Unmanned vehicle dynamic path optimization method and system based on artificial intelligence

By constructing an abnormal fluctuation rendering road network and a risk chain effect prediction engine, combined with user intent prediction and multi-vehicle path conflict analysis, and adopting intelligent traffic iterative game and risk buffering strategies, the path planning problem of autonomous vehicles in complex environments is solved, and efficient and safe dynamic path optimization is achieved.

CN121768227APending Publication Date: 2026-03-31JIANGXI COLLEGE OF APPLIED TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202511973179.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-25
Publication Date
2026-03-31

AI Technical Summary

Technical Problem

Existing autonomous vehicle path planning methods lack the ability to accurately perceive and respond to real-time conditions and vehicle behavior dynamics, resulting in insufficient path planning and difficulty in meeting the path optimization needs in complex and ever-changing environments. Furthermore, the lack of multi-vehicle collaboration and global optimization mechanisms leads to low vehicle driving efficiency and increased safety hazards.

Method used

By using artificial intelligence-based methods, we construct a road network rendering system for abnormal fluctuations, a risk chain effect prediction engine, a user intent prediction situation map, and a multi-vehicle path conflict probability calculation. Combined with reinforcement learning and multi-agent collaborative algorithms, we conduct intelligent traffic iterative game to generate the optimal vehicle conflict game optimization path and perform risk buffer path planning.

Benefits of technology

It enables efficient and safe path planning in complex environments, improves the dynamic response capability and passenger safety of autonomous vehicles, reduces the probability of path conflicts and the risk of emergencies, and enhances the robustness of the system and user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121768227A_ABST
    Figure CN121768227A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of path planning, in particular to an unmanned vehicle dynamic path optimization method and system based on artificial intelligence. The method comprises the following steps: collecting regional state data based on real-time navigation information, carrying out abnormal traffic flow fluctuation analysis, carrying out abnormal road section visual rendering, and constructing an abnormal fluctuation rendering road network; performing abnormal disturbance propagation chain reaction mining on the abnormal fluctuation rendering road network, performing dynamic risk effect modeling, and constructing a risk linkage effect prediction engine; performing user behavior intention deep analysis and riding intention situation prediction to obtain a user intention prediction situation map; and performing adjacent vehicle path pre-judgment based on the user intention prediction situation map, and performing path conflict probability calculation to obtain a multi-vehicle path conflict probability. According to the invention, the robustness of the unmanned driving system to deal with sudden situations and complex dynamic environments is improved, and the riding safety and the path accessibility are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and in particular to a method and system for dynamic path optimization of unmanned vehicles based on artificial intelligence. Background Technology

[0002] Autonomous vehicles utilize sensors, communication, and computing platforms to achieve real-time perception and intelligent decision-making regarding themselves and their surroundings. However, with the increasing scale of autonomous vehicles and the growing complexity of roads, achieving efficient and safe path planning has become one of the key challenges hindering the widespread application of autonomous driving technology. Traditional path planning methods largely rely on preset rules and static maps, lacking the ability to accurately perceive and respond to real-time conditions, vehicle dynamics, and unexpected events, making it difficult to meet the dynamic path optimization needs of autonomous vehicles in complex and ever-changing environments.

[0003] Existing route planning schemes often face problems such as frequent multi-vehicle route conflicts, insufficient risk prediction, and delayed response to route adjustments, leading to low vehicle driving efficiency, increased congestion, and even potential safety hazards. Furthermore, traditional planning methods are mostly based on a single-vehicle perspective, lacking multi-vehicle collaboration and global optimization mechanisms, making it difficult to effectively coordinate dynamic interactions between vehicles and reducing the overall system's intelligence level. Existing route optimization technologies largely rely on rule-driven or simple heuristic algorithms, failing to fully mine and utilize the potential information in complex data, and struggling to achieve real-time dynamic adjustments and risk control. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a dynamic path optimization method and system for autonomous vehicles based on artificial intelligence, thereby resolving at least one of the aforementioned technical issues.

[0005] To achieve the above objectives, this invention provides a dynamic path optimization method for autonomous vehicles based on artificial intelligence, comprising the following steps:

[0006] Step S1: Based on real-time navigation information, collect regional status data, perform abnormal traffic flow fluctuation analysis, and visualize and render abnormal road sections to construct an abnormal fluctuation rendering road network.

[0007] Step S2: Min the chain reaction of abnormal disturbance propagation in the abnormal fluctuation rendering road network, and model the dynamic risk effect to build a risk chain effect prediction engine.

[0008] Step S3: Obtain user riding behavior perception information, perform in-depth analysis of user behavior intent and riding intent situation prediction to obtain user intent prediction situation map;

[0009] Step S4: Identify other vehicle perception information in the real-time road network of the area, predict the paths of nearby vehicles based on the user intent prediction situation map, and calculate the probability of path conflict to obtain the probability of multi-vehicle path conflict.

[0010] Step S5: Perform conflict optimization analysis based on the multi-vehicle path conflict probability, and conduct intelligent traffic iterative game to extract the optimal vehicle conflict game optimization path;

[0011] Step S6: Based on the risk chain effect prediction engine, predict the impact of sudden events and plan the risk buffer path for the optimal vehicle conflict game optimization path to perform vehicle dynamic path optimization.

[0012] This specification provides an artificial intelligence-based dynamic path optimization system for autonomous vehicles, used to execute the artificial intelligence-based dynamic path optimization method for autonomous vehicles as described above, including:

[0013] The abnormal fluctuation module is used to collect regional status data based on real-time navigation information, perform abnormal traffic flow fluctuation analysis, visualize and render abnormal road sections, and construct an abnormal fluctuation rendering road network.

[0014] The chain reaction module is used to mine the chain reaction of abnormal disturbance propagation in the abnormal fluctuation rendering road network, and to model the dynamic risk effect to build a risk chain effect prediction engine.

[0015] The intent prediction module is used to acquire user riding behavior perception information, perform in-depth analysis of user behavior intent and riding intent situation prediction to obtain user intent prediction situation map.

[0016] The conflict probability calculation module is used to identify the perception information of other vehicles in the real-time road network of the area, predict the path of nearby vehicles based on the user intent prediction situation map, and calculate the path conflict probability to obtain the multi-vehicle path conflict probability.

[0017] The conflict game module is used to perform conflict optimization analysis based on the probability of multi-vehicle path conflict, and to conduct intelligent traffic iteration game to extract the optimal vehicle conflict game optimization path.

[0018] The risk buffer planning module is used to predict the impact of sudden events based on the risk chain effect prediction engine, and to plan risk buffer paths for the optimal vehicle conflict game optimization path in order to perform vehicle dynamic path optimization operations.

[0019] The beneficial effects of this invention are specifically as follows: By continuously collecting regional flow data from real-time navigation information and combining it with parameters such as sudden speed changes, abnormal traffic flow, and congestion, abnormal traffic flow fluctuations and spatiotemporal dynamic feature recognition are carried out. This enhances the autonomous driving system's environmental perception of "abnormal states," enabling rapid identification, intuitive visualization, and dynamic presentation of problem road sections, providing a higher-resolution disturbance baseline layer for path optimization. A disturbance causal chain diffusion analysis is performed on abnormal fluctuations to identify the cascading effects of "butterfly effects" such as congestion, queue backflow, and signal failure. This allows for early insight into the potential impact of local anomalies on a larger road network, moving from "point-like analysis" to "chain-like prediction," providing scientific support for proactive risk avoidance and regional path adjustment. User travel behavior perception data (such as destination frequency, preferred routes, and stopping strategies) is collected and combined with voice interaction, gesture input, or historical trajectories to construct a personalized user intent model and generate a visualized user intent prediction situation map. This achieves alignment and interactive understanding between the autonomous driving decision-making system and the user's intent, avoiding deviations in path planning where "the system is reasonable but the user is not satisfied," and improving the degree of human-machine consensus.

[0020] The system analyzes real-time perception data (position, speed, turn signal signals, etc.) of other vehicles within the area, combines this data with the user's intent situational map, predicts the possible driving paths of neighboring vehicles, and quantifies the probability of potential conflict paths with the user's own vehicle, forming a multi-vehicle path conflict probability distribution map. This enhances the dynamic game-playing ability and interactive understanding of autonomous vehicles, moving beyond reliance solely on static rules or safety distance judgments to anticipate potential interaction risks based on intent matching and behavior prediction. Based on the multi-vehicle path conflict probability, it generates conflict avoidance optimization strategies and conducts multi-round game simulations to construct an intelligent traffic game model. Through algorithms such as reinforcement learning and multi-agent collaboration, it selects the system's optimal (rather than the local shortest) traffic path and extracts the current optimal vehicle conflict game optimization path. It considers not only the current state but also the impact of future conflict risks and other vehicle behaviors, achieving a more robust and safer path selection mechanism and avoiding the risks associated with frequent lane changes and sudden braking. Combining the aforementioned risk chain effect prediction engine, this system predicts unforeseen events (such as accidents, road closures, and emergency traffic diversions) on the optimal path in the game theory framework. A risk buffer path planning strategy with safety redundancy is designed and ultimately used for vehicle dynamic path optimization and real-time deployment adjustments. This constructs a path planning mechanism with self-healing and adaptive capabilities, significantly improving the robustness of the autonomous driving system in dealing with sudden situations and complex dynamic environments, and enhancing passenger safety and path accessibility. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the steps of an artificial intelligence-based dynamic path optimization method for autonomous vehicles according to the present invention.

