Unmanned vehicle-based automatic object detection, obstacle avoidance and intelligent identification system at azimuth end

Abstract

The invention discloses an unmanned vehicle azimuth end-based object autonomous detection obstacle avoidance intelligent identification system, which comprises the following modules: a sensing data acquisition module, which is used for acquiring and preprocessing multi-source sensing data and generating an environment sensing data set; the feature extraction module is used for inputting the environment perception data set into the Motion-DETR dynamic perception model to generate a target object feature set; the perception graph construction module is used for constructing a dynamic environment perception graph; the threat evaluation module is used for calculating threat levels, generating an azimuth end threat level matrix and establishing a dynamic risk distribution diagram; the path extension optimization module is used for dynamically selecting an extension direction by adopting a path extension optimization method introducing dynamic risk cost; the path cost optimization module is used for dynamically optimizing the expansion structure according to the principle of minimum accumulated path cost value; and the path output module is used for generating an optimal safe driving path in the dynamic environment. The invention provides an intelligent obstacle avoidance path planning method fusing perception mapping and risk optimization.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving and artificial intelligence perception and decision-making technology, and in particular to an intelligent identification system for autonomous detection and obstacle avoidance of objects at the orientation end of an unmanned vehicle. Background Art

[0002] In the field of intelligent driving technology, how to achieve efficient autonomous detection of surrounding objects and safe obstacle avoidance by unmanned vehicles in dynamic and complex environments has always been a core issue in the interdisciplinary research of autonomous driving, robotic perception, and intelligent decision-making. With the continuous expansion of the application scenarios of unmanned vehicles, the requirements for real-time and accurate environmental perception and robustness of autonomous decision-making are constantly increasing. Traditional obstacle avoidance methods that rely on rule-driven or single perception modes are gradually unable to meet the requirements of efficient and safe driving in complex dynamic environments. To improve the environmental understanding and obstacle avoidance decision-making capabilities of unmanned vehicles in complex roads and scenes, academia and industry have introduced advanced perception and decision-making technologies based on deep learning, dynamic modeling, and intelligent reasoning. In particular, new neural network models represented by Transformer have become a key direction in the research of intelligent driving environment perception and path planning in recent years due to their ability to capture spatiotemporal dynamic features in large-scale data.

[0003] Existing environmental perception and obstacle avoidance methods for unmanned vehicles are generally based on lidar point cloud data, camera visual information, or millimeter-wave radar signals, combined with classic object detection networks (such as Faster R-CNN and the YOLO series) to identify objects in static environments, and obstacle avoidance based on path planning algorithms such as A* search, D*Lite, and traditional RRT. Some methods introduce motion prediction modules in the perception stage, such as linear motion estimation based on Kalman filters and short-term trajectory prediction based on trajectory regression networks, which can roughly predict dynamic targets to a certain extent. However, most of these technologies are limited to the perception of environmental states at a single time step and lack fine-grained modeling of future motion trends and threat changes of target objects. This leads to problems such as obstacle avoidance delays, frequent path conflicts, or planning failures in complex dynamic environments, making it difficult to meet the dual requirements of safety and real-time performance in highly dynamic traffic environments.

[0004] When it comes to understanding dynamic environments, existing technologies typically perform object recognition based on static image feature extraction, ignoring the critical role of object motion and trend information in risk assessment. Even when short-term trajectory prediction is introduced, it is often independent of threat level modeling, failing to organically integrate motion trends with local threat perception and lacking a unified dynamic risk distribution modeling mechanism. This fragmented approach makes the system prone to the risk of local path infeasibility, decision lag, and even emergency braking when encountering high-density, irregularly moving target groups, seriously impacting driving safety and stability.

[0005] In terms of path planning optimization, existing methods based on Rapid Random Trees (RRT) or its improved versions (such as RRT* and Informed RRT*) mostly perform extended searches based on static environmental maps. The path cost function is only related to the spatial location of obstacles. They lack a real-time incorporation of dynamic threat information and a cost adjustment mechanism, making it difficult to generate forward-looking safe driving paths in complex dynamic environments. Especially when faced with fast-approaching dynamic obstacles, traditional RRT, lacking a dynamic risk perception mechanism, easily leads to local optimality or dead zones in the extended direction, significantly degrading obstacle avoidance performance.

[0006] In terms of environmental modeling granularity, existing methods generally use a global unified cost map or coarse-grained region partitioning, failing to perform fine-grained dynamic threat modeling tailored to the autonomous vehicle's own positional characteristics. In reality, during actual driving, the importance of different directions (such as the front, left front, and right rear) to driving decisions varies significantly. A single global cost field cannot reflect this local threat heterogeneity, making it difficult to accurately prioritize key threat directions during path planning, reducing obstacle avoidance efficiency and the rationality of driving decisions.

[0007] In terms of perception models, although some research has introduced Transformer-based target detection networks, these have primarily focused on static object detection, and modeling the continuous motion trends of dynamic targets remains insufficient. Existing publicly available technologies have yet to systematically integrate dynamic perception networks with azimuth-end threat classification, local risk modeling, and dynamic path optimization. They lack a complete system integrating perception and decision-making with a dynamic feedback loop. Furthermore, existing systems generally lack mechanisms for real-time path adjustment based on dynamic environmental perception updates. This results in delayed path replanning and a sudden increase in security risks in the event of drastic environmental changes.

[0008] Therefore, existing technologies lack an intelligent autonomous obstacle avoidance system that can accurately extract the target's motion trends based on dynamic perception models, dynamically establish future motion trajectories and risk distribution models based on azimuth-based threat level classification, and embed dynamic threat prediction results into the path extension and optimization process in real time. This is particularly true in complex, multi-target dynamic environments, high-density traffic flows, or irregularly moving targets. Existing technologies still have significant deficiencies in terms of spatiotemporal consistency in environmental understanding, real-time dynamic threat response, and robust path planning, necessitating the development of new technical solutions to address these challenges. Summary of the Invention

[0009] One purpose of the present invention is to propose an intelligent identification system for autonomous detection and obstacle avoidance of objects at the azimuth end of an unmanned vehicle. The present invention integrates dynamic environment perception modeling and a fast random tree optimization method that introduces dynamic risk cost to construct an intelligent autonomous obstacle avoidance system based on dynamic threat prediction at the azimuth end. The motion trend characteristics of each target object in the environment are extracted by introducing the Motion-DETR dynamic perception model, and a dynamic environment perception map is constructed in combination with the current position of the unmanned vehicle. The azimuth end threat level matrix is generated according to the azimuth division rule, and a dynamic risk distribution model is further established through trajectory prediction and local risk modeling. The threat cost correction factor is dynamically introduced in the path planning stage, and the path extension strategy is optimized in real time. It has the advantages of high environmental perception accuracy, fast dynamic threat response, strong intelligent obstacle avoidance decision-making and high path planning robustness. It is suitable for intelligent decision-making tasks of autonomous detection, obstacle avoidance and safe driving of unmanned vehicles in complex dynamic environments.

[0010] The autonomous object detection and obstacle avoidance intelligent identification system for an unmanned vehicle according to an embodiment of the present invention includes the following modules:

[0011] The perception data acquisition module is used to collect multi-source perception data of the environment around the unmanned vehicle, perform time synchronization, spatial alignment and standardization preprocessing, and generate an environmental perception data set;

[0012] The feature extraction module is used to input the environmental perception data set into the Motion-DETR dynamic perception model, extract the information of each target object in the environment, and generate the target object feature set;

[0013] A perception map construction module is used to construct a dynamic environment perception map based on the target object feature set and with the current position of the unmanned vehicle as the center, and divide the dynamic environment perception map into multiple orientation end areas according to preset orientation division rules;

[0014] The threat assessment module is used to calculate the threat level in each end-to-end area based on the target object's information, generate an end-to-end threat level matrix, and predict the future movement trajectory based on the target object's movement trend to establish a dynamic risk distribution map;

[0015] The path extension optimization module is used to take the dynamic risk distribution map as input, set the current position of the unmanned vehicle as the starting point of path planning, and adopt the path extension optimization method that introduces dynamic risk cost to dynamically select the extension direction;

[0016] The path cost optimization module is used to update the cumulative path cost of each extended node in real time during the path extension process, and dynamically optimize the extension structure based on the principle of minimizing the cumulative path cost;

[0017] The path output module is used to trace back from the end node along the parent node link to the starting node to generate the optimal safe driving path in a dynamic environment.

[0018] Through a modular design, this system organically integrates six core functions: environmental perception, dynamic graph modeling, threat assessment, path expansion, path cost optimization, and path output, creating an intelligent obstacle avoidance system architecture with complete closed-loop capabilities. Clear data flows and functional boundaries between modules ensure functional coordination and consistent response in complex dynamic scenarios. This modular design not only improves the system's maintainability and scalability, but also provides strong support for subsequent algorithm optimization and system iteration, greatly enhancing the system's adaptability and engineering capabilities across different platforms and scenarios.

