probability area coordinates and their probability values. This component evaluates the tendency of the planned path to approach or cross high-risk areas. Calculate the distance from each point on the path to the nearest high-risk area boundary. When the distance is less than the 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 the path segment directly crosses a high-risk area, the risk cost component is significantly increased based on the length of the crossed area and the average risk probability value of the area. The calculated path safety cost component (denoted as ), path efficiency cost component (denoted as ) and the path risk cost component (denoted as ) for fusion. The fusion adopts the weighted summation method:

[0038] in: It is the final dynamic path cost vector, which may contain cost values ​​for different segments or different evaluation points on the path; 、 , are the weight coefficients of the path safety cost component, the efficiency cost component, and the risk cost component, respectively. These weight coefficients are preset or configurable parameters, whose values reflect different priority trade-offs between safety, driving efficiency, and risk avoidance in path planning. For example, in a complex urban environment, the safety weight and the risk weight may be set higher; while in a highway scenario, the efficiency weight may be relatively higher. The dynamic path cost vector obtained after weighted fusion provides a quantitative basis for subsequent path planning strategy optimization.

[0039] Embodiment 4: The path planning strategy optimization step takes the dynamic path cost vector as the core driving input. This vector quantifies the comprehensive cost of different path choices or path points in different dimensions (safety, efficiency, risk). This vector input is fed into a strategy learning model. The model is usually 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 goal of the model is to output a path planning strategy that defines how to generate a specific driving path under the current environmental state and cost constraints. The model is trained or optimized online through a cumulative cost minimization algorithm. This algorithm evaluates the expected cumulative cost of the candidate strategy proposed by the model over multiple time steps in the future or until the task goal point. The calculation of the expected cumulative cost takes into account the discount factor, i.e., the influence of future costs on current decisions gradually diminishes. The algorithm adjusts the internal parameters of the strategy learning model using optimization methods such as stochastic gradient descent, evolutionary strategies, or sampling-based optimization, aiming to find a set of parameters that minimize the expected cumulative cost corresponding to the strategy output by the model. After parameter update, 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 logic for generating path point sequences, such as: when a static obstacle is perceived on the left side, the generated path point sequence should be offset to the right by a certain distance; when a dynamic target is detected approaching in front, the path point sequence should plan an avoidance trajectory in advance, possibly including a combination of lateral offset and longitudinal deceleration points; when approaching a turn at an intersection, the path point sequence should follow specific turn radius and angle of entry rules. The speed control parameters define the recommended speed or acceleration curve to be adopted in different sections of the path, such as: on a curve section, the speed parameter should be reduced to below a set value; when a path point approaches a high-risk probability area, the speed parameter should be set to a cautious passing speed; on a straight section with no obstacles, the speed parameter can be increased to an efficiency-first speed. These rules and parameters together constitute the optimized driving decisions.

[0040] The real-time dynamic path generation step is 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. Based on the pathpoint generation rules, the module generates a set of candidate pathpoint sequences. For example, in a specific scenario: the current environmental state encoding vector indicates a stationary obstacle occupying part of the lane 100 meters ahead, a vehicle in the adjacent lane on the left traveling in the same direction at a high speed, and a curb on the right. Based on the pathpoint generation rule that states "When encountering an obstacle in the same direction and a vehicle on the left, generate a pathpoint sequence for slightly maneuvering right and preparing to use another lane to avoid it," the module might generate the following sequence: the first 20 pathpoints (corresponding to the next 2 seconds) involve slightly maneuvering right within the current lane, maintaining a safe distance from the right curb. The next 20 pathpoints (corresponding to another 2 seconds) generate a pathpoint sequence for cutting left into the centerline of the adjacent lane after confirming that the vehicle on the left has passed. The final 20 pathpoints (corresponding to another 2 seconds) generate a pathpoint sequence for returning to the centerline of the original lane. Meanwhile, speed control parameters are set according to a set of rules: maintain the current speed during the initial fine-tuning phase; slightly decelerate when entering an adjacent lane; and resume the original speed after returning to the original lane. Table 1 shows some candidate waypoint sequences and their associated speed control parameters for this scenario.