[0022] Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1.

[0023] Figure 3 This is a detailed flowchart illustrating the implementation steps of step S2;

[0024] Figure 4 This is a flowchart illustrating the detailed implementation steps of step S3. Detailed Implementation

[0025] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0026] This application provides a method and system for dynamic path optimization of autonomous vehicles based on artificial intelligence. The executing entities of the method and system include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, and network upload devices that can be considered general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio-visual management system, an information management system, and a cloud data management system.

[0027] Please see Figures 1 to 4 This invention provides a dynamic path optimization method for autonomous vehicles based on artificial intelligence, comprising the following steps:

[0028] Step S1: Based on real-time navigation information, collect regional status data, perform abnormal traffic flow fluctuation analysis, and visualize and render abnormal road sections to construct an abnormal fluctuation rendering road network.

[0029] Step S2: Min the chain reaction of abnormal disturbance propagation in the abnormal fluctuation rendering road network, and model the dynamic risk effect to build a risk chain effect prediction engine.

[0030] Step S3: Obtain user riding behavior perception information, perform in-depth analysis of user behavior intent and riding intent situation prediction to obtain user intent prediction situation map;

[0031] Step S4: Identify other vehicle perception information in the real-time road network of the area, predict the paths of nearby vehicles based on the user intent prediction situation map, and calculate the probability of path conflict to obtain the probability of multi-vehicle path conflict.

[0032] Step S5: Perform conflict optimization analysis based on the multi-vehicle path conflict probability, and conduct intelligent traffic iterative game to extract the optimal vehicle conflict game optimization path;

[0033] Step S6: Based on the risk chain effect prediction engine, predict the impact of sudden events and plan the risk buffer path for the optimal vehicle conflict game optimization path to perform vehicle dynamic path optimization.

[0034] In the embodiments of the present invention, see Figure 1 This is a flowchart illustrating the steps of an AI-based dynamic path optimization method for autonomous vehicles according to the present invention. In this example, the steps of the AI-based dynamic path optimization method for autonomous vehicles include:

[0035] Step S1: Based on real-time navigation information, collect regional status data, perform abnormal traffic flow fluctuation analysis, and visualize and render abnormal road sections to construct an abnormal fluctuation rendering road network.

[0036] In this embodiment, key indicators such as traffic flow, average vehicle speed, and road occupancy are acquired within a 5-kilometer radius using multi-source data interfaces, including floating car data from the city management center, roadside sensor networks, and the Gaode Map API, with a sampling frequency of 30 seconds. The system establishes a 15-dimensional feature vector including traffic density, speed variance, and traffic flow mutation rate. An anomaly detection mechanism is triggered when the traffic density exceeds 0.8 (saturation flow ratio) or the speed variance is greater than twice the standard deviation. Anomaly traffic flow fluctuation analysis employs a sliding window algorithm, setting a time window of 5 minutes and a spatial window of 500 meters. The Z-score anomaly detection index is calculated to identify states deviating from the normal distribution by more than 3 standard deviations. The system uses color-coding visualization technology, marking normal road sections in green (congestion index 0-0.3), slightly abnormal road sections in yellow (congestion index 0.3-0.6), and severely abnormal road sections in red (congestion index 0.6-1.0), constructing a real-time updated anomaly fluctuation rendering road network. This rendered road network integrates a GIS geographic information system, supports multi-level zoom display, and provides basic data for visualized abnormal states for subsequent steps.

[0037] Step S2: Min the chain reaction of abnormal disturbance propagation in the abnormal fluctuation rendering road network, and model the dynamic risk effect to build a risk chain effect prediction engine.

[0038] In this embodiment, a directed graph model is constructed, with each road segment as a node and the connectivity between road segments as the edge weight. The propagation attenuation coefficient α = 0.85 and the distance attenuation parameter β = 0.02 / km are set. When the system identifies critical propagation paths, the influence propagation intensity of each path is calculated as I = I0 × α^t × e^(-β × d), where I0 is the initial disturbance intensity, t is the propagation time step, and d is the propagation distance. Dynamic risk effect modeling adopts the Markov chain Monte Carlo method, establishing a state transition matrix to describe the evolution of abnormal states in the time dimension. Paths with a state transition probability threshold of 0.7 or higher are set as high-risk propagation chains. The system is trained using historical data and finds that when the congestion index of the upstream road segment exceeds 0.8, the probability of congestion occurring within 500 meters downstream reaches 0.76 within 10 minutes and rises to 0.89 within 15 minutes. The risk chain effect prediction engine integrates a deep learning LSTM network with 128 hidden layer nodes and a learning rate of 0.001. It can predict the propagation range and impact intensity of abnormal disturbances within the next 20 minutes, providing a forward-looking risk assessment for path planning.

[0039] Step S3: Obtain user riding behavior perception information, perform in-depth analysis of user behavior intent and riding intent situation prediction to obtain user intent prediction situation map;

[0040] In this embodiment, an infrared camera captures changes in passenger facial expressions at 30fps, extracting micro-expression feature vectors including 42 dimensions such as eye movement units AU01-AU07 and mouth movement units AU12-AU27. A pressure-sensing seat monitors the frequency of passenger posture changes; under normal conditions, the posture adjustment frequency is 2-3 times / minute, increasing to 8-12 times / minute in emergency situations. A speech recognition system analyzes passenger dialogue, extracting emotional inclination markers such as keywords like "anxious," "hurry up," and "not in a hurry." Deep analysis of user behavior intent employs a multilayer perceptron (MLP) network. The input layer receives 67 dimensions of data, including physiological, behavioral, and speech features. The hidden layers consist of three layers with 256, 128, and 64 neurons respectively. The output layer generates three intent classification probabilities: "time-sensitive," "comfort-first," and "cost-first." The ride intention situation prediction uses a time series analysis algorithm to establish a dynamic model of intention changes. When the probability of detecting a time-sensitive intention exceeds 0.8, the system adjusts the route planning weight to a configuration scheme of 70% for optimal time, 20% for comfort, and 10% for cost.

[0041] Step S4: Identify other vehicle perception information in the real-time road network of the area, predict the paths of nearby vehicles based on the user intent prediction situation map, and calculate the probability of path conflict to obtain the probability of multi-vehicle path conflict.

[0042] In this embodiment, a 360-degree perception system consisting of an onboard LiDAR, millimeter-wave radar, and a high-definition camera scans participants within a 200-meter radius at a frequency of 20Hz. LiDAR point cloud data processing employs a PointNet++ neural network to achieve accurate classification and localization of targets such as vehicles, pedestrians, and bicycles, achieving a detection accuracy of 95.2% and a localization error controlled within ±0.15 meters. The system predicts the trajectories of other vehicles using a Kalman filter algorithm, setting a prediction time window of 8 seconds. It calculates the confidence level of future trajectories by combining the vehicle's current position, speed, acceleration, and other motion state parameters. When predicting the paths of nearby vehicles based on the user intent prediction situation map, the system establishes a multi-vehicle interaction model, considering the mutual influence between vehicles. When a vehicle ahead is detected to be decelerating at a rate exceeding -2 m / s², the system predicts that the probability of that vehicle changing lanes or stopping is 0.73. The path conflict probability calculation adopts the spatiotemporal occupancy grid method, which divides the road space into 2m×2m grid cells and calculates the probability that different vehicles will occupy the same grid cell in the future. When the overlap probability exceeds 0.4, it is considered that there is a potential conflict risk, and when it exceeds 0.7, the emergency avoidance mechanism is triggered.