[0019] According to an embodiment of the present invention, a method for intelligently identifying objects based on autonomous detection and obstacle avoidance at the azimuth end of an unmanned vehicle includes the following steps:

[0020] S1. Collect multi-source perception data of the environment around the unmanned vehicle, and perform time synchronization, spatial alignment, and standardization preprocessing on the multi-source perception data to generate an environmental perception dataset;

[0021] S2. Input the environmental perception dataset into the Motion-DETR dynamic perception model to extract the information of each target object in the environment and form a target object feature set;

[0022] S3. Based on the target object feature set, a dynamic environment perception map is constructed with the current position of the unmanned vehicle as the center, and the dynamic environment perception map is divided into multiple azimuth end areas according to a preset azimuth division rule;

[0023] S4. Calculate the local threat score in each end-to-end area based on the target object information and generate an end-to-end threat level matrix;

[0024] S5. Based on the azimuth threat level matrix and the target object's motion trend information, predict the target object's future motion trajectory and establish a dynamic risk distribution map;

[0025] S6. Based on the dynamic risk distribution map, the current position of the unmanned vehicle is set as the starting point for path planning. The RRT path optimization method with dynamic risk cost is used to perform path extension search. The future motion trajectory and threat level of the target object are used as path cost correction factors to update the dynamic risk cost path extension tree.

[0026] S7. During the path expansion search process, the accumulated cost of each node is updated in real time, and the expansion direction is dynamically selected based on the minimum cost principle to generate the optimal safe driving path in a dynamic environment.

[0027] By constructing a full-process processing framework from multi-source perception data collection, feature extraction, environmental mapping, risk modeling to path planning and output, the Motion-DETR dynamic perception model and the risk distribution modeling method based on threat level and motion trend are introduced to achieve efficient linkage between environmental perception and path decision-making. This solution can simultaneously identify environmental changes and potential risk factors in a dynamic multi-target environment, and deeply integrate risk control with path feasibility calculation through the dynamic generation and real-time optimization of the path extension tree. Compared with traditional path planning methods that rely on static maps or single-target modeling, the present invention significantly improves the safety, consistency and environmental adaptability of path planning in complex working conditions such as multi-source interference, motion conflicts and spatial changes, meeting the key requirements of high-reliability obstacle avoidance and autonomous navigation of unmanned vehicles.

[0028] Optionally, the S1 specifically includes:

[0029] S11. Collect multi-source perception data of the environment around the unmanned vehicle, wherein the multi-source perception data includes visual image data and three-dimensional point cloud data;

[0030] S12, perform time synchronization processing on multi-source perception data, and extract the acquisition timestamp t of the visual image data c The acquisition time stamp t of the 3D point cloud data r , using the nearest neighbor matching strategy, select t c and t r The pair of visual image data and 3D point cloud data with the smallest absolute time difference is paired;

[0031] S13, assigning a unified synchronization timestamp to the paired visual image data and the three-dimensional point cloud data, wherein the unified synchronization timestamp is set to the acquisition timestamp t of the visual image data. c The acquisition time stamp t of the 3D point cloud data r The average value of

[0032] S14, perform spatial alignment processing on the multi-source perception data after time synchronization, based on the external parameter matrix T ext The 3D point cloud data in the radar coordinate system is transformed into the camera coordinate system using the intrinsic parameter matrix K. The transformed 3D point cloud coordinates are projected onto the image plane using the intrinsic parameter matrix to obtain the corresponding 2D pixel positions. Points that exceed the image plane boundary after projection are removed to achieve the fusion of visual image data and 3D point cloud data in a unified spatial reference system.

[0033] S15. Standardizing the multi-source perception data after time synchronization and spatial alignment, wherein the visual image data is normalized to normalize the pixel values to the interval [0, 1], and the three-dimensional point cloud data is normalized according to the set ranging range to normalize the ranging values to the interval [0, 1];

[0034] S16. The standardized multi-source perception data are fused and organized according to a unified spatial index rule to generate an environmental perception data set.

[0035] By preprocessing raw sensor data through temporal synchronization, spatial alignment, and standardization, the system addresses the issues of inconsistent sampling frequencies, inconsistent coordinate systems, and large scale fluctuations in multi-source sensor data, ensuring the accuracy and uniformity of sensor data from the source. In particular, the use of spatial affine mapping and linear interpolation for alignment and infill effectively improves the accuracy of the description of the continuous motion of dynamic objects, providing a high-quality input foundation for subsequent target detection and graph modeling, significantly enhancing the stability and reliability of the system's overall performance.

[0036] Optionally, the S2 specifically includes:

[0037] S21, inputting the environmental perception dataset into the Motion-DETR dynamic perception model;

[0038] S22, through the convolutional neural network set in the visual feature extractor module, the visual image data in the environmental perception dataset is encoded and processed, the visual space feature information is extracted, and the visual feature tensor F is generated. img ;

[0039] S23, through the multi-layer perceptron network set in the point cloud feature extractor module, the three-dimensional point cloud data in the environmental perception data set is encoded point by point, the three-dimensional point cloud spatial geometric features are extracted, and the point cloud feature tensor F is generated. pc ;

[0040] S24, respectively, the visual feature tensor F img And the point cloud feature tensor F pc Introduce the self-attention weight matrix for feature reweighting:

[0041]

[0042] Among them, F′ img is the weighted visual feature tensor, F′ pc is the weighted point cloud feature tensor, is the self-attention weighted matrix corresponding to the visual feature tensor, is the self-attention weighted matrix corresponding to the point cloud feature tensor;

[0043] S25, weighted visual feature tensor F' img and the weighted point cloud feature tensor F′ pc Perform additive fusion processing to generate fused feature tensors through the fusion module:

[0044] F fusion =α·F′ img +(1-α)·F′ pc ;

[0045] Among them, F fusion is the fusion feature tensor, α is the fusion weight coefficient;

[0046] S26, fusion feature tensor F fusion The input is sent to the dynamic feature modeling module, and the multi-head self-attention mechanism set in the dynamic feature modeling module is used to model the dynamic association relationship between feature units, and output the target object feature set O, which includes category information, center position coordinates, velocity vector and motion trend characteristics:

[0047]

[0048] Among them, i is the feature vector of the i-th target object, c i is the category information of the i-th target object, (x i ,y i ) is the center position coordinate of the i-th target object, is the velocity vector of the i-th target object in the x and y directions, θ i is the motion trend feature of the i-th target object.

[0049] By using the Motion-DETR dynamic perception model to perform target detection and feature extraction on standardized perception data, the system can simultaneously achieve spatial positioning and motion state estimation, outputting a high-dimensional feature set that comprehensively characterizes key elements such as the target object's category, position, speed, and direction. This model combines the long-range dependency modeling capabilities of the Transformer structure with the detection efficiency advantages of DETR, improving target recognition accuracy while maintaining real-time performance. It has particularly good recognition and separation capabilities for complex scenarios such as high-speed moving targets, small targets, and overlapping multiple targets, enhancing the system's adaptability to dynamic environments.

[0050] Optionally, the S3 specifically includes:

[0051] S31, based on the target object feature set O, extract the center position coordinates (x i ,y i ), velocity vector and motion trend characteristics θi , taking the current position coordinates (x0, y0) of the unmanned vehicle as a reference, according to the center position coordinates and velocity vector of the target object, according to the set short-term prediction step size, the predicted position is calculated, and then the dynamic polar radius is determined based on the Euclidean distance between the predicted position and the current position of the unmanned vehicle, and the dynamic polar angle is determined based on the direction relationship between the position of the target object and the unmanned vehicle after the short-term prediction;

[0052] S32: Filter out valid target objects that meet the perception radius conditions based on the dynamic polar radius, and construct a valid target object subset O. ′ The perception radius is the maximum perception distance threshold set with the current position of the unmanned vehicle as the center, limiting the processing object to a physical space area that the unmanned vehicle can perceive and respond to in real time;

[0053] S33, with the current position of the unmanned vehicle as the center, based on the valid target object subset O ′ ,construct a dynamic environment perception map, where the nodes of the dynamic environment perception map are valid target objects, and the node attributes include the predicted position, velocity vector and motion trend characteristics of the valid target objects;

[0054] S34, according to the dynamic polar angle corresponding to the node of the dynamic environment perception map, the dynamic environment perception map is divided into N azimuth end areas according to the preset direction division rule, and each azimuth end area is numbered A j , where j∈{1,2,...,N}.