[0041] Table 1: Candidate path point sequence and speed control parameters Waypoint sequence index Waypoint 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 to avoid obstacles 10 (100.0,4.0) 2.0 8.0 Approaching an obstacle, keep a safe distance to the right 15 (150.0,3.0) 3.0 7.5 Cut into the adjacent lane on the left and start slowing down 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 Prepare to return to the original lane 30 (300.0,3.5) 6.0 8.0 Return to the center of the original lane and resume speed After the candidate pathpoint 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 pathpoints is calculated and divided by the time difference to obtain the required average speed. This speed is then checked to see if it is within the permitted speed range for the autonomous vehicle. The local curvature of the pathpoint sequence (estimated by the angle formed by three consecutive points) is calculated to see if it is less than the maximum permissible curvature for the autonomous vehicle, preventing sharp turns that exceed the vehicle's physical limits. The rate of change of direction between adjacent pathpoints is also checked to see if it is less than the maximum permissible value. Next, static obstacle collision detection is performed: the trajectory formed by the pathpoint sequence (usually represented by line segments or simple geometric shapes such as capsules) is checked for geometric intersection with the static obstacle map of the environment constructed based on real-time perception. If an intersection is detected, the pathpoint or segment is deemed to pose a collision risk. Finally, dynamic obstacle spatiotemporal conflict detection is performed: the predicted trajectories of dynamic objects (such as vehicles and pedestrians) in the environment are obtained (usually a time-stamped pathpoint sequence). The candidate pathpoint sequence of the vehicle is compared with the predicted trajectory sequence of each dynamic object in the same time coordinate system. At each timestamp, the area occupied by the vehicle (taking into account the vehicle's outline) is calculated to determine whether it overlaps with the area occupied by the dynamic target. If overlap is detected at a future time point, a spatiotemporal conflict is determined. The dynamic feasibility verification module performs these checks on the entire candidate waypoint sequence. For waypoints that violate dynamic constraints, the module may perform local adjustments, such as smoothing curvature or limiting the rate of velocity change. For waypoints that collide with static obstacles, the module attempts to search for safe alternatives nearby or fine-tune the waypoint position. For waypoint sequence segments that experience spatiotemporal conflicts with dynamic targets, the module may attempt to adjust the timestamps (i.e., velocity profiles) of the waypoints in that segment to offset the conflict or adjust the spatial position of the waypoints to avoid them. If these adjustments fail or the conflict is severe, the candidate segment may be discarded, triggering replanning or emergency measures. Waypoint sequences that pass all verifications or, after adjustments, satisfy all constraints are assembled into the final real-time dynamic path. The path is an ordered list, where each element is a path instruction, including the precise target location coordinates (X, Y), the timestamp (T) of the expected arrival time at that location, and the recommended speed (V) or more specific control instructions (such as the expected acceleration) at that point or section. This timestamped path instruction sequence is output to the autonomous vehicle's underlying motion control system, which is responsible for tracking the sequence and generating specific throttle, brake, and steering control signals to drive the vehicle along the planned real-time dynamic path.

[0042] Example 5: The adaptive path adjustment step constitutes the closed-loop optimization mechanism of the entire path planning system. Its core is to utilize execution feedback data from real-time dynamic paths to continuously adjust the internal parameters of the dynamic encoding model and policy learning model to adapt to environmental changes or correct model deviations. This step begins with the systematic collection of execution feedback data. As the unmanned vehicle executes the real-time dynamic path, multiple dimensions of actual operational data are recorded and aggregated. This data includes: the coordinate sequence of the unmanned vehicle's actual driving trajectory points and their corresponding timestamps, which are 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, which are compared with the planned speed instructions; 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 indicators triggered by the safety monitoring system, such as the time, type, and intensity of events such as automatic emergency braking activation and forced avoidance maneuvers; and global task progress indicators, such as the length of the path traveled, the remaining path length, and the estimated time of arrival deviation. This feedback data constitutes a comprehensive record of the actual execution of the planned path.

[0043] Based on the collected execution feedback data, the parameter offsets of the dynamic encoding model and the policy learning model are calculated. This offset reflects the systematic deviation between the model's predictions or decision outputs and actual observations under the current parameter settings. The calculation process involves defining loss functions tailored to different models and different feedback types. For the dynamic encoding model, a key loss term is the error in the environment state prediction. For example, the predicted component of the environment state encoding vector output by the model, such as the obstacle position or dynamic target velocity, is compared with the corresponding value actually perceived during execution, and the difference between the predicted and actual values ​​is calculated (such as the Euclidean distance or absolute difference). Another loss term may focus on the accuracy of the predicted path sequence, comparing the spatial deviation between the predicted path point sequence generated by the model and the actual driving trajectory point sequence 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 execution feedback data. For example, the actual driving time reflects the efficiency cost, the actual minimum safe distance reflects the safety cost, and the occurrence of an emergency reflects the risk cost. The actual cost calculated based on the feedback is compared with the cumulative cost estimated by the policy model during decision making. In addition, specialized loss terms are defined for safety margin violations (e.g., when the actual minimum safety distance falls below a threshold set during planning). These loss functions compute the gradients of their respective model parameters or directly calculate parameter adjustments. The calculation is typically performed on feedback data within a sliding time window to smooth transient noise and obtain more stable offset estimates.