[0043] Step S5: Perform conflict optimization analysis based on the multi-vehicle path conflict probability, and conduct intelligent traffic iterative game to extract the optimal vehicle conflict game optimization path;

[0044] In this embodiment, the system employs the multi-objective optimization algorithm NSGA-III to handle complex conflict resolution issues. The conflict optimization objective function design includes three dimensions: safety index S, efficiency index E, and comfort index C, with safety weighted at 0.6, efficiency weighted at 0.3, and comfort weighted at 0.1. The intelligent passage iterative game uses the evolutionary stable strategy (ESS) theory, treating each vehicle as a rational decision-maker and searching for a Nash equilibrium solution in the game matrix. The system sets the game payoff function as R = αT + βS + γF, where T is the travel time payoff, S is the safe distance payoff, and F is the fuel consumption payoff, with weight coefficients α = 0.5, β = 0.3, and γ = 0.2. The iterative game process uses the reinforcement learning Q-learning algorithm with a learning rate of 0.1 and a discount factor of 0.9, achieving convergence after an average of 150 iterations. The optimal vehicle conflict game optimization path is extracted through Pareto optimality screening; when any solution in the solution set is not inferior to other solutions in all objectives, that solution is selected as a candidate optimal solution. Experimental data show that this game-theoretic optimization method can reduce the probability of multi-vehicle conflicts from the original 0.23 to 0.07, while ensuring an average traffic efficiency increase of 15.8%, providing a safe and efficient conflict solution for autonomous vehicles.

[0045] Step S6: Based on the risk chain effect prediction engine, predict the impact of sudden events and plan the risk buffer path for the optimal vehicle conflict game optimization path to perform vehicle dynamic path optimization.

[0046] In this embodiment, the system establishes an impact assessment model for 12 types of emergencies, including accidents, road construction, severe weather, and large-scale events. Different impact radius and duration parameters are set for each type of event; for example, a minor accident has an impact radius of 800 meters and a duration of 45 minutes, while a major accident has an impact radius of 2500 meters and a duration of 180 minutes. The system integrates multi-source information, including accident reports from traffic police, weather warnings from meteorological departments, and construction announcements from municipal departments, to establish an emergency event probability prediction model. When the predicted probability exceeds 0.6, emergency route planning is initiated. Risk buffer route planning uses an improved version of the dynamic A* algorithm, adding risk cost and time window constraints to the traditional route cost. The system sets a risk buffer coefficient k = 1.2-1.8, dynamically adjusting it according to the risk level, and simultaneously establishes multiple alternative routes to ensure system robustness. When the risk index of the main route exceeds the threshold of 0.75, the system automatically switches to the alternative route with the lowest risk index. The execution of vehicle dynamic route optimization adopts a hierarchical control architecture, with the upper layer responsible for global route planning and risk assessment, and the lower layer responsible for local route adjustment and real-time obstacle avoidance. The system exchanges information with surrounding vehicles and facilities through the vehicle-to-everything (V2X) communication protocol, with data transmission latency controlled within 50ms, ensuring the real-time and accuracy of path optimization decisions, and ultimately achieving the goal of dynamic path optimization for autonomous vehicles based on artificial intelligence.

[0047] In this embodiment, see Figure 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include:

[0048] The system collects regional status data based on real-time navigation information; it then performs anomaly cleaning on the regional status data to obtain optimized regional status data.

[0049] Set the evolution time range, and identify road segments within the optimized area status data, marking multiple road segments within the area;

[0050] Road connectivity is identified for multiple regional road segments, and road topology correlation fitting is performed to construct a real-time regional road network;

[0051] Multi-dimensional feature mining is performed on a segment-by-segment basis on the real-time road network in the region to generate a multi-dimensional feature stream for each segment;

[0052] Analyze the flow change patterns of the multi-dimensional feature flow to generate a road segment status map;

[0053] Based on the road segment status map, abnormal traffic flow fluctuation analysis is performed, and abnormal road segments are visualized and rendered to construct an abnormal fluctuation rendering road network.

[0054] In this embodiment, data is collected through a multi-source real-time navigation system, primarily including the Gaode Map API, Baidu Map API, V2X communication module, vehicle GPS, geomagnetic detector, and road video recognition equipment. This data covers key indicators such as vehicle speed, traffic flow, lane occupancy rate, congestion index, events, and travel time. The data collection frequency is set to once every 10 seconds. The experimental area covers major urban roads and secondary roads, spanning approximately 25 square kilometers, involving about 1800 road segments and 500 intersections. To improve data integrity and stability, an edge computing architecture is introduced, deploying some data preprocessing modules in the intersection controller to prioritize filtering invalid and empty data packets. Furthermore, the data collection process incorporates historical traffic distribution data for each road segment, marking abnormal fluctuations in real time to lay the foundation for the next stage of data cleaning. The entire process ensures that information flow is perceived by the system with millisecond-level latency, effectively meeting the high-timeliness requirements of autonomous driving path decision-making. For missing values ​​caused by equipment failure, communication interruption, or GPS drift, a sliding time window interpolation strategy (window size 30 seconds) and a substitution strategy based on the historical average of neighboring road segments are used to fill in the gaps. Secondly, a statistical anomaly detection model (Z-score method) is used to identify data points deviating from the mean by more than three standard deviations; data with logical contradictions (such as vehicle speed of 0 but high traffic volume) are identified and removed using a rule engine. Furthermore, an outlier confidence model is constructed based on the frequency of anomaly occurrences to completely mask data from frequently malfunctioning equipment. The entire cleaning process adopts a batch-stream integrated data stream processing architecture (based on Apache Flink), and in experiments comparing the original and cleaned data, the accuracy improved by approximately 9.6%, and data stability improved by 13.8%. This process not only ensures the accurate reflection of state data but also provides a more solid data foundation for subsequent pattern mining.

[0055] The analysis timeframe was set to 30 minutes, employing a sliding time window mechanism (5-minute window cycle) to dynamically update the status and capture drastic changes in traffic flow characteristics within a short period. Spatially, by combining optimized GPS trajectory data with high-precision electronic maps (such as OpenStreetMap or road network layers provided by the city management center), an improved HMM map matching algorithm was used to correct the location of all vehicle trajectories and identify road segments. The map matching accuracy reached over 96%. Subsequently, based on the identification results, the system uniquely numbered and labeled the roads within the area, including road type (urban arterial road, secondary arterial road, etc.), speed limit, direction, number of lanes, and signal type. Approximately 1800 active road segments were identified within the experimental area, and about 350 road segments with significant status fluctuations were selected as the objects for subsequent in-depth analysis. Through this step, the system successfully established an evolutionary time-space analysis model, providing accurate modeling units for dynamic mapping and path optimization.

[0056] The system structurally integrates marked independent road segments to construct a topological network that reflects real-world road traffic logic. First, using intersections as nodes and road segments as edges, a preliminary connected graph is built. Next, the system analyzes the turning relationships and traffic rules (including no-left-turn, restricted sections, U-turn restrictions, etc.) between each edge and adjacent nodes to construct a directional graph. To improve connectivity modeling accuracy, an edge prediction model from graph neural networks is introduced. The potential connection probabilities between road segments are learned through historical traffic trajectories (5 million sampled vehicle trajectory sequences), and topology fitting and fine-tuning are performed based on connection weights. Then, graph embedding technology maps structural features to a vector space, optimizing the representation of connection strength between roads. In the experimental scenario, a road topology graph with nearly 3000 directed edges and 800 nodes is generated, with a graph density of approximately 0.0024 and an average node degree of 3.7. This real-time road network can be automatically reconstructed as the state is updated, serving as the core structural foundation for autonomous vehicle path reasoning algorithms and dynamic obstacle avoidance decisions. Deep feature mining is performed on each road segment to generate a multi-dimensional feature stream that can be used for behavior prediction and route planning. Feature dimensions include: real-time average vehicle speed (km / h), number of vehicles per unit time (vehicles / min), vehicle spacing change rate, occupancy rate (%), delay index, congestion level (levels: 0-5), and wave index. Data is aggregated in 5-minute intervals to construct a time-series feature chain (i.e., feature stream) to describe the continuous evolution trend of road conditions. The system uses recurrent neural networks (such as GRU) to model the historical data of each road segment and extracts key change patterns through feature embedding. In the experimental setup, seven main dimensions were constructed for each road segment, generating over 2 million data points daily. The feature stream is further normalized and standardized, converted into a tensor sequence with a unified input format for easy use by downstream analysis models. Finally, a complete "regional feature stream database" is formed, supporting multiple key modules such as subsequent prediction, route risk assessment, and anomaly identification.