[0055] A dynamic environmental perception map is constructed based on a set of target object features. A polar-coordinate-based azimuth-based region modeling mechanism is introduced to enhance the spatial resolution of threat information from different directions within the perception map. This map structure not only preserves the spatial dependencies between objects in the environment but also, through a partitioning strategy, enables an independent representation of local azimuth risks. This approach is particularly suitable for analyzing the structure of densely sensed areas along the unmanned vehicle's forward trajectory. Compared to traditional planar perception maps, this method improves the response accuracy of azimuth-based obstacle avoidance strategies and provides a spatial structural foundation for subsequent threat assessment and path planning.

[0056] Optionally, the S4 specifically includes:

[0057] S41, based on the dynamic environment perception map, for each azimuth end area A j Extract regional nodes and construct regional node subset O j , where the regional node subset O j Each regional node o i Including predicted position, velocity vector, motion trend characteristics and dynamic polar diameter;

[0058] S42, for each regional node o i The stored velocity vector is calculated for the region node oi The velocity modulus s i ;

[0059] S43, based on each regional node o i Current movement trend characteristics θ i and the motion trend characteristic θ at the previous moment i (t-1) , calculate the rate of change of movement trend Δθ i ;

[0060] S44, statistics area node o i The number of neighbor nodes n within the sensing radius i as local density features;

[0061] S45. For regional node o i , based on the dynamic polar radius r′ i , speed modulus s i , motion trend characteristics θ i , the rate of change of movement trend Δθ i And the local density feature n i , calculate the local threat score:

[0062] μ i =σ(w1·f r (r′ i )+w2·f s (s i )+w3·f θ (θ i )+w4·f Δθ (Δθ i )+w5·f d (n i ));

[0063] Among them, μ i is the local threat score, σ(·) is the Sigmoid normalized activation function, which compresses the local threat score to (0,1), and f r (r′ i ) is the dynamic polar diameter r′ i Normalized feature processing results, f s (s i ) is the velocity modulus s i Normalized feature processing results, f θ (θ i ) is the motion trend feature θ i Normalized feature processing results, f Δθ (Δθ i ) is the rate of change of motion trend Δθ i Normalized feature processing results, f d (n i) is the local density feature n i Normalized feature processing results, w1, w2, w3, w4, w5 are the local threat score fusion weighting coefficients;

[0064] S45, in each azimuth end area A j Based on the local threat score μ i , initialize the azimuth threat level matrix M j , the local threat score μ i Fill in the threat level matrix M at the azimuth end according to the regional node index order j Corresponding position, complete the azimuth threat level matrix M j The generation of the azimuth end area A j Associated storage.

[0065] The system quantitatively analyzes the speed, direction, and distance of target objects within each azimuth region, calculates the threat level, and constructs a hierarchy matrix. It also combines target trajectory trends to predict potential risk areas in the future, ultimately establishing a dynamic risk distribution model. This model enables proactive assessment and dynamic adjustment of potential collision risks for autonomous vehicles in multiple locations, significantly improving its response to unexpected obstacles or high-threat targets. By continuously updating the hierarchy matrix, the system maintains sensitivity to dynamic changes in the environment, enhancing the proactive and defensive nature of its obstacle avoidance strategy.

[0066] Optionally, the S5 specifically includes:

[0067] S51, based on the azimuth end threat level matrix M j , extract each azimuth end area A j The local threat score of the inner region node at time step t Current Location Velocity vector and movement trend characteristics Calculate the trajectory position in the next K time steps and generate the predicted trajectory set P i :

[0068]

[0069] S52, based on the predicted trajectory set P i The predicted location of each regional node in and the local threat score of the regional node at the current time step t Will Directly extend to each future time step t+k and set the future local threat score

[0070] S53, at each time step t+k in the future, based on the predicted trajectory set Pi The predicted location of each regional node in Local density feature n i and future local threat scores Fusion generates dynamic risk distribution map R (t) :

[0071]

[0072] Among them, R (t+k) is the dynamic risk distribution diagram for each time step t+k in the future, and R(x,y,t+k) is the dynamic risk distribution diagram R for each time step t+k in the future. (t+k) The risk intensity of each location point in O ′ is a subset of valid target objects, is the predicted trajectory set P i The predicted location of each regional node in The Euclidean distance from the reference point (x, y), α is the Euclidean distance influence weight coefficient, and γ is the local node density influence weight coefficient.

[0073] By integrating the azimuth threat level matrix with motion trend information, the system constructs a dynamic risk distribution map with temporal prediction capabilities. This system also incorporates a risk evolution mechanism before path decisions are made. This ensures that path planning is based not only on static risk but also on future risk trends, strengthening the temporal dimension of path safety. This mechanism, particularly in multi-target interaction or dense environments, can proactively avoid potential high-risk intersections, effectively reducing the probability of path conflicts and planning failures, and achieving dynamic obstacle avoidance within a time window.

[0074] Optionally, the RRT path optimization method introducing dynamic risk cost specifically includes:

[0075] During the extended sampling process, the extended sampling point (x s ,y s ) Limited to low-risk areas Internal generation, the low-risk domain is defined as the area that satisfies R(x, y, t) ≤ δ, where δ is the set risk threshold;

[0076] In the process of path expansion, based on the dynamic risk distribution map R (t) , initialize the dynamic risk cost path expansion tree Τ, search and expand the sampling point (x s ,y s ) The node v with the closest Euclidean distance near , generating an extended path segment, the extended path segment points to the extended sampling point (x s ,y s ), and calculate v near With the new node v newThe Euclidean distance d(v near ,v new ), real-time selection of dynamic evaluation benchmarks (x m ,y m ) corresponding to the risk intensity R(x m ,y m ,t), R(x m ,y m ,t) and d(v near ,v new ) to form the basic dynamic risk cost C risk (v near ,v new ):

[0077] C risk (v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t));

[0078] Among them, ρ0 is the dynamic risk level weighting coefficient, which is used to regulate the response sensitivity of the extended path segment to changes in environmental risks;

[0079] Based on the predicted trajectory set P i , calculate the path cost correction factor Φ(x m ,y m ,t), the path cost correction factor Φ(x m ,y m ,t) is determined by the trajectory deviation TrjBias(x m ,y m ) and future local threat scores Weighted combination:

[0080]

[0081] Among them, ρ1 is the trajectory deviation weighting coefficient, ρ2 is the threat level weighting coefficient;

[0082] Basic dynamic risk cost C risk (v near ,v new ) and the path cost correction factor Φ(x m ,y m ,t) form the evaluation value Q(v near ,v new ):

[0083] Q(v near ,v new)=d(v near ,v new )×(1+ρ0R(x m ,y m ,t))×(1+Φ(x m ,y m ,t));

[0084] In the process of path extension, the evaluation value Q(v near ,v new ) minimization principle, dynamically select the expansion direction and expansion base point, update the dynamic risk cost path expansion tree Τ, and continuously optimize the path expansion process by refreshing the basic dynamic risk cost status and path cost correction factor status in real time.

[0085] By introducing dynamic risk and path cost correction factors, the system establishes a composite path segment assessment mechanism. This upgrades the traditional path extension cost from a single Euclidean distance to a multi-factor weighted indicator that integrates risk, trajectory deviation, and threat level, enhancing the context-awareness of path selection. This mechanism assigns differentiated safety weights to different path segments, guiding routes to avoid high-risk areas early in path extension, improving the proactiveness of expansion strategies and the speed of path convergence. This is particularly suitable for applications requiring a high degree of responsiveness in dynamic, high-frequency environments.

[0086] Optionally, the S6 specifically includes:

[0087] S61, based on the dynamic risk distribution map R (t) , introduce the dynamic risk cost RRT path optimization method to perform path extension search, set the current position of the unmanned vehicle (x0, y0) as the path extension search starting point, and initialize the dynamic risk cost path extension tree Τ;

[0088] S62. Define low-risk domains:

[0089]

[0090] Among them, R(x,y,t) is the dynamic risk distribution map R of each time step t+k in the future (t+k) The risk intensity of each location point in the ,δ is the set risk threshold; in the low risk domain Extended sampling is performed within the sample, and extended sampling points (x s ,y s );

[0091] S63. In the dynamic risk cost path expansion tree T, search and expand the sampling point (x s ,y s ) The node v with the closest Euclidean distance near , and v nearAs the extended base point, an extended path segment is generated, and the extended path segment points to the extended sampling point (x s ,y s ), select the middle reference point (x m ,y m ) as a dynamic evaluation benchmark, and the target object's future motion trajectory and threat level as path cost correction factors, and perform real-time evaluation:

[0092] Q(v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t))×(1+Φ(x m ,y m ,t));

[0093] Among them, Q(v near ,v new ) is the evaluation value of the extended path segment, v new is a new node, d(v near ,v new ) is the Euclidean distance between two nodes, ρ0 is the risk level weighting coefficient, R(x m ,y m ,t) is the middle reference point (x m ,y m ) in the dynamic risk distribution map R (t) The corresponding risk intensity, path cost correction factor Φ(x m ,y m ,t) is defined as:

[0094]

[0095] Among them, ρ1 is the trajectory deviation weighting coefficient, TrjBias(x m ,y m ) is (x m ,y m ) and the predicted trajectory set P i The deviation between the predicted positions of the nodes in each region, ρ2 is the threat level weighting coefficient, To predict the trajectory set P at each time step t+k in the future i The future local threat score of each regional node in the;

[0096] S64, based on the evaluation value Q(v near ,v new ), select the new node v with the best performance in the current extended path segment new, update the dynamic risk cost path expansion tree Τ.