[0044] After parameter offset calculation is complete, incremental parameter updates are performed on the dynamic encoding model and 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 encoding model, the calculated environment state prediction error loss gradient is used to update the model parameters using a one-step gradient descent update with a small learning rate, fine-tuning the model's perceptual feature extraction and state encoding capabilities. For the policy learning model, the accumulated cost-bias loss gradient or policy gradient estimate is used to update the policy parameters using gradient descent, similarly favoring decisions with lower actual costs. Another approach is fine-tuning, which leaves most model parameters unchanged and only uses new feedback data to perform small-scale retraining of the relevant submodules or specific layers for identified deviations (such as systematic underestimation of the distance to a certain type of obstacle). A third approach is batch updates based on the experience replay buffer. Historical execution feedback data (including environment state, adopted policy, actual results, and cost) is stored in a replay buffer. A batch of data is periodically sampled from the buffer to calculate the loss and update the model parameters. This approach can leverage a richer historical experience, but the update latency is relatively large. After the incremental update process is completed, the updated dynamic encoding model and the updated policy learning model are output. These updated models are put into operation immediately or in the next planning cycle.

[0045] The updated dynamic encoding model is used in the subsequent dynamic encoding of the environment state. The 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 a dynamic target in rain and fog, to generate a more accurate encoding vector for the environment state. The updated policy learning model is used in the subsequent path planning strategy optimization step. This model incorporates feedback on the actual effectiveness of historical decisions. For example, in a scenario like a slow-moving vehicle ahead, if past aggressive overtaking strategies resulted in high costs (such as triggering emergency braking), while following strategies are less costly, the model will adjust its internal parameters to favor following strategies in similar scenarios in the future, thereby outputting a more robust path planning strategy that better meets the actual environmental constraints. Through this continuous, data-driven model parameter adjustment, the entire path planning system can continuously adapt to the specific operating environment characteristics (such as traffic flow patterns in a specific city or perception challenges in specific weather conditions) over the long term, gradually correcting the model's initial limitations or biases, and improving the system's overall performance and adaptability in complex dynamic environments. This adaptive adjustment process continues throughout the autonomous vehicle's operational lifecycle, forming a complete closed loop from planning and execution to feedback and learning.

[0046] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0047] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, 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 multi-source sensor data acquisition step: Based on the unmanned vehicle driving environment, multi-source sensor data is collected in real time, and the multi-source sensor data includes visual image data, lidar point cloud data and millimeter wave radar reflection data; environmental state dynamic encoding step: Based on the multi-source sensor data, the environmental state encoding vector is generated through a dynamic encoding model; the dynamic encoding model is a Transformer based on the attention mechanism The encoder input is the multi-scale environmental features output by the feature extraction network of the fused sensor data, and the output is a fixed-dimensional environmental state encoding vector. The encoder captures the spatial correlation and dynamic dependency between the multi-scale environmental features through the self-attention mechanism, retains the spatial position information of the environmental features through position encoding, and realizes the dimensionality reduction and fusion of the multi-scale environmental features through multi-layer nonlinear transformation; the environmental state encoding vector includes the position, size and density parameters of the obstacles corresponding to the spatial obstacle distribution characteristics and the instantaneous speed, movement direction and acceleration trend of the dynamic target corresponding to the dynamic target motion characteristics; the predicted path generation step: based on the environmental state encoding vector, the predicted path sequence of the unmanned vehicle is generated; the multi-dimensional path difference analysis step: the predicted path sequence is compared with the real-time environmental data for multi-dimensional differences to generate a path difference coefficient tensor; the risk area probability positioning step: based on the path difference coefficient tensor, a risk probability distribution map is generated through spatial environment topology analysis: the dynamic path cost calculation step: based on the risk probability distribution map and the environmental state encoding vector, the dynamic path cost vector is calculated; Path planning strategy optimization step: 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, with the dynamic path cost vector as input and the optimized path planning strategy parameters as output; the 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 cumulative cost calculation, and iteratively updates the policy network parameters to meet the cumulative cost minimization constraint; the policy network comprises 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 speed control parameter range; Real-time dynamic path generation step: Based on the path planning strategy, a real-time dynamic path for the unmanned 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 the path point coordinates, the minimum safe distance between the path point and the obstacle in the static obstacle scene is not less than 1.5 meters, and the lateral offset between the path point and the predicted trajectory of the dynamic target in the dynamic target scene is not less than 0.8 meters; the speed control parameters include recommended speed ranges for different scenarios, 6-10 meters / second for straight obstacle-free scenes, no more than 5 meters / second for curved scenes, and no more than 3 meters / second for high-risk area scenes.