[0057] For each road segment, a time window trend extraction algorithm (using a sliding window + Savitzky-Golay filter) is applied to extract indicators such as rate of change, acceleration, and fluctuation amplitude. Then, unsupervised learning clustering (using K-means++ and DBSCAN joint modeling) is used to group similar changing features into typical patterns, such as stable smooth flow, periodic congestion, sudden congestion, gradual deterioration, and rapid recovery. For each state, its spatial distribution and temporal evolution trend are further analyzed, and the clustering results are mapped back onto the road topology map to form a state map. In the experimental scenario, the clustering model was used to model the features of the past 30 days, achieving a model accuracy of 93.4%, with an average daily map update frequency of 144 times (once every 10 minutes). The final generated map has three capabilities: spatial visualization, semantic interpretation, and structural analysis, serving as a foundational reference layer for constructing more advanced intelligent scheduling and game-theoretic path optimization. The CUSUM (Cumulative Sum Control Chart) and BOS (Bayesian Online Segmentation) algorithms are introduced to detect sudden fluctuations, primarily identifying short-term changes such as sudden drops in vehicle speed, sudden increases in traffic flow, and surges in delays. The detection sensitivity is controlled within the range of 0.85 to 0.95, and the average response time for anomaly identification is less than 12 seconds. Then, an anomaly scoring mechanism (combining change rate, spatial density, and duration) is constructed to quantitatively score and classify each anomalous road segment (mild, moderate, severe). Simultaneously, anomaly information is mapped to the topology network diagram in real time and rendered visually using color, line width, and flashing frequency, forming an "anomaly fluctuation rendering road network." In experimental demonstrations, the system was validated against anomaly events in a district of Beijing, successfully identifying three sudden congestion events during the morning rush hour on June 14, 2025, with an accuracy rate of 94.1%. The final result provides clear risk avoidance zone identification and path avoidance instructions for autonomous driving systems, serving as a key supporting module for dynamic risk warning and intelligent path planning.

[0058] In this embodiment, the specific steps for performing abnormal traffic flow fluctuation analysis based on road segment state maps, visualizing and rendering abnormal road segments, and constructing an abnormal fluctuation rendering road network are as follows:

[0059] Real-time traffic flow for each road segment is calculated based on the road segment status map;

[0060] The real-time traffic flow is analyzed to identify the correlation between road segments, so as to obtain the traffic flow correlation features between road segments;

[0061] The real-time traffic flow is calculated for multi-time-point traffic flow fluctuations to generate a traffic flow fluctuation curve;

[0062] Based on the traffic flow correlation characteristics, the traffic flow fluctuation curve is used to perform in-depth identification of key road segment intersections in the road network, and key road segment intersections are marked.

[0063] The probability of abnormal fluctuations at different scales is calculated for key road intersections, and the degree of impact of abnormal traffic flow is calculated to obtain anomaly indexes for multiple key road intersections.

[0064] Based on the aforementioned anomaly level index, the abnormal fluctuations of the real-time road network in the region are visualized and rendered to construct an abnormal fluctuation rendering road network.

[0065] In this embodiment, based on the constructed road segment status map, various features are integrated to extract and quantify the real-time traffic flow index for each road. Traffic flow is a key parameter for measuring road capacity and load, defined as the number of vehicles passing through a certain cross-section per unit time. The system adopts an estimation model based on multi-sensor fusion, jointly modeling data such as V2X backhaul data, roadside perception camera statistics, GPS trajectory point counts, and signal cycle traffic estimation. Specifically, vehicle-mounted GPS data uses a clustering-based trajectory filtering method to retain only trusted terminals, and camera count data uses the YOLOv5 model for vehicle type identification and classification statistics. A sliding time window (60 seconds window width, 5-second update cycle) is used to perform real-time alignment and weighted fusion of various data types to construct a standardized real-time traffic flow sequence. In the experimental area, approximately 1800 roads were tested, with an average data processing volume exceeding 200,000 data points per minute. This method can control the instantaneous traffic flow calculation error within ±6 vehicles / minute, and the traffic flow estimation accuracy can reach 94.7% during peak periods. The calculation results will serve as the basic input data for subsequent fluctuation and correlation modeling. Because traffic flow in the road network exhibits continuity, directionality, and time lag, the traffic flow of adjacent road segments should demonstrate a certain degree of statistical correlation. This step employs the Granger Causality Test and Dynamic Time Warping (DTW) to assess the strength of traffic flow association between road segments. By analyzing historical traffic flow trends over 5 to 15 minutes, the system performs causal modeling of traffic propagation along three paths: upstream → current → downstream. Simultaneously, to account for differences in road turning behavior, a turning probability matrix (derived from historical traffic trajectories) is introduced to correct and estimate traffic flow at divergence points. Within the experimental area, the system identified approximately 3600 pairs of highly coupled road segments, with 48.3% exhibiting a Pearson correlation coefficient greater than 0.7. The output is a "traffic correlation matrix" between each pair of road segments, used to characterize the traffic propagation path within the structure, providing support for key intersection identification and anomaly propagation modeling.

[0066] Using a 5-minute time granularity, multiple sliding windows (5 minutes, 15 minutes, and 30 minutes) were set to sample, smooth, and differralize the traffic flow sequence for each road, extracting short-term fluctuation trends in traffic flow. During modeling, statistical methods such as coefficient of variation (CV), skewness, standard deviation, and first-order difference sequence analysis were employed to characterize traffic flow stability and volatility. Furthermore, to enhance the responsiveness of the fluctuation curve to abnormal signals, the system integrated wavelet transform (DWT) for multi-scale frequency domain analysis of time-series data, effectively identifying sudden changes within short periods. The fluctuation curve is output as a two-dimensional time-flow curve, which can be further analyzed for slope and trend clustering. Experiments showed that the average fluctuation amplitude of traffic flow during peak hours was 2.3 times that during off-peak hours, and the average detection delay of abrupt changes within a 15-minute time window did not exceed 12 seconds. The generated fluctuation curve not only reflects the rhythm of road load changes but can also be used as a prerequisite for sudden event analysis and key road segment identification. The hub nodes have the most significant impact on overall stability. This step employs an "influence propagation model" based on graph neural networks (GNNs). Using a topological graph as a foundation, it comprehensively evaluates nodes (intersections) across multiple dimensions, including their position within the network, traffic flow stability, traffic concentration, and downstream influence radius. The modeling process calculates the betweenness centrality, traffic sensitivity (reflected by the probability of traffic mutation propagation), and potential ability of each node to cause network instability. In an experiment, a model was created for a district in Beijing, identifying the top 10 intersections with the highest impact scores during evening rush hour. Traffic changes at these intersections could affect at least four adjacent intersections within an average of 6.2 minutes. The system uses these results to label key intersections and construct an "influence path map," marking potential high-risk propagation sources. This identification mechanism provides crucial decision-making support for autonomous driving systems, enabling them to adjust routes and avoid risks before they spread.

[0067] Multi-scale anomaly probability modeling was performed on key nodes, including short-term (0-5 minutes), medium-term (5-15 minutes), and long-term (15-30 minutes) models, using Hidden Markov Models (HMMs) and Bayesian anomaly inference methods, respectively. The input for each scale model consisted of past period traffic flow fluctuation characteristics, upstream and downstream coupling levels, and records of sudden events (such as temporary road closures and traffic light malfunctions). Secondly, the system introduced an "Anomaly Impact Index" (AII), calculated based on the spatial range (influence radius) and duration of the impact that can be transmitted under anomaly conditions at each node. For example, if an intersection affects three downstream road segments within 5 minutes, with each segment experiencing a traffic decrease of more than 20%, its AII value will be assessed as high-level. In the experimental scenario, anomaly severity indices for 120 key nodes were calculated and normalized. The results showed that nodes with high impact indices are often located in structurally complex areas such as transfer hubs and ramp merging points. This index system provides a quantitative input basis for subsequent route avoidance planning.

[0068] After completing the anomaly modeling and indicator evaluation of key nodes, the system finally performs visualization anomaly rendering on the real-time road network of the region, constructing a dynamic anomaly fluctuation layer. This layer uses the road topology as the base map and the node anomaly degree index as the superimposed variable, expressing multiple dimensions through color depth, line width, and flashing frequency. For example, road segments with an anomaly impact index (AII) above 0.8 are highlighted in red, and their downstream impact range is rendered, with fluctuation trends represented by dynamic lines. The system supports three-level view switching: micro (single road segment), meso (intersection group), and macro (regional road network), facilitating multi-level linkage analysis. During the visualization process, WebGL technology and a GIS map engine are integrated to ensure smooth rendering even in high-density road network environments. In experimental deployment, the rendering layer was successfully constructed in the Beijing CBD test area, with an average system response latency of 410ms and a dynamic refresh rate of 1 time / second. The final generated rendered road network not only provides a high-risk area early warning map for autonomous driving systems but also serves as a visualization input platform for path game theory and risk avoidance algorithms.

[0069] In this embodiment, see Figure 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include:

[0070] The abnormal fluctuations in the rendered road network are identified to identify the abnormal fluctuation propagation links and extract the evolution and propagation paths of abnormal fluctuations.

[0071] The abnormal fluctuation evolution and propagation path is subjected to disturbance propagation chain reaction mining to generate road network disturbance propagation chain reaction characteristics.