[0097] During the path expansion process, a dynamic risk-driven expansion search strategy, combined with sampling restrictions in low-risk areas and the principle of minimizing the overall path segment evaluation value, effectively improves the robustness and safety of the expanded path in dynamic environments. The system can filter high-risk areas in real time during the growth of the expansion tree and dynamically adjust the expansion direction based on the risk cost, significantly reducing the probability of the path falling into local extremes. Furthermore, this strategy, coupled with the output of the Motion-DETR model, enables rapid response and path reconstruction to newly emerging targets, enhancing the system's real-time obstacle avoidance capabilities and path adaptability.

[0098] Optionally, the S7 specifically includes:

[0099] During the path expansion search process, the accumulated cost value of each sampling node is updated in real time, and the expansion direction is dynamically selected according to the minimum cost principle to generate the optimal safe driving path in a dynamic environment.

[0100] S71, the new node v with the best performance in the current extended path segment new As the new extension base point, based on the evaluation value Q(v near ,v new ), calculate the path extension search starting point v start to v new The cumulative path cost C total (v new ):

[0101]

[0102] Among them, P(v start ,v new ) is v start to v new The set of path segments;

[0103] S72. During the path expansion process, the cumulative path cost value C of each expansion base point in the dynamic risk cost path expansion tree T is updated in real time. total (v new ), and with the current expansion base point v near For reference, from v near Filter the expansion base point v with the minimum cumulative path cost value from the set of adjacent expandable nodes best , as the next path expansion direction;

[0104] S73, compare v new The cumulative path cost C total (v new ) and the sum of the cumulative path costs of the candidate parent node to, if there is a candidate parent node valt So that through v alt The cumulative path cost value is smaller, then update v new , optimizes the connection relationship of the dynamic risk cost path extension tree Τ;

[0105] S74, when the path extension reaches the end node v global Or when the extension termination condition is met, from the end node v global Start by tracing back along the parent node to the path extension search starting point v start , connect each node in turn to generate the optimal safe driving path based on the minimum cumulative path cost in a dynamic environment.

[0106] The system introduces a dynamic update mechanism for cumulative path costs during the expansion search process, and generates the final path according to the principle of minimum cumulative cost during path backtracking, achieving a complete closed loop from "single expansion optimization" to "global path optimization." By continuously recording the cumulative cost of each expansion node and adjusting the parent node structure, the system effectively avoids the residual impact of early low-quality paths and improves the global rationality of the final path. This mechanism is particularly suitable for scenarios requiring frequent path adjustments in dynamic risk environments, ensuring that each path output is optimal within the current environment, enhancing system stability and practicality.

[0107] The beneficial effects of the present invention are:

[0108] The present invention aims to solve the key problem of realizing azimuth-end object detection and obstacle avoidance recognition for unmanned vehicles in complex dynamic environments, and constructs an intelligent recognition system based on dynamic perception graph modeling and path planning optimization. By collecting multi-source heterogeneous sensor data and completing synchronization, alignment and standardized preprocessing in time and space dimensions, a highly consistent environmental perception data set is generated, ensuring the robustness and real-time performance of the perception input. In the feature extraction stage, the system introduces the Motion-DETR dynamic perception model, combines target detection with trajectory association strategies, and realizes the precise extraction and structured modeling of different types of dynamic target objects, making up for the problems of missed and false detection of targets in dynamic scenes by traditional static image perception methods.

[0109] Furthermore, the present invention constructs a dynamic environmental perception map centered on the unmanned vehicle, and uses multi-directional division rules to map environmental information to multiple directional end areas. Combined with characteristic information such as the speed and direction of the target objects in each area, the threat level is calculated and a dynamic risk distribution model is established, thereby realizing spatial distribution modeling and forward warning of potential risk areas, and effectively improving the perception of target movement trends and dynamic changes in risks during the path planning stage. In the path planning optimization process, the system introduces a dual mechanism of dynamic risk cost and path cost correction factor, embeds the target object's movement trend deviation and threat level into the path cost calculation, and dynamically adjusts the path expansion strategy. This mechanism significantly improves the system's path adaptability and safety in the face of uncertain environmental changes, and breaks through the problem that the traditional RRT path expansion method cannot be globally optimized in complex obstacle-dense areas.

[0110] Furthermore, during the path search process, this invention constructs a path extension tree through a dynamic update mechanism that accumulates path cost values. It then dynamically optimizes the extension structure using a minimum cost criterion, ultimately backtracking to generate the optimal and safest driving path. Compared to traditional static path assessment strategies, this solution integrates local environmental risks and global cost information in real time, dynamically revising the extension structure and outputting an overall optimal path, thereby improving the global rationality and local safety of the navigation path. Ultimately, the system outputs the optimal path as a driving command, providing highly secure and robust intelligent obstacle avoidance and path decision support for unmanned vehicles.

[0111] In summary, the present invention constructs an end-to-end intelligent obstacle avoidance system architecture by integrating multi-source perception, dynamic graph modeling, path cost optimization and intelligent expansion strategy. It solves the problems of low obstacle avoidance accuracy, delayed response, and unstable path of existing unmanned vehicle systems in dynamic multi-target environments. It has strong environmental adaptability, path feasibility and deployment practicality, and significantly improves the navigation and decision-making capabilities of intelligent vehicle systems in complex actual scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0113] Figure 1 This is a schematic diagram of the structure of the intelligent recognition system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end proposed by the present invention;

[0114] Figure 2 This is the overall flow chart of the intelligent identification method for autonomous detection and obstacle avoidance of objects at the unmanned vehicle's azimuth end proposed by the present invention. DETAILED DESCRIPTION

[0115] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0116] refer to Figure 1 The intelligent recognition system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's positioning terminal includes the following modules:

[0117] The perception data acquisition module is used to collect multi-source perception data of the environment around the unmanned vehicle, perform time synchronization, spatial alignment and standardization preprocessing, and generate an environmental perception data set;

[0118] The feature extraction module is used to input the environmental perception data set into the Motion-DETR dynamic perception model, extract the information of each target object in the environment, and generate the target object feature set;

[0119] A perception map construction module is used to construct a dynamic environment perception map based on the target object feature set and with the current position of the unmanned vehicle as the center, and divide the dynamic environment perception map into multiple orientation end areas according to preset orientation division rules;

[0120] The threat assessment module is used to calculate the threat level in each end-to-end area based on the target object's information, generate an end-to-end threat level matrix, and predict the future movement trajectory based on the target object's movement trend to establish a dynamic risk distribution map;

[0121] The path extension optimization module is used to take the dynamic risk distribution map as input, set the current position of the unmanned vehicle as the starting point of path planning, and adopt the path extension optimization method that introduces dynamic risk cost to dynamically select the extension direction;

[0122] The path cost optimization module is used to update the cumulative path cost of each extended node in real time during the path extension process, and dynamically optimize the extension structure based on the principle of minimizing the cumulative path cost;

[0123] The path output module is used to trace back from the end node along the parent node link to the starting node to generate the optimal safe driving path in a dynamic environment.

[0124] refer to Figure 2 The method for autonomously detecting and avoiding obstacles based on the unmanned vehicle's position-side object intelligent identification includes the following steps:

[0125] S1. Collect multi-source perception data of the environment around the unmanned vehicle, and perform time synchronization, spatial alignment, and standardization preprocessing on the multi-source perception data to generate an environmental perception dataset;

[0126] S2. Input the environmental perception dataset into the Motion-DETR dynamic perception model to extract the information of each target object in the environment and form a target object feature set;

[0127] S3. Based on the target object feature set, a dynamic environment perception map is constructed with the current position of the unmanned vehicle as the center, and the dynamic environment perception map is divided into multiple azimuth end areas according to a preset azimuth division rule;

[0128] S4. Calculate the local threat score in each end-to-end area based on the target object information and generate an end-to-end threat level matrix;

[0129] S5. Based on the azimuth threat level matrix and the target object's motion trend information, predict the target object's future motion trajectory and establish a dynamic risk distribution map;

[0130] S6. Based on the dynamic risk distribution map, the current position of the unmanned vehicle is set as the starting point for path planning. The RRT path optimization method with dynamic risk cost is used to perform path extension search. The future motion trajectory and threat level of the target object are used as path cost correction factors to update the dynamic risk cost path extension tree.