2. The method for dynamic path planning of an unmanned vehicle based on multi-source sensor fusion according to claim 1, characterized in that: The multi-source sensor data real-time collection step specifically includes: The visual image data is subjected to adaptive illumination 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-series aligned and fused to generate fused sensor data.

3. The unmanned vehicle dynamic path planning method based on multi-source sensor fusion according to claim 2 is characterized in that: The step of dynamically encoding the environmental state specifically includes: The fused sensor data is input into the feature extraction network to extract multi-scale environmental features. The feature extraction network is a multi-branch heterogeneous network, which includes a visual feature branch, a point cloud feature branch and a radar feature branch: the visual feature branch adopts a 5-layer convolutional neural network CNN, and the input is the corrected visual data. The image edge texture, local contour and global semantic features are extracted through a 3×3 convolution kernel, a convolution operation with a step size of 1 and a 2×2 maximum pooling operation, and the output dimension is a 256-dimensional visual feature vector; the point cloud feature branch adopts a PointNet network structure, and the input is the filtered point cloud data. The point cloud coordinates are aligned through the T-Net module and the point cloud spatial distribution features are extracted through the MLP layer, and the output dimension is a 256-dimensional point cloud feature vector; the radar feature branch adopts a 2-layer temporal convolutional network TCN, and the input is the time series sequence of the compensated reflection data. The convolution kernel captures the temporal dependency of radar data and outputs a 128-dimensional radar feature vector. The feature extraction network combines the above visual feature vectors, point cloud feature vectors, and radar feature vectors into a 640-dimensional multi-scale environmental feature through feature splicing. The scale levels include Meter-level detail features, Meter-level local features, Meter-level global features; the multi-scale environmental features are subjected to dimensionality reduction coding processing through 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 method for dynamic path planning of an unmanned vehicle 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 smoothed and optimized to generate a predicted path sequence; the predicted path sequence includes a path point coordinate sequence in a time dimension.

5. The method for dynamic path planning of an unmanned vehicle based on multi-source sensor fusion according to claim 4, characterized in that: The multidimensional path difference analysis step specifically includes: Extracting a path point coordinate sequence from the predicted path sequence and an actual obstacle position sequence from the real-time environment data; calculating a spatial offset between the path point coordinate sequence and the actual obstacle position sequence; calculating a matching degree between a path execution delay amount and a dynamic target motion feature in a time dimension; and integrating the spatial offset, the path execution delay amount, and the matching degree of the dynamic target motion feature into a normalized tensor to generate a path difference coefficient tensor.

6. The method for dynamic path planning of an unmanned vehicle based on multi-source sensor fusion according to claim 5, characterized in that: The risk area probability positioning step specifically includes: Construct a spatial environment topology map of the unmanned vehicle's driving environment, wherein the spatial environment topology map includes road node connection relationships and obstacle density parameters; map the path difference coefficient tensor to the corresponding nodes of the spatial environment topology map; calculate the risk diffusion path through a graph structure propagation algorithm; and generate a risk probability distribution map based on the risk diffusion path, wherein the risk probability distribution map marks the coordinates of high-risk probability areas.

7. The method for dynamic path planning of an unmanned vehicle based on multi-source sensor fusion according to claim 6, characterized in that: The dynamic path cost calculation step specifically includes: Extract the coordinates of the high-risk probability area in the risk probability distribution map; calculate the path safety cost component based on the spatial obstacle distribution characteristics in the environmental state coding vector; calculate the path efficiency cost component based on the dynamic target motion characteristics in the environmental state coding vector; calculate the path risk cost component based on the coordinates of the high-risk probability area; and perform weighted fusion of 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 method for dynamic path planning of an unmanned vehicle based on multi-source sensor fusion according to claim 7, characterized in that: The path planning strategy optimization step specifically includes: The dynamic path cost vector is input into a strategy learning model; the path planning strategy is updated through a cumulative cost minimization algorithm to generate an optimized path planning strategy; the optimized path planning strategy includes a path point generation rule and a speed control parameter.

9. The method for dynamic path planning of an unmanned vehicle 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 candidate path point sequence is generated; a dynamic feasibility verification process is performed on the candidate path point sequence to generate a real-time dynamic path; the real-time dynamic path includes a path instruction sequence marked with a timestamp.

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

Images