[0072] Based on the chain reaction characteristics of road network disturbance propagation, potential risk prediction is performed to obtain road network potential risk prediction data.

[0073] The chain reaction intensity is assessed based on the road network potential risk prediction data to obtain the risk chain intensity assessment value;

[0074] The disturbance amplification factor is calculated based on the risk chain intensity assessment value to generate the disturbance amplification factor for each road segment;

[0075] Based on the disturbance amplification coefficient and the abnormal fluctuation evolution and propagation path, dynamic risk effect modeling is performed, and a risk chain effect prediction engine is constructed.

[0076] In this embodiment, the propagation path of abnormal fluctuations in the rendered road network is identified, clarifying the evolution path of abnormal events from their origin to their affected terminals. First, the abnormal fluctuation rendered road network is abstracted as a directed graph structure, with nodes corresponding to road intersections or road segments, and edges representing flow directions. Using time-series anomaly detection results, combined with the intensity of abnormal traffic flow fluctuations and the propagation time window (set to 5 minutes, with a 1-minute step), potential adjacent edges for abnormal propagation are screened. Specifically, this involves calculating the time lag correlation of abnormal states between upstream and downstream nodes and using a cross-correlation function to evaluate the synchronicity and transmissibility of abnormal fluctuations between two road segments. Subsequently, a graph traversal algorithm (DFS) is applied to extend the path along the propagation trend, extracting the full-link path from the anomaly origin downstream. A method based on Temporal Graph Neural Network (TGNN) is used, combining the time series of traffic fluctuations in road segments and anomaly event markers, to model the propagation mechanism of disturbances along the path. TGNN can capture spatiotemporal dependencies, revealing the evolution law of the impact intensity of disturbances from their source to their terminal. In practice, the input consists of multi-dimensional feature flows (flow rate, speed, headway, etc.) and abnormal fluctuation indicators along the path, while the output is the disturbance propagation weight and response delay for each link in the path. The system optimizes model parameters by training on historical sudden congestion event data (including a peak-hour dataset from Beijing from 2019 to 2024, covering over 5000 abnormal propagation paths) to accurately reconstruct the disturbance response chain. The generated chain reaction features include disturbance intensity gradient, propagation efficiency, and decay rate, providing quantitative indicators for subsequent risk prediction. Furthermore, the model can be updated in real time, supporting dynamic capture of road condition disturbance evolution.

[0077] By constructing a deep learning model based on spatiotemporal features (such as a Spatio-Temporal Attention Network), the system integrates disturbance intensity, propagation path, and external factors (weather, event announcements, etc.) to predict risks in the short term (5-30 minutes). The model input is a chain reaction feature sequence, and the output is the potential risk probability distribution for each road segment, reflecting its likelihood of experiencing congestion or accidents. During the prediction process, the model utilizes a multi-level attention mechanism to strengthen the impact assessment of critical paths and important nodes. By comparing with historical actual accidents and congestion records, the model achieved an average prediction accuracy of 89.4% and a recall rate of 86.1% on the test dataset. The generated potential risk prediction data provides a basis for prediction for the autonomous driving path optimization system, enabling early warning and dynamic path adjustment. A quantitative assessment of the risk chain reaction intensity is performed on the potential risk prediction results. By defining a risk propagation model, the risk probability of each road segment is weighted and superimposed with the risk accumulation relationship of other road segments along its propagation path to calculate a comprehensive risk chain intensity assessment value. The evaluation method is based on risk diffusion theory, combining the propagation matrix and risk threshold function from graph theory to achieve cascaded risk calculation among multiple nodes. Specifically, the system uses a Spectral Graph Convolutional Network to simulate the risk propagation path and calculate the "weighted risk influence" of each node in the risk propagation chain. In the experiment, the distribution of risk chain strength effectively distinguished between high-risk clusters and low-risk areas, with the average chain strength value between 0 and 1. Road segments exceeding 0.7 were defined as risk outbreak zones. This evaluation provides a quantitative reference for risk avoidance in path planning, ensuring that autonomous vehicles avoid high-chain risk areas.

[0078] The disturbance amplification factor is an extension of the risk cascading strength, comprehensively calculated by combining the road segment's capacity, historical fluctuation amplitude, and network location (such as hub status). Specifically, a multi-factor weighted model is used, with factors including risk cascading strength (40%), traffic saturation (30%), historical abnormal event frequency (20%), and road network centrality index (10%). This model normalizes the amplification factor to a uniform range of 0-2; a higher coefficient indicates a higher sensitivity of the road segment to disturbances and a greater likelihood of triggering cascading congestion. Data from the Beijing test area shows that the average amplification factor for hub road segments is 1.62, while for non-hub road segments it is 0.78. This factor is used to dynamically adjust path planning strategies, enabling autonomous driving systems to prioritize avoiding road segments with high disturbance amplification risk, improving safety and traffic efficiency. By combining the disturbance amplification factor with the abnormal fluctuation evolution and propagation path, a comprehensive dynamic risk effect modeling framework is constructed—the risk cascading effect prediction engine. This engine takes multi-source data as input and integrates graph neural networks (GNNs), time-series prediction models, and multi-factor interaction analysis modules to achieve multi-dimensional and dynamic prediction of future road network risks. The engine dynamically simulates the propagation process of abnormal events along their propagation paths and adjusts the risk transmission intensity based on the amplification factor, updating the risk status in real time. The model has real-time online learning capabilities, enabling it to adapt to environmental changes and unexpected event feedback. Experimental results show that this prediction engine effectively reduced congestion events by 15.3% and improved vehicle traffic efficiency by 11.7% in autonomous vehicle path planning. The engine provides powerful risk warning and path optimization decision support for autonomous driving systems and is a core module for achieving intelligent safety management and dynamic path adjustment.

[0079] In this embodiment, reference Figure 4 The diagram below illustrates the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include:

[0080] Real-time perception of vehicle passenger behavior is achieved based on the vehicle-mounted perception system to obtain user passenger behavior perception information.

[0081] Deep analysis of user behavior intent is performed on perceived user riding behavior information to extract deep features of user behavior intent;

[0082] Based on the deep features of user behavior intent, we calculate driving urgency, passenger comfort preferences, and travel purpose to obtain user behavior demand features;

[0083] Acquire real-time in-vehicle interactive voice, perform deep speech semantic understanding and sentiment recognition, and generate latent demand mining features;

[0084] Based on user behavior and latent needs characteristics, ride intention patterns are predicted to obtain a user intention prediction pattern map.

[0085] In this embodiment, the in-vehicle perception system includes a high-definition camera, infrared sensors, and an inertial measurement unit (IMU). It utilizes multi-sensor fusion technology to dynamically capture passenger movements, postures, and facial expressions. Specifically, images captured by the camera are input into a behavior recognition model based on a convolutional neural network (CNN) (such as 3D-CNN or spatiotemporal convolutional networks) to achieve real-time classification of passenger action categories (such as sitting, turning the head, and gestures), with an accuracy rate exceeding 92%. Simultaneously, the infrared sensors detect changes in the passenger's physiological state (such as body temperature and respiratory rate), and the inertial measurement unit monitors seat micro-vibrations to assist in identifying seating comfort and tension levels. Sensor data is fused in a time synchronization module to form a unified behavior perception information stream. Experiments verify that in typical urban road driving scenarios, the system can capture 98% of passenger behavior changes in real time, with data latency controlled within 100ms, meeting the timely response requirements of dynamic route optimization to user status. After obtaining the passenger behavior perception information, the system uses deep learning methods to analyze the behavioral intent. A temporal model based on Long Short-Term Memory (LSTM) networks and attention mechanisms is employed, combined with Graph Convolutional Networks (GCNs) to encode the spatial structure of behavior and analyze the underlying intentions of passenger behavior. Specifically, behavioral data is input as a time series into the LSTM network to capture the evolution trend and temporal dependencies of behavior, while the attention mechanism enhances the expressive power of behavioral features at key moments. GCNs are used to analyze the spatial dependencies between movements of different parts of the passenger's body, improving the semantic representation of features. Through a multi-task learning framework, the model simultaneously predicts the passenger's intention category (e.g., relaxation, anxiety, desire to change lanes, etc.) and its corresponding intensity. The model was trained based on over 100,000 real-world passenger behavior data points and, after five-fold cross-validation, achieved an intention recognition accuracy of 88.7%. The deep features extracted in this step serve as the core input for subsequent passenger demand analysis, significantly improving the depth and accuracy of user state understanding.