[0131] S7. During the path expansion search process, the accumulated cost of each node is updated in real time, and the expansion direction is dynamically selected based on the minimum cost principle to generate the optimal safe driving path in a dynamic environment.

[0132] In this embodiment, S1 specifically includes:

[0133] S11. Collect multi-source perception data of the environment around the unmanned vehicle, wherein the multi-source perception data includes visual image data and three-dimensional point cloud data;

[0134] S12, perform time synchronization processing on multi-source perception data, and extract the acquisition timestamp t of the visual image data c The acquisition time stamp t of the 3D point cloud data r , using the nearest neighbor matching strategy, select t c and t r The pair of visual image data and 3D point cloud data with the smallest absolute time difference is paired;

[0135] S13, assigning a unified synchronization timestamp to the paired visual image data and the three-dimensional point cloud data, wherein the unified synchronization timestamp is set to the acquisition timestamp t of the visual image data. c The acquisition time stamp t of the 3D point cloud data r The average value of

[0136] S14, perform spatial alignment processing on the multi-source perception data after time synchronization, based on the external parameter matrix T extThe 3D point cloud data in the radar coordinate system is transformed into the camera coordinate system using the intrinsic parameter matrix K. The transformed 3D point cloud coordinates are projected onto the image plane using the intrinsic parameter matrix to obtain the corresponding 2D pixel positions. Points that exceed the image plane boundary after projection are removed to achieve the fusion of visual image data and 3D point cloud data in a unified spatial reference system.

[0137] S15. Standardizing the multi-source perception data after time synchronization and spatial alignment, wherein the visual image data is normalized to normalize the pixel values to the interval [0, 1], and the three-dimensional point cloud data is normalized according to the set ranging range to normalize the ranging values to the interval [0, 1];

[0138] S16. The standardized multi-source perception data are fused and organized according to a unified spatial index rule to generate an environmental perception data set.

[0139] In this embodiment, S2 specifically includes:

[0140] S21, inputting the environmental perception dataset into the Motion-DETR dynamic perception model;

[0141] S22, through the convolutional neural network set in the visual feature extractor module, the visual image data in the environmental perception dataset is encoded and processed, the visual space feature information is extracted, and the visual feature tensor F is generated. img ;

[0142] S23, through the multi-layer perceptron network set in the point cloud feature extractor module, the three-dimensional point cloud data in the environmental perception data set is encoded point by point, the three-dimensional point cloud spatial geometric features are extracted, and the point cloud feature tensor F is generated. pc ;

[0143] S24, respectively, the visual feature tensor F img And the point cloud feature tensor F pc Introduce the self-attention weight matrix for feature reweighting:

[0144]

[0145] Among them, F′ img is the weighted visual feature tensor, F′ pc is the weighted point cloud feature tensor, is the self-attention weighted matrix corresponding to the visual feature tensor, is the self-attention weighted matrix corresponding to the point cloud feature tensor;

[0146] S25, weighted visual feature tensor F' img and the weighted point cloud feature tensor F′ pcPerform additive fusion processing to generate fused feature tensors through the fusion module:

[0147] F fusion =α·F′ img +(1-α)·F′ pc ;

[0148] Among them, F fusion is the fusion feature tensor, α is the fusion weight coefficient;

[0149] S26, fusion feature tensor F fusion The input is sent to the dynamic feature modeling module, and the multi-head self-attention mechanism set in the dynamic feature modeling module is used to model the dynamic association relationship between feature units, and output the target object feature set O, which includes category information, center position coordinates, velocity vector and motion trend characteristics:

[0150]

[0151] Among them, i is the feature vector of the i-th target object, c i is the category information of the i-th target object, (x i ,y i ) is the center position coordinate of the i-th target object, is the velocity vector of the i-th target object in the x and y directions, θ i is the motion trend feature of the i-th target object.

[0152] In this embodiment, S3 specifically includes:

[0153] S31, based on the target object feature set O, extract the center position coordinates (x i ,y i ), velocity vector and motion trend characteristics θ i , taking the current position coordinates (x0, y0) of the unmanned vehicle as a reference, according to the center position coordinates and velocity vector of the target object, according to the set short-term prediction step size, the predicted position is calculated, and then the dynamic polar radius is determined based on the Euclidean distance between the predicted position and the current position of the unmanned vehicle, and the dynamic polar angle is determined based on the direction relationship between the position of the target object and the unmanned vehicle after the short-term prediction;

[0154] S32: Filter out valid target objects that meet the perception radius conditions based on the dynamic polar radius, and construct a valid target object subset O. ′ The perception radius is the maximum perception distance threshold set with the current position of the unmanned vehicle as the center, limiting the processing object to a physical space area that the unmanned vehicle can perceive and respond to in real time;

[0155] S33, with the current position of the unmanned vehicle as the center, based on the valid target object subset O ′ ,construct a dynamic environment perception map, where the nodes of the dynamic environment perception map are valid target objects, and the node attributes include the predicted position, velocity vector and motion trend characteristics of the valid target objects;

[0156] S34, according to the dynamic polar angle corresponding to the node of the dynamic environment perception map, the dynamic environment perception map is divided into N azimuth end areas according to the preset direction division rule, and each azimuth end area is numbered A j , where j∈{1,2,...,N}.

[0157] In this embodiment, the S4 specifically includes:

[0158] S41, based on the dynamic environment perception map, for each azimuth end area A j Extract regional nodes and construct regional node subset O j , where the regional node subset O j Each regional node o i Including predicted position, velocity vector, motion trend characteristics and dynamic polar diameter;

[0159] S42, for each regional node o i The stored velocity vector is calculated for the region node o i The velocity modulus s i ;

[0160] S43, based on each regional node o i Current movement trend characteristics θ i Compared with the movement trend characteristics of the previous moment Calculate the rate of change of movement trend Δθ i ;

[0161] S44, statistics area node o i The number of neighbor nodes n within the sensing radius i as local density features;

[0162] S45. For regional node o i , based on the dynamic polar radius r′ i , speed modulus s i , motion trend characteristics θ i , the rate of change of movement trend Δθ i And the local density feature n i , calculate the local threat score:

[0163] μ i =σ(W1·f r (r′ i )+w2·fs (s i )+w3·f θ (θ i )+w4·f Δθ (Δθ i )+w5·f d (n i ));

[0164] Among them, μ i is the local threat score, σ(·) is the Sigmoid normalized activation function, which compresses the local threat score to (0,1), and f r (r′ i ) is the dynamic polar diameter r′ i Normalized feature processing results, f s (s i ) is the velocity modulus s i Normalized feature processing results, f θ (θ i ) is the motion trend feature θ i Normalized feature processing results, f Δθ (Δθ i ) is the rate of change of motion trend Δθ i Normalized feature processing results, f d (n i ) is the local density feature n i Normalized feature processing results, w1, w2, w3, w4, w5 are the local threat score fusion weighting coefficients;

[0165] S45, in each azimuth end area A j Based on the local threat score μ i , initialize the azimuth threat level matrix M j , the local threat score μ i Fill in the threat level matrix M at the azimuth end according to the regional node index order j Corresponding position, complete the azimuth threat level matrix M j The generation of the azimuth end area A j Associated storage.

[0166] In this embodiment, the S5 specifically includes:

[0167] S51, based on the azimuth end threat level matrix M j , extract each azimuth end area A j The local threat score of the inner region node at time step t Current Location Velocity vector and movement trend characteristics Calculate the trajectory position in the next K time steps and generate the predicted trajectory set P i :

[0168]

[0169] S52, based on the predicted trajectory set P i The predicted location of each regional node in and the local threat score of the regional node at the current time step t Will Directly extend to each future time step t+k and set the future local threat score

[0170] S53, at each time step t+k in the future, based on the predicted trajectory set P i The predicted location of each regional node in Local density feature n i and future local threat scores Fusion generates dynamic risk distribution map R (t) :

[0171]

[0172] Among them, R (t+k) is the dynamic risk distribution diagram for each time step t+k in the future, and R(x,y,t+k) is the dynamic risk distribution diagram R for each time step t+k in the future. (t+k) The risk intensity of each location point in O ′ is a subset of valid target objects, is the predicted trajectory set P i The predicted location of each regional node in The Euclidean distance from the reference point (x, y), α is the Euclidean distance influence weight coefficient, and γ is the local node density influence weight coefficient.

