Highway traffic flow prediction and accident identification method and system

Abstract

The invention relates to an expressway traffic flow prediction and accident identification method and system, and relates to the technical field of intelligent detection, and the method comprises the steps: firstly, carrying out the real-time monitoring of an expressway through a mobile robot, obtaining the traffic flow feature data, and obtaining the traffic flow prediction data of a specified time period through feature extraction analysis and fusion; and then, according to the traffic flow prediction data and traffic flow data acquired in real time, accident identification and accident grading are carried out, when an accident on the highway is identified, dynamic path planning is carried out by using an algorithm and combining the data identified in real time, and based on the planned path, the mobile robot is controlled to execute a dredging task on the lane of the highway. Through robot automatic monitoring and an intelligent algorithm, whole-process unmanned operation of traffic flow prediction and accident identification and disposal is realized, manual intervention is reduced, the monitoring efficiency and the safety of workers are improved, and the operation cost is reduced.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection technology, and in particular to a method and system for predicting highway traffic flow and identifying accidents. Background Art

[0002] With the large-scale construction of expressway networks and the continued growth of traffic volume, the safe operation and efficient management of expressways face increasingly severe challenges. As a key component of the country's comprehensive three-dimensional transportation network, the efficiency and safety of expressways directly impact socio-economic development and the public's travel experience. Traditional expressway management models rely primarily on manual inspections, fixed surveillance cameras, and limited traffic guidance facilities. These models suffer from issues such as insufficient real-time performance, delayed emergency response, and incomplete traffic situation awareness, making them unable to meet the demands of refined management in complex traffic environments. Therefore, the development of expressway robotic systems with intelligent monitoring and dynamic traffic diversion capabilities has become a key direction for improving traffic management efficiency and ensuring safe and smooth roads.

[0003] In recent years, the rapid development of technologies such as artificial intelligence, the Internet of Things, autonomous driving, and 5G communications has injected new impetus into intelligent highway management. By integrating multimodal sensing devices, intelligent decision-making algorithms, and autonomous actuators, intelligent highway monitoring and traffic control robots can collect and process road conditions, traffic flow information, and emergency events in real time. Combined with cloud-based traffic big data analysis, they dynamically generate and automatically execute optimal traffic control strategies, thereby building an integrated "perception-decision-execution" intelligent traffic management system. The application of this intelligent equipment can effectively address the shortcomings of traditional management models and enhance highway management efficiency and emergency response capabilities in scenarios such as daily operations, accident handling, and severe weather response.

[0004] Existing highway traffic flow monitoring often uses fixed cameras combined with traditional algorithm models. These cameras, mounted on either side of the road or on gantries, record road conditions and detect vehicles, then calculate traffic flow based on corresponding algorithms. Furthermore, some highways employ accident identification methods, which primarily rely on pre-configured accident identification and handling systems. These systems rely on manual patrols, with personnel regularly inspecting the roads at designated locations to conduct traffic diversion and handle incidents upon discovery. This approach is incapable of automatically and accurately identifying incidents. Consequently, existing technologies suffer from low efficiency in automated highway traffic flow monitoring and an inability to automatically identify and handle accidents. Summary of the Invention

[0005] This application provides a method and system for predicting traffic flow and identifying accidents on highways. This addresses the existing challenges of drones and other devices, which are limited by their endurance and anti-interference capabilities, making them difficult to operate stably in inclement weather. This application integrates multimodal sensors, such as rain- and fog-resistant cameras and millimeter-wave radars, into a ground mobile robot for automated monitoring. Combined with a data fusion algorithm, this method significantly improves monitoring reliability in complex environments. Furthermore, while achieving traffic flow prediction, it further enables unmanned accident identification and road planning, improving monitoring efficiency and addressing the low efficiency of existing automated highway traffic flow monitoring and the inability to automatically identify and handle accidents.

[0006] In a first aspect, the present application provides a method for predicting highway traffic flow and identifying accidents, comprising:

[0007] Normalizing the acquired traffic flow characteristic data of the highway to obtain pre-processed target traffic flow characteristic data, wherein the traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow of the highway in real time during dynamic patrol;

[0008] Performing temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and performing spatial feature convolution processing on the acquired highway section data to obtain spatial feature information;

[0009] Based on the time feature information and the space feature information, the traffic flow prediction data for the specified time period is obtained by splicing and fusing;

[0010] Perform accident identification and accident classification assessment based on the traffic flow prediction data and the real-time traffic flow data obtained in real time, and determine accident information, wherein the accident information includes accident analysis results, accident impact range, and accident level;

[0011] Through a dynamic path planning algorithm, dynamic path planning is