[0086] A multi-dimensional analysis of the deep characteristics of behavioral intentions is used to construct a user behavior demand feature model. This model combines the intensity of behavioral intentions, temporal variations, and users' personal historical ride data to quantify three key indicators: driving urgency (reflecting the user's urgency regarding travel time), passenger comfort preferences (including requirements for stability, noise, and in-vehicle environment), and travel purpose (such as business, leisure, emergency medical care, etc.). Specifically, a multilayer perceptron (MLP) is used to fuse various feature vectors, outputting continuous indicator values, all within a range of 0-1. Driving urgency is aided by detecting the frequency of passengers frequently checking their phones, exhibiting anxious actions, and using verbal commands. Comfort preferences are determined by combining seat posture stability and voice emotion data. Travel purpose is inferred using historical ride data and temporal and geographical information. The model was tested on real-world data from 2000 users, with prediction errors for driving urgency and comfort preferences controlled at 0.07 and 0.05, respectively, and a travel purpose classification accuracy of 85%. This user behavior demand feature provides multi-dimensional user profile support for the path adjustment strategy of autonomous driving systems. The system uses an in-vehicle microphone array to collect user interaction voice data in real time, performing deep semantic analysis and sentiment recognition. The voice data first undergoes endpoint detection and noise suppression, then is transcribed into text by a pre-trained end-to-end speech recognition model. Semantic understanding is performed using the Bidirectional Encoder Representation Model (BERT) in Natural Language Processing (NLP), combined with sentiment analysis models (such as a Transformer-based sentiment classification network) to identify emotional tendencies (e.g., anxiety, pleasure, dissatisfaction) in the statements. Furthermore, the system uses keyword extraction and topic modeling methods to uncover users' unexpressed but latent needs, such as route preferences and special service requests. In a laboratory setting, testing on 5000 voice samples yielded a semantic understanding accuracy of 91.2% and a sentiment tendency recognition accuracy of 88.5%. Implicit needs are represented by vectors and, together with user behavioral needs features, are input into the subsequent ride intention and posture prediction model, greatly enriching the perceived dimensions of user status.

[0087] Using the aforementioned user behavior and latent demand characteristics as input, the model predicts the trend of user intent changes within a certain timeframe. The model employs a multimodal fusion architecture, combining a Temporal Convolutional Network (TCN) and a multi-head attention mechanism to dynamically capture and predict user intent states. The output is an intent prediction state map, displaying the user's future state in dimensions such as driving urgency, comfort preference, and travel purpose in a multidimensional probability distribution. The time prediction range is generally set to the next 10 minutes, with a step size of 1 minute. Model training is based on the fusion of over 500,000 historical behavioral and speech data points. Experimental results show that the accuracy of state prediction reaches 90.5%, with prediction latency controlled within 150ms, meeting the requirements of real-time path optimization for autonomous driving. The user intent prediction state map provides dynamic feedback for vehicle path decision-making, supporting precise and personalized improvements in the riding experience and path scheduling optimization.

[0088] In this embodiment, step S4 includes the following steps:

[0089] Identify other vehicle perception information in the real-time road network of the area;

[0090] Based on the other vehicle perception information, analyze the real-time location and predict the behavior path;

[0091] Based on the real-time location and estimated behavior path, driving prediction is performed in the current road network environment to generate vehicle driving trajectories within the road network.

[0092] Based on the user intent prediction situation map, the initial path planning is performed on the abnormal fluctuation rendering road network to obtain the fluctuation avoidance path;

[0093] Based on the vehicle driving trajectory and fluctuation avoidance path within the road network, the path of adjacent vehicles is predicted, and the path conflict probability is calculated to obtain the multi-vehicle path conflict probability.

[0094] In this embodiment, vehicle-mounted sensors and vehicle-to-everything (V2X) technology are used to identify and collect perception information of other vehicles in the area in real time, providing a data foundation for path prediction and conflict analysis. Vehicle-mounted radar, LiDAR, high-definition cameras, and ultrasonic sensors capture dynamic information such as speed, acceleration, position, and lane occupancy of surrounding vehicles in real time. Simultaneously, using V2X communication protocols, the positioning and status broadcast data of nearby vehicles are acquired in real time. Sensor data is processed through a multi-sensor fusion algorithm (based on Kalman filtering and particle filtering methods) to achieve high-precision dynamic tracking and identification of vehicles within the area. The system can achieve a vehicle positioning error of less than 0.3 meters and an update frequency of over 10Hz in complex urban environments, ensuring real-time perception of the surrounding dynamic environment. Through this step, a multi-vehicle real-time perception database covering the target area is constructed, laying a solid data foundation for subsequent path prediction and conflict analysis. Using historical vehicle trajectory data and current motion status, combined with a deep learning-based trajectory prediction model (such as a trajectory prediction network based on Long Short-Term Memory (LSTM)), the future position of vehicles is predicted in a time series. The model takes into account current speed, acceleration, steering angle, and surrounding environmental features. The output is a sequence of possible vehicle trajectories for the next 5-10 seconds. In real-world urban road tests, the trajectory prediction accuracy achieved an average error of less than 0.5 meters. To improve the predictive reliability, the system integrates rule constraints and road network structure information, employing a multimodal prediction method to generate multiple possible paths and calculate the confidence probability of each path. Furthermore, the model incorporates an attention mechanism to enhance its predictive capabilities for key road sections and intersections, particularly improving the accuracy of identifying turning behavior at intersections to over 88%. The predicted behavioral paths provide crucial information for obstacle avoidance and path planning for autonomous vehicles in dynamic environments.

[0095] The real-time locations and predicted paths of all neighboring vehicles are input into a unified road network environment driving prediction module to generate a comprehensive driving trajectory map of multiple vehicles within the area. This module employs a multi-agent simulation method, combining vehicle interaction models (such as social force models and game theory models) to simulate the mutual influence and avoidance behaviors between vehicles, improving the realism and reliability of the driving trajectories. During the simulation, the system dynamically adjusts the paths of each vehicle, considering traffic light status, road speed limits, and congestion conditions to ensure the physical rationality of the trajectories. Through high-frequency updates with a time step of 0.1 seconds, the system generates a continuous sequence of vehicle trajectories for the next 10 seconds. Experimental results show that the trajectories generated using this method better reflect the real driving state than single-vehicle predictions, reducing trajectory errors by approximately 15% and increasing the collision warning advance time between vehicles by 20%. This driving trajectory within the road network provides a dynamic environment model for autonomous driving path optimization, supporting multi-vehicle cooperative obstacle avoidance and path conflict assessment. Based on the risk assessment indicators of the road network rendered by abnormal fluctuations, a weighted road network map is constructed, where road weights comprehensively consider the degree of anomaly, user intent requirements, and real-time traffic flow. By employing heuristic search algorithms (such as the improved A* algorithm) combined with a dynamic weight adjustment strategy, a low-risk path that meets user needs is planned. The path planning process dynamically integrates driving urgency and comfort preferences from the user's travel intentions, adjusting path priorities and lane-changing strategies accordingly. Tests show that this initial fluctuation avoidance path can reduce the average path anomaly risk index by more than 20%, while keeping the path length increase within 5%, thus meeting both user safety and efficiency requirements. This path provides baseline path information for subsequent prediction of adjacent vehicle paths and calculation of conflict probabilities.

[0096] By integrating the driving trajectories of neighboring vehicles and the fluctuating avoidance paths of user vehicles, the system calculates and analyzes the probability of multi-vehicle path conflicts. The system first performs spatiotemporal overlap detection on vehicle paths, identifying path intersections and potential conflict areas. A probabilistic graphical model is used to evaluate the spatiotemporal probability distribution of path conflicts, combined with a vehicle dynamics model to estimate the likelihood of conflict occurrence. To more accurately assess conflict risk, the system introduces a Bayesian inference method to model path selection uncertainty and environmental changes, outputting the conflict probability between each pair of vehicles. Experimental data shows that at typical urban intersections, the accuracy of conflict probability prediction reaches 87%, with a false alarm rate of less than 12%. Furthermore, the system adjusts dynamic weights based on the conflict probability and feeds them back to the path optimization module, enabling intelligent avoidance and coordination of vehicle paths. This multi-vehicle path conflict probability calculation effectively improves the cooperative driving safety and path optimization accuracy of autonomous driving systems.

[0097] In this embodiment, the specific steps of step S5 are as follows:

[0098] Conflict optimization analysis is performed based on the probability of multi-vehicle path conflict, and intelligent passage iterative game is conducted to obtain multiple conflict game optimization paths.

[0099] Dynamic driving simulation is performed on the multiple conflict game optimization paths to generate driving simulation data for each optimization path;

[0100] Based on the driving simulation data, a comprehensive path evaluation is performed to obtain a comprehensive path evaluation value for each path;

[0101] The optimal vehicle conflict game optimization path is extracted based on the comprehensive path evaluation value.