[0173] In this embodiment, the RRT path optimization method introducing dynamic risk cost specifically includes:

[0174] During the extended sampling process, the extended sampling point (x s ,y s ) Limited to low-risk areas Internal generation, the low-risk domain is defined as the area that satisfies R(x, y, t) ≤ δ, where δ is the set risk threshold;

[0175] In the process of path expansion, based on the dynamic risk distribution map R (t) , initialize the dynamic risk cost path expansion tree Τ, search and expand the sampling point (x s ,y s ) The node v with the closest Euclidean distancenear , generating an extended path segment, the extended path segment points to the extended sampling point (x s ,y s ), and calculate v near With the new node v new The Euclidean distance d(v near ,v new ), real-time selection of dynamic evaluation benchmarks (x m ,y m ) corresponding to the risk intensity R(x m ,y m ,t), R(x m ,y m ,t) and d(v near ,v new ) to form the basic dynamic risk cost C risk (v near ,v new ):

[0176] C risk (v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t));

[0177] Among them, ρ0 is the dynamic risk level weighting coefficient, which is used to regulate the response sensitivity of the extended path segment to changes in environmental risks;

[0178] Based on the predicted trajectory set P i , calculate the path cost correction factor Φ(x m ,y m ,t), the path cost correction factor Φ(x m ,y m ,t) is determined by the trajectory deviation TrjBias(x m ,y m ) and future local threat scores Weighted combination:

[0179]

[0180] Among them, ρ1 is the trajectory deviation weighting coefficient, ρ2 is the threat level weighting coefficient;

[0181] Basic dynamic risk cost C risk (v near ,v new ) and the path cost correction factor Φ(x m ,y m,t) form the evaluation value Q(v near ,v new ):

[0182] Q(v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t))×(1+Φ(x m ,y m ,t));

[0183] In the process of path extension, the evaluation value Q(v near ,v new ) minimization principle, dynamically select the expansion direction and expansion base point, update the dynamic risk cost path expansion tree Τ, and continuously optimize the path expansion process by refreshing the basic dynamic risk cost status and path cost correction factor status in real time.

[0184] In this embodiment, S6 specifically includes:

[0185] S61, based on the dynamic risk distribution map R (t) , introduce the dynamic risk cost RRT path optimization method to perform path extension search, set the current position of the unmanned vehicle (x0, y0) as the path extension search starting point, and initialize the dynamic risk cost path extension tree Τ;

[0186] S62. Define low-risk domains:

[0187]

[0188] Among them, R(x,y,t) is the dynamic risk distribution map R of each time step t+k in the future (t+k) The risk intensity of each location point in the ,δ is the set risk threshold; in the low risk domain Extended sampling is performed within the sample, and extended sampling points (x s ,y s );

[0189] S63. In the dynamic risk cost path expansion tree T, search and expand the sampling point (x s ,y s ) The node v with the closest Euclidean distance near , and v near As the extended base point, an extended path segment is generated, and the extended path segment points to the extended sampling point (x s ,y s ), select the middle reference point (x m ,y m) as a dynamic evaluation benchmark, and the target object's future motion trajectory and threat level as path cost correction factors, and perform real-time evaluation:

[0190] Q(v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t))×(1+Φ(x m ,y m ,t));

[0191] Among them, Q(v near ,v new ) is the evaluation value of the extended path segment, v new is a new node, d(v near ,v new ) is the Euclidean distance between two nodes, ρ0 is the risk level weighting coefficient, R(x m ,y m ,t) is the middle reference point (x m ,y m ) in the dynamic risk distribution map R (t) The corresponding risk intensity, path cost correction factor Φ(x m ,y m ,t) is defined as:

[0192]

[0193] Among them, ρ1 is the trajectory deviation weighting coefficient, TrjBias(x m ,y m ) is (x m ,y m ) and the predicted trajectory set P i The deviation between the predicted positions of the nodes in each region, ρ2 is the threat level weighting coefficient, To predict the trajectory set P at each time step t+k in the future i The future local threat score of each regional node in the;

[0194] S64, based on the evaluation value Q(v near ,v new ), select the new node v with the best performance in the current extended path segment new , update the dynamic risk cost path expansion tree Τ.

[0195] In this embodiment, the S7 specifically includes:

[0196] During the path expansion search process, the accumulated cost value of each sampling node is updated in real time, and the expansion direction is dynamically selected according to the minimum cost principle to generate the optimal safe driving path in a dynamic environment.

[0197] S71, the new node v with the best performance in the current extended path segment new As the new extension base point, based on the evaluation value Q(v near ,v new ), calculate the path extension search starting point v start to v new The cumulative path cost C total (v new ):

[0198]

[0199] Among them, P(v start ,v new ) is v start to v new The set of path segments;

[0200] S72. During the path expansion process, the cumulative path cost value C of each expansion base point in the dynamic risk cost path expansion tree T is updated in real time. total (v new ), and with the current expansion base point v near For reference, from v near Filter the expansion base point v with the minimum cumulative path cost value from the set of adjacent expandable nodes best , as the next path expansion direction;

[0201] S73, compare v new The cumulative path cost C total (v new ) and the sum of the cumulative path costs of the candidate parent node to, if there is a candidate parent node v alt So that through v alt The cumulative path cost value is smaller, then update v new , optimizes the connection relationship of the dynamic risk cost path extension tree Τ;

[0202] S74, when the path extension reaches the end node v global Or when the extension termination condition is met, from the end node v global Start by tracing back along the parent node to the path extension search starting point v start , connect each node in turn to generate the optimal safe driving path based on the minimum cumulative path cost in a dynamic environment.

[0203] Example 1:

[0204] In order to verify the feasibility of the present invention in implementation, the present invention was applied to an autonomous driving test conducted in a certain intelligent manufacturing park. In an intelligent manufacturing park, unmanned vehicles face typical high-dynamic mixed scenarios, including interfering factors such as staggered pedestrians, multiple types of transport vehicles, autonomous robots, and temporarily stacked goods. These factors present the complex characteristics of high dynamism, spatial uncertainty, and target diversity. Traditional static map-based and rule-based path planning systems are difficult to operate stably in such unstructured environments. Frequent path interruptions, obstacle avoidance failures, repeated planning delays, and other problems often occur, resulting in poor overall system stability, untimely response, and even potential safety hazards.

[0205] The project team conducted comparative testing in four typical areas of the park: the park's main road, a storage intersection, the factory perimeter, and a storage lane. These areas represent a progression of scenarios from low risk to high interference. The system, integrated and deployed on a general-purpose medium-sized logistics unmanned vehicle platform, features high-precision lidar, millimeter-wave radar, and an RGB-D binocular camera, coupled with a high-performance edge computing module for online reasoning and path planning.

[0206] During the operational process, the perception data acquisition module first synchronizes the time and spatial coordinates of multi-source sensors to construct a unified-scale environmental perception dataset. The Motion-DETR dynamic perception model performs real-time target recognition and feature extraction in the current scene, outputting information including the center point coordinates, velocity vector, type label, and pose estimation for each target, forming a target object feature set. The system constructs a dynamic environmental perception map with the unmanned vehicle as the pole and divides it into azimuth-end regions using polar coordinate partitioning. Each region independently performs a threat level assessment to construct an azimuth-end threat level matrix.

[0207] Based on this matrix and the predicted future motion trends of the target object, the system further develops a dynamic risk distribution model. This model maps the temporal evolution of risk into a two-dimensional space, supporting dynamic risk analysis across multiple regions within a short time window and providing a priori conditions for subsequent path planning. During the path expansion phase, the system employs an improved RRT algorithm that integrates dynamic risk costs with path cost correction factors in path cost calculations to avoid expansion into high-risk areas, improving path safety and convergence speed.

[0208] Each path expansion is completed within 50ms. The system updates the cumulative path cost of the expanded nodes in real time, executes the optimal parent node selection strategy, and backtracks after the expansion is complete to determine the optimal safe path at the current moment. If the dynamic risk distribution suddenly changes (such as the sudden approach of a high-threat target), the system automatically triggers path reconstruction to ensure that the route remains within a feasible, low-risk trajectory. The entire planning and backtracking process maintains high concurrent response, without affecting the simultaneous execution of perception and path decision-making.

[0209] The data collected and organized during the experiment are shown in the following table.

[0210] Table 1 Performance comparison results of the present invention and the traditional method

[0211]

[0212]

[0213] The table above compares the core performance of the traditional method and the proposed system in four typical test scenarios. In the scenario of a campus main road with a low proportion of dynamic targets, the traditional method was able to largely complete the obstacle avoidance task, but suffered from poor path stability, delayed response, and high planning time. The proposed system achieved a 100% obstacle avoidance success rate, reduced planning latency by 41.5%, and required no path reconstruction, demonstrating excellent operational stability.

[0214] At intersections in storage areas and at the perimeter of factory buildings, where dynamic interactions are more complex, the traditional approach's obstacle avoidance success rate dropped to 85.4% and 78.1%, respectively. Path reconstructions were repeated as often as four or five times, and the overall time consumption exceeded 65ms. The proposed system, through a dynamic risk guidance mechanism, proactively avoids potentially high-risk paths, maintaining an obstacle avoidance success rate above 91%. The average path planning time was kept under 50ms, and the number of path reconstructions was significantly reduced. Overall path coherence and response efficiency were superior to those of traditional approaches.