[0102] In this embodiment, the path conflict problem is modeled as a multi-agent non-cooperative game, where each vehicle is a player, its policy space is a set of possible paths, and the objective function comprehensively considers path length, conflict risk, and passenger comfort. The Nash equilibrium solution method from game theory is employed, dynamically adjusting path selection to reduce the overall conflict probability by iteratively updating the strategies of each vehicle. The iterative process utilizes reinforcement learning algorithms, such as Deep Q-Networks (DQN) or Multi-Agent Reinforcement Learning (MARL) frameworks, trained in a simulated environment to ensure convergence to a stable policy. In experiments, the algorithm converged in approximately 50 iterations in a 100-vehicle simulated scenario, with the conflict probability reduced by about 30% compared to the initial path. Finally, multiple conflict game-optimized paths are output, representing the optimal driving schemes for vehicles under coordinated driving, effectively improving safety and traffic efficiency. Fine-grained simulation verification is performed on each path using a high-precision dynamic driving simulation platform. The simulation environment includes vehicle dynamics models, sensor models, and complex road scenarios, implemented using simulation software (such as CarSim or SUMO combined with deep learning modules). During the simulation, vehicles travel along game-theoretic optimized paths, dynamically simulating changes in acceleration, steering angle, speed, and interactions with other vehicles. Simulation data includes path time consumption, energy consumption, ride comfort indicators (such as acceleration change rate), collision warning frequency, and actual occurrences of path conflicts. The simulation frequency is set to 100Hz to ensure nuanced dynamic response. In tests at multiple complex intersections and high-density scenarios, the simulation data provides objective quantification of path safety and driving smoothness, laying the foundation for comprehensive path evaluation.

[0103] Based on simulation results, a multi-index comprehensive evaluation system was established to quantitatively score each path. Evaluation indicators cover path safety (collision risk rate), traffic efficiency (total travel time and path length), ride comfort (acceleration and steering smoothness), and energy consumption. The Analytic Hierarchy Process (AHP) and a weighted scoring model were used to normalize each indicator and then weighted to synthesize a comprehensive evaluation value. Weighting coefficients were set according to actual user preferences and autonomous driving safety regulations, such as 40% for safety, 30% for efficiency, 20% for comfort, and 10% for energy consumption. For example, path A has a safety score of 0.9, efficiency of 0.85, comfort of 0.8, and energy consumption of 0.75, resulting in a comprehensive score of 0.845. Experiments show that this evaluation method is highly correlated with actual user satisfaction in multi-scenario verification, with a correlation coefficient of 0.87. This comprehensive evaluation value provides an objective basis for optimal path selection, ensuring the selection of a path scheme that balances safety, efficiency, and comfort. Based on the comprehensive path evaluation value, multiple conflict game optimization paths were ranked, and the path with the highest score was selected as the optimal vehicle conflict game optimization path. This route balances user travel needs, safety risk control, and driving efficiency, possessing the highest practical value. The route selection employs a multi-objective optimization method combined with a ranking algorithm to ensure the fastest selection of the optimal solution under different weight preferences. In practical applications, the system supports dynamic weight adjustment to adapt to different user travel intentions and environmental changes. Experimental tests show that compared to traditional static route planning schemes, this optimal route reduces the probability of conflict by approximately 28%, shortens travel time by 5%, and improves passenger comfort by 12%. This route is directly used for dynamic path execution in autonomous vehicles, improving overall driving intelligence and user satisfaction.

[0104] In this embodiment, step S6 includes the following steps:

[0105] Real-time regional information flow is collected and input into the risk chain effect prediction engine to predict the impact of sudden events, thereby generating the probability and degree of impact of sudden events.

[0106] Dynamic risk buffer path planning is performed on the optimal vehicle conflict game optimization path based on the probability and impact of the emergency, so as to obtain a risk-matched buffer planning path.

[0107] Based on the buffer planning path, real-time dynamic adjustment control of the path is performed, and the path switching timing window is calculated to generate a risk buffer avoidance model.

[0108] The risk buffering and avoidance model is optimized through multi-vehicle collaborative calculation and cloud-based iterative sharing to perform vehicle dynamic path optimization.

[0109] In this embodiment, a multi-source real-time data stream acquisition system integrates data from vehicle-mounted sensors, roadside cameras, intelligent signal systems, and vehicle-to-everything (V2X) platforms to form a high spatiotemporal resolution information stream within the region. A preprocessing module cleans and fuses the data, eliminating noise and missing data. Subsequently, a risk chain reaction prediction engine uses a deep neural network combined with a graph neural network (GNN) to model the regional road network state, capturing the mutual influence relationships between road network nodes and edges. The engine focuses on identifying the triggers of sudden events (such as abnormal traffic surges, accident alarms, etc.) and uses a time-series prediction model (such as a temporal convolutional network TCN) trained on a historical event database to predict the probability of sudden events occurring in the near future (5-15 minutes). The model also assesses the potential chain reactions caused by sudden events, calculating the scope and severity of impact through risk propagation simulation. In experiments, the model achieved an event prediction accuracy of 85% in the actual test area, with the impact severity assessment error controlled within 10%. The output probability and impact severity of sudden events provide key risk indicators for subsequent dynamic path adjustments. A risk-matching buffer path planning is then performed on the determined optimal vehicle conflict game path. The specific process involves mapping the probability and impact of emergencies to risk weights for each road segment in the road network, and dynamically adjusting the edge weights of the path planning graph based on these risk weights. A set of risk buffer paths is generated using a risk-sensitive dynamic programming algorithm, combined with real-time status and user travel needs. These paths are designed to reduce risk by avoiding high-risk areas while remaining close to the optimal path. A fuzzy logic system is introduced during the planning process to balance path safety and driving efficiency, ensuring passenger comfort. In the experimental environment, the buffer paths increased coverage in emergency areas by approximately 25%, reduced the path risk index by more than 20%, and kept path length growth within 5%. This buffer planning path can dynamically respond to changes in sudden risks, improving the safety and reliability of autonomous vehicles.

[0110] By monitoring risk indicators and the vehicle's current location in real time, the system dynamically assesses whether to trigger a route change. The calculation of the route change timing window incorporates a multi-factor decision model, covering the vehicle's current speed, remaining path length, risk change rate, and surrounding density. A Bayesian decision framework is used to fuse historical risk data with real-time perception, achieving accurate estimation of the route change trigger probability. This ensures that the change timing neither prematurely triggers unnecessary lane changes nor delays risk exposure. A smooth transition path generation algorithm is employed in the change strategy to guarantee driving smoothness and passenger comfort during the route change process. The system-generated risk buffer avoidance model outputs change suggestions and dynamic path planning results in real time. Simulation results show that the risk buffer model can shorten the route change response time by approximately 15%, reduce the duration of sudden risk exposure, and improve overall driving safety.

[0111] Multi-vehicle collaboration utilizes a distributed computing framework to share real-time risk data and path adjustment strategies, avoiding local decision-making conflicts and congestion. Federated learning technology, while ensuring privacy and security, enables iterative cloud-based training and optimization of model parameters, improving the adaptability and robustness of overall path planning. The cloud platform dynamically adjusts the weights of the risk assessment model and path planning algorithm based on real-time big data analysis, achieving globally optimal path scheduling. In multi-vehicle collaboration mode, the system's path optimization performance is approximately 18% better than single-vehicle optimization, path conflict rate is reduced by 22%, and dynamic response speed is increased by 20%. This cloud-based iterative optimization mechanism ensures that autonomous vehicles achieve efficient and safe dynamic path optimization in complex environments, improving the overall system operating efficiency.

[0112] In this embodiment, an AI-based autonomous vehicle dynamic path optimization system is provided, used to execute the AI-based autonomous vehicle dynamic path optimization method described above, including:

[0113] The abnormal fluctuation module is used to collect regional status data based on real-time navigation information, perform abnormal traffic flow fluctuation analysis, visualize and render abnormal road sections, and construct an abnormal fluctuation rendering road network.

[0114] The chain reaction module is used to mine the chain reaction of abnormal disturbance propagation in the abnormal fluctuation rendering road network, and to model the dynamic risk effect to build a risk chain effect prediction engine.

[0115] The intent prediction module is used to acquire user riding behavior perception information, perform in-depth analysis of user behavior intent and riding intent situation prediction to obtain user intent prediction situation map.

[0116] The conflict probability calculation module is used to identify the perception information of other vehicles in the real-time road network of the area, predict the path of nearby vehicles based on the user intent prediction situation map, and calculate the path conflict probability to obtain the multi-vehicle path conflict probability.

[0117] The conflict game module is used to perform conflict optimization analysis based on the probability of multi-vehicle path conflict, and to conduct intelligent traffic iteration game to extract the optimal vehicle conflict game optimization path.