[0215] In the warehouse lanes with the most restricted scenarios, the present invention effectively predicts the path blocking trend through the dynamic threat level matrix in a high-density target environment, detouring around low-risk areas in advance, increasing the obstacle avoidance success rate to 97.3%, and stabilizing the path reconstruction within 1 time, effectively ensuring the continuity of path decision-making and safety redundancy capabilities.

[0216] In summary, field testing on multiple road types and mixed dynamic target environments in a typical industrial park demonstrates that the proposed intelligent obstacle avoidance and autonomous object detection system for unmanned vehicles significantly addresses the core issues of traditional path planning methods, such as response lag, poor path stability, lack of risk perception, and insufficient path optimization granularity under dynamic target interference. Test results demonstrate that the system maintains a high obstacle avoidance success rate, low path reconstruction times, and stable path planning delays under varying target density, dynamic ratio, and scene complexity, demonstrating excellent path coherence and adaptability to dynamic environments. In particular, the system demonstrates strong robustness and high stability in complex areas with high dynamics and high-frequency interference, fully demonstrating the efficient collaboration between its perception, modeling, and decision-making modules. Relying on the Motion-DETR perception model, dynamic risk distribution modeling, and an extended mechanism that introduces a path cost correction factor, the system outperforms existing solutions in terms of path generation quality, depth of environmental perception, and overall resource efficiency. Comprehensive evaluation shows that the present invention not only excels in obstacle avoidance safety, system real-time performance, and structural flexibility, but also has good engineering feasibility and potential for implementation and promotion. It is suitable for various intelligent driving application scenarios such as smart logistics, unmanned warehousing vehicles, and industrial park inspections, and has important engineering promotion value and industrial transformation prospects.

[0217] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. Based on the autonomous object detection and obstacle avoidance intelligent identification system of the unmanned vehicle, it is characterized by: Includes the following modules: The perception data acquisition module is used to collect multi-source perception data of the environment around the unmanned vehicle, perform time synchronization, spatial alignment and standardization preprocessing, and generate an environmental perception data set; The feature extraction module is used to input the environmental perception data set into the Motion-DETR dynamic perception model, extract the information of each target object in the environment, and generate the target object feature set; A perception map construction module is used to construct a dynamic environment perception map based on the target object feature set and with the current position of the unmanned vehicle as the center, and divide the dynamic environment perception map into multiple orientation end areas according to preset orientation division rules; The threat assessment module is used to calculate the threat level in each end-to-end area based on the target object's information, generate an end-to-end threat level matrix, and predict the future movement trajectory based on the target object's movement trend to establish a dynamic risk distribution map; The path extension optimization module is used to take the dynamic risk distribution map as input, set the current position of the unmanned vehicle as the starting point of path planning, and adopt the path extension optimization method that introduces dynamic risk cost to dynamically select the extension direction; The path cost optimization module is used to update the cumulative path cost of each extended node in real time during the path extension process, and dynamically optimize the extension structure based on the principle of minimizing the cumulative path cost; The path output module is used to trace back from the end node along the parent node link to the starting node to generate the optimal safe driving path in a dynamic environment.

2. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 1 is characterized in that: The modules are implemented as follows: S1. Collect multi-source perception data of the environment around the unmanned vehicle, and perform time synchronization, spatial alignment, and standardization preprocessing on the multi-source perception data to generate an environmental perception dataset; S2. Input the environmental perception dataset into the Motion-DETR dynamic perception model to extract the information of each target object in the environment and form a target object feature set; S3. Based on the target object feature set, a dynamic environment perception map is constructed with the current position of the unmanned vehicle as the center, and the dynamic environment perception map is divided into multiple azimuth end areas according to a preset azimuth division rule; S4. Calculate the local threat score in each end-to-end area based on the target object information and generate an end-to-end threat level matrix; S5. Based on the azimuth threat level matrix and the target object's motion trend information, predict the target object's future motion trajectory and establish a dynamic risk distribution map; S6. Based on the dynamic risk distribution map, the current position of the unmanned vehicle is set as the starting point for path planning. The RRT path optimization method with dynamic risk cost is used to perform path extension search. The future motion trajectory and threat level of the target object are used as path cost correction factors to update the dynamic risk cost path extension tree. S7. During the path expansion search process, the accumulated cost of each node is updated in real time, and the expansion direction is dynamically selected based on the minimum cost principle to generate the optimal safe driving path in a dynamic environment.

3. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: Said S1 specifically includes: S11. Collect multi-source perception data of the environment around the unmanned vehicle, wherein the multi-source perception data includes visual image data and three-dimensional point cloud data; S12, perform time synchronization processing on multi-source perception data, and extract the acquisition timestamp t of the visual image data c The acquisition time stamp t of the 3D point cloud data r , using the nearest neighbor matching strategy, select t c and t r The pair of visual image data and 3D point cloud data with the smallest absolute time difference is paired; S13, assigning a unified synchronization timestamp to the paired visual image data and the three-dimensional point cloud data, wherein the unified synchronization timestamp is set to the acquisition timestamp t of the visual image data. c The acquisition time stamp t of the 3D point cloud data r The average value of S14, perform spatial alignment processing on the multi-source perception data after time synchronization, based on the external parameter matrix T ext The 3D point cloud data in the radar coordinate system is transformed into the camera coordinate system using the intrinsic parameter matrix K. The transformed 3D point cloud coordinates are projected onto the image plane using the intrinsic parameter matrix to obtain the corresponding 2D pixel positions. Points that exceed the image plane boundary after projection are removed to achieve the fusion of visual image data and 3D point cloud data in a unified spatial reference system. S15. Standardizing the multi-source perception data after time synchronization and spatial alignment, wherein the visual image data is normalized to normalize the pixel values to the interval [0, 1], and the three-dimensional point cloud data is normalized according to the set ranging range to normalize the ranging values to the interval [0, 1]; S16. The standardized multi-source perception data are fused and organized according to a unified spatial index rule to generate an environmental perception data set.

4. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The S2 specifically includes: S21, inputting the environmental perception dataset into the Motion-DETR dynamic perception model; S22, through the convolutional neural network set in the visual feature extractor module, the visual image data in the environmental perception dataset is encoded and processed, the visual space feature information is extracted, and the visual feature tensor F is generated. img ; S23, through the multi-layer perceptron network set in the point cloud feature extractor module, the three-dimensional point cloud data in the environmental perception data set is encoded point by point, the three-dimensional point cloud spatial geometric features are extracted, and the point cloud feature tensor F is generated. pc ; S24, respectively, the visual feature tensor F img And the point cloud feature tensor F pc Introduce the self-attention weight matrix for feature reweighting: Among them, F′ img is the weighted visual feature tensor, F′ pc is the weighted point cloud feature tensor, is the self-attention weighted matrix corresponding to the visual feature tensor, is the self-attention weighted matrix corresponding to the point cloud feature tensor; S25, weighted visual feature tensor F' img and the weighted point cloud feature tensor F′ pc Perform additive fusion processing to generate fused feature tensors through the fusion module: F fusion =α·F′ img +(1-α)·F′ pc ; Among them, F fusion is the fusion feature tensor, α is the fusion weight coefficient; S26, fusion feature tensor F fusion The input is sent to the dynamic feature modeling module, and the multi-head self-attention mechanism set in the dynamic feature modeling module is used to model the dynamic association relationship between feature units, and output the target object feature set O, which includes category information, center position coordinates, velocity vector and motion trend characteristics: Among them, i is the feature vector of the i-th target object, c i is the category information of the i-th target object, (x i ,y i ) is the center position coordinate of the i-th target object, is the velocity vector of the i-th target object in the x and y directions, θ i is the motion trend feature of the i-th target object.

5. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The S3 specifically includes: S31, based on the target object feature set O, extract the center position coordinates (x i ,y i ), velocity vector and motion trend characteristics θ i , taking the current position coordinates (x0, y0) of the unmanned vehicle as a reference, according to the center position coordinates and velocity vector of the target object, according to the set short-term prediction step size, the predicted position is calculated, and then the dynamic polar radius is determined based on the Euclidean distance between the predicted position and the current position of the unmanned vehicle, and the dynamic polar angle is determined based on the direction relationship between the position of the target object and the unmanned vehicle after the short-term prediction; S32. Filter out valid target objects that meet the perception radius condition based on the dynamic polar radius and construct a valid target object subset O′. The perception radius is a maximum perception distance threshold set with the current position of the unmanned vehicle as the center, limiting the processing object to a physical space area that the unmanned vehicle can perceive and respond to in real time; S33. Build a dynamic environment perception map based on the valid target object subset O′ with the current position of the unmanned vehicle as the center, where the nodes of the dynamic environment perception map are valid target objects, and the node attributes include the predicted position, velocity vector, and motion trend characteristics of the valid target objects; S34, according to the dynamic polar angle corresponding to the node of the dynamic environment perception map, the dynamic environment perception map is divided into N azimuth end areas according to the preset direction division rule, and each azimuth end area is numbered A j , where j∈{1,2,...,N}.

6. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The S4 specifically includes: S41, based on the dynamic environment perception map, for each azimuth end area A j Extract regional nodes and construct regional node subset O j , where the regional node subset O j Each regional node o i Including predicted position, velocity vector, motion trend characteristics and dynamic polar diameter; S42, for each regional node o i The stored velocity vector is calculated for the region node o i The velocity modulus s i ; S43, based on each regional node o i Current movement trend characteristics θ i Compared with the movement trend characteristics of the previous moment Calculate the rate of change of movement trend Δθ i ; S44, statistics area node o i The number of neighbor nodes n within the sensing radius i as local density features; S45. For regional node o i , based on the dynamic polar radius r′ i , speed modulus s i , motion trend characteristics θ i , the rate of change of movement trend Δθ i And the local density feature n i , calculate the local threat score: m i =σ(W1·f r (r′ i )+w2·f s (s i )+w3·f θ (i i )+w4·f Δθ (Dth i )+w5·f d (n i )); Among them, μ i is the local threat score, σ(·) is the Sigmoid normalized activation function, which compresses the local threat score to (0,1), and f r (r′ i ) is the dynamic polar diameter r′ i Normalized feature processing results, f s (s i ) is the velocity modulus s i Normalized feature processing results, f θ (θ i ) is the motion trend feature θ i Normalized feature processing results, f Δθ (Δθ i ) is the rate of change of motion trend Δθ i Normalized feature processing results, f d (n i ) is the local density feature n i Normalized feature processing results, w1, w2, w3, w4, w5 are the local threat score fusion weighting coefficients; S45, in each azimuth end area A j Based on the local threat score μ i , initialize the azimuth threat level matrix M j , the local threat score μ i Fill in the threat level matrix M at the azimuth end according to the regional node index order j Corresponding position, complete the azimuth threat level matrix M j The generation of the azimuth end area A j Associated storage.

7. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The S5 specifically includes: S51, based on the azimuth end threat level matrix M j , extract each azimuth end area A j The local threat score of the inner region node at time step t Current Location Velocity vector and movement trend characteristics Calculate the trajectory position in the next K time steps and generate the predicted trajectory set P i : S52, based on the predicted trajectory set P i The predicted location of each regional node in and the local threat score of the regional node at the current time step t Will Directly extend to each future time step t+k and set the future local threat score S53, at each time step t+k in the future, based on the predicted trajectory set P i The predicted location of each regional node in Local density feature n i and future local threat scores Fusion generates dynamic risk distribution map R (t) : Among them, R (t+k) is the dynamic risk distribution diagram for each time step t+k in the future, and R(x,y,t+k) is the dynamic risk distribution diagram R for each time step t+k in the future. (t+k) The risk intensity of each location point in , O′ is the subset of valid target objects, is the predicted trajectory set P i The predicted location of each regional node in The Euclidean distance from the reference point (x, y), α is the Euclidean distance influence weight coefficient, and γ is the local node density influence weight coefficient.

8. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The RRT path optimization method introducing dynamic risk cost specifically includes: During the extended sampling process, the extended sampling point (x s ,y s ) Limited to low-risk areas Internal generation, the low-risk domain is defined as the area that satisfies R(x, y, t) ≤ δ, where δ is the set risk threshold; In the process of path expansion, based on the dynamic risk distribution map R (t) , initialize the dynamic risk cost path expansion tree Τ, search and expand the sampling point (x s ,y s ) The node v with the closest Euclidean distance near , generating an extended path segment, the extended path segment points to the extended sampling point (x s ,y s ), and calculate v near With the new node v new The Euclidean distance d(v near ,v new ), real-time selection of dynamic evaluation benchmarks (x m ,y m ) corresponding to the risk intensity R(x m ,y m ,t), R(x m ,y m ,t) and d(v near ,v new ) to form the basic dynamic risk cost C risk (v near ,v new ): C risk (v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t)); Among them, ρ0 is the dynamic risk level weighting coefficient, which is used to regulate the response sensitivity of the extended path segment to changes in environmental risks; Based on the predicted trajectory set P i , calculate the path cost correction factor Φ(x m ,y m ,t), the path cost correction factor Φ(x m ,y m ,t) is determined by the trajectory deviation TrjBias(x m ,y m ) and future local threat scores Weighted combination: Among them, ρ1 is the trajectory deviation weighting coefficient, ρ2 is the threat level weighting coefficient; Basic dynamic risk cost C risk (v near ,v new ) and the path cost correction factor Φ(x m ,y m ,t) form the evaluation value Q(v near ,v new ): Q(v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t))×(1+Φ(x m ,y m ,t)); In the process of path extension, the evaluation value Q(v near ,v new ) minimization principle, dynamically select the expansion direction and expansion base point, update the dynamic risk cost path expansion tree Τ, and continuously optimize the path expansion process by refreshing the basic dynamic risk cost status and path cost correction factor status in real time.

9. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The S6 specifically includes: S61, based on the dynamic risk distribution map R (t) , introduce the dynamic risk cost RRT path optimization method to perform path extension search, set the current position of the unmanned vehicle (x0, y0) as the path extension search starting point, and initialize the dynamic risk cost path extension tree Τ; S62. Define low-risk domains: Among them, R(x,y,t) is the dynamic risk distribution map R of each time step t+k in the future (t+k) The risk intensity of each location point in the , δ is the set risk threshold; in the low risk domain Extended sampling is performed within the sample, and extended sampling points (x s ,y s ); S63. In the dynamic risk cost path expansion tree T, search and expand the sampling point (x s ,y s ) The node v with the closest Euclidean distance near , and v near As the extended base point, an extended path segment is generated, and the extended path segment points to the extended sampling point (x s ,y s ), select the middle reference point (x m ,y m ) as a dynamic evaluation benchmark, and the target object's future motion trajectory and threat level as path cost correction factors, and perform real-time evaluation: Q(v near ,v new )=d(v near ,v new )×(1+ρ0R(x m ,y m ,t))×(1+Φ(x m ,y m ,t)); Among them, Q(v near ,v new ) is the evaluation value of the extended path segment, v new is a new node, d(v near ,v new ) is the Euclidean distance between two nodes, ρ0 is the risk level weighting coefficient, R(x m ,y m ,t) is the middle reference point (x m ,y m ) in the dynamic risk distribution map R (t) The corresponding risk intensity, path cost correction factor Φ(x m ,y m ,t) is defined as: Among them, ρ1 is the trajectory deviation weighting coefficient, TrjBias(x m ,y m ) is (x m ,y m ) and the predicted trajectory set P i The deviation between the predicted positions of the nodes in each region, ρ2 is the threat level weighting coefficient, To predict the trajectory set P at each time step t+k in the future i The future local threat score of each regional node in the; S64, based on the evaluation value Q(v near ,v new ), select the new node v with the best performance in the current extended path segment new , update the dynamic risk cost path expansion tree Τ.

10. The intelligent identification system for autonomous object detection and obstacle avoidance based on the unmanned vehicle's azimuth end according to claim 2 is characterized in that: The S7 specifically includes: During the path expansion search process, the accumulated cost value of each sampling node is updated in real time, and the expansion direction is dynamically selected according to the minimum cost principle to generate the optimal safe driving path in a dynamic environment. S71, the new node v with the best performance in the current extended path segment new As the new extension base point, based on the evaluation value Q(v near ,v new ), calculate the path extension search starting point v start to v new The cumulative path cost C total (v new ): Among them, P(v start ,v new ) is v start to v new The set of path segments; S72. During the path expansion process, the cumulative path cost value C of each expansion base point in the dynamic risk cost path expansion tree T is updated in real time. total (v new ), and with the current expansion base point v near For reference, from v near Filter the expansion base point v with the minimum cumulative path cost value from the set of adjacent expandable nodes best , as the next path expansion direction; S73, compare v new The cumulative path cost C total (v new ) and the sum of the cumulative path costs of the candidate parent node to, if there is a candidate parent node v alt So that through v alt The cumulative path cost value is smaller, then update v new , optimizes the connection relationship of the dynamic risk cost path extension tree Τ; S74, when the path extension reaches the end node v global Or when the extension termination condition is met, from the end node v global Start by tracing back along the parent node to the path extension search starting point v start , connect each node in turn to generate the optimal safe driving path based on the minimum cumulative path cost in a dynamic environment.

Images