[0118] The risk buffer planning module is used to predict the impact of sudden events based on the risk chain effect prediction engine, and to plan risk buffer paths for the optimal vehicle conflict game optimization path in order to perform vehicle dynamic path optimization operations.

[0119] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0120] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. An artificial intelligence-based dynamic path optimization method for an unmanned vehicle, characterized by, Comprise the following steps: Step S1: based on real-time navigation information acquisition area state data, abnormal traffic fluctuation analysis is carried out, and abnormal road section visualization rendering is carried out, and abnormal fluctuation rendering road network is constructed; Step S2: abnormal fluctuation rendering road network is carried out abnormal disturbance propagation chain reaction mining, and dynamic risk effect modeling is carried out, and risk chain effect prediction engine is constructed; Step S3: obtain user ride behavior perception information, carry out user behavior intention deep analysis and ride intention situation prediction, to obtain user intention prediction situation map; Step S4: identify other vehicle perception information of real-time road network in the region, and based on the user intention prediction situation map, the adjacent vehicle path is predicted, and the path conflict probability is calculated, to obtain the multi-vehicle path conflict probability; Step S5: based on the multi-vehicle path conflict probability, the conflict optimization analysis is carried out, and the intelligent traffic iteration game is carried out, and the optimal vehicle conflict game optimization path is extracted; Step S6: based on the risk chain effect prediction engine, the influence of the sudden event is predicted, and the risk buffer path planning is carried out on the optimal vehicle conflict game optimization path, to execute the vehicle dynamic path optimization work. 2.The AI-based dynamic path optimization method for a self-driving vehicle according to claim 1, wherein, The specific steps of step S1 are: Based on real-time navigation information acquisition area state data; the abnormal data cleaning is carried out on the area state data, to obtain the optimized area state data; Set the evolution time range, and identify the road sections in the region for the optimized area state data, and mark a plurality of regional road sections; The road connectivity of the plurality of regional road sections is identified, and the road topological correlation fitting is carried out, to construct the real-time road network in the region; The multi-dimensional feature flow of each road section is generated by carrying out multi-dimensional feature mining on the real-time road network in the region; The flow change rule analysis is carried out on the multi-dimensional feature flow, to generate the road section state atlas; Based on the road section state atlas, abnormal traffic fluctuation analysis is carried out, and abnormal road section visualization rendering is carried out, to construct the abnormal fluctuation rendering road network. 3.The AI-based dynamic path optimization method for a self-driving vehicle according to claim 2, wherein, The specific steps of the abnormal traffic fluctuation analysis based on the road section state atlas, and the abnormal road section visualization rendering, and the construction of the abnormal fluctuation rendering road network are as follows: Based on the road section state atlas, the real-time traffic flow of each road section is calculated; The real-time traffic flow is associated with the road section to obtain the traffic flow correlation characteristics between road sections; The traffic flow fluctuation curve is generated by calculating the traffic flow fluctuation at multiple time points; According to the traffic flow correlation characteristics, the depth identification of the key road section of the traffic flow fluctuation curve is carried out, and the key road section is marked; The abnormal fluctuation probability of the key road section is calculated in different scales, and the abnormal traffic influence degree is calculated, so as to obtain the abnormal degree index of a plurality of key road sections; Based on the abnormal degree index, the abnormal fluctuation visualization rendering of the real-time road network in the region is carried out, to construct the abnormal fluctuation rendering road network. 4.The AI-based dynamic path optimization method for a self-driving vehicle according to claim 1, wherein, The specific steps of step S2 are: Abnormal fluctuation propagation link identification is carried out on the abnormal fluctuation rendering road network, and the abnormal fluctuation evolution propagation path is extracted; The disturbance propagation chain reaction feature is generated by carrying out disturbance propagation chain reaction mining on the abnormal fluctuation evolution propagation path; Potential risk prediction is carried out based on the road network disturbance propagation chain reaction feature, to obtain road network potential risk prediction data; The potential risk prediction data of the road network is evaluated for the chain reaction intensity to obtain a risk chain intensity evaluation value; The disturbance amplification coefficient is calculated according to the risk chain intensity evaluation value to generate the disturbance amplification coefficient of each road section; According to the disturbance amplification coefficient and the abnormal fluctuation evolution propagation path, the dynamic risk effect modeling is performed to construct a risk chain effect prediction engine. 5.The AI-based dynamic path optimization method for a self-driving vehicle according to claim 1, wherein, The specific steps of step S3 are: Based on the vehicle-mounted sensing system, the real-time riding behavior of the vehicle is perceived to obtain user riding behavior perception information; The user riding behavior perception information is deeply analyzed to extract user behavior intention deep features; Based on the user behavior intention deep features, the driving urgency, passenger comfort preference and travel purpose are calculated to obtain user behavior demand features; Real-time interactive voice in the vehicle is obtained, and deep voice semantic understanding and emotional tendency recognition are performed to generate implicit demand mining features; According to the user behavior demand features and the implicit demand mining features, the riding intention situation prediction is performed to obtain a user intention prediction situation map. 6.The AI-based dynamic path optimization method for a self-driving vehicle according to claim 1, wherein, The specific steps of step S4 are: Identify other vehicle perception information of the real-time road network in the region; According to the other vehicle perception information, the real-time position and estimated behavior path are analyzed; Based on the real-time position and estimated behavior path, the current road network environment driving prediction is performed to generate the vehicle driving track in the road network; According to the user intention prediction situation map, the initial path planning is performed on the abnormal fluctuation rendering road network to obtain a fluctuation avoidance path; Based on the vehicle driving track in the road network and the fluctuation avoidance path, the adjacent vehicle path is pre-judged, and the path conflict probability is calculated to obtain the multi-vehicle path conflict probability.

7. The artificial intelligence based dynamic path optimization method for a self-driving vehicle of claim 1, wherein, The specific steps of step S5 are: Based on the multi-vehicle path conflict probability, the conflict optimization analysis is performed, and the intelligent passing iterative game is performed to obtain a plurality of conflict game optimization paths; The dynamic driving simulation is performed on the plurality of conflict game optimization paths to generate driving simulation data of each optimization path; Based on the driving simulation data, the comprehensive path evaluation is performed to obtain a comprehensive path evaluation value of each path; Based on the comprehensive path evaluation value, the optimal vehicle conflict game optimization path is extracted.

8. The artificial intelligence-based dynamic path optimization method for a self-driving vehicle of claim 1, wherein, The specific steps of step S6 are: Collect real-time regional information flow and input into the risk chain effect prediction engine, and perform sudden event influence prediction to generate a sudden event probability and an influence degree value; According to the sudden event probability and the influence degree value, the dynamic risk buffer path planning is performed on the optimal vehicle conflict game optimization path to obtain a risk-matched buffer planning path; Based on the buffer planning path, the real-time path dynamic adjustment control is performed, and the path switching opportunity window is calculated to generate a risk buffer avoidance model; The risk buffer avoidance model is subjected to multi-vehicle collaborative calculation and cloud iterative sharing optimization to perform vehicle dynamic path optimization work.

9. An artificial intelligence based dynamic path optimization system for autonomous vehicles, characterized in that, The method for performing the artificial intelligence-based dynamic path optimization of the unmanned vehicle according to claim 1 comprises: An abnormal fluctuation module is configured to collect regional state data based on real-time navigation information, analyze abnormal traffic fluctuation, and perform abnormal road section visualization rendering to construct an abnormal fluctuation rendering road network; A chain effect module is configured to mine abnormal disturbance propagation chain reactions of an abnormal fluctuation rendered road network and perform dynamic risk effect modeling to construct a risk chain effect prediction engine; An intention prediction module is configured to acquire user riding behavior perception information, perform user behavior intention deep analysis and riding intention situation prediction to obtain a user intention prediction situation map; A conflict probability calculation module is configured to identify other vehicle perception information of a real-time road network in a region, perform adjacent vehicle path prediction based on the user intention prediction situation map, and perform path conflict probability calculation to obtain multi-vehicle path conflict probability; A conflict game module is configured to perform conflict optimization analysis based on the multi-vehicle path conflict probability, perform intelligent passing iterative game, and extract an optimal vehicle conflict game optimization path; A risk buffer planning module is configured to perform emergency event influence prediction based on the risk chain effect prediction engine, and perform risk buffer path planning on the optimal vehicle conflict game optimization path to perform vehicle dynamic path optimization work.

Citation Information

Cited By

  • Road intelligent traffic test platform and test method based on human-vehicle-road cloud cooperation

    CN122116647A

  • Intelligent Transportation Testing Platform and Testing Method Based on Human-Vehicle-Road-Cloud Collaboration

    CN122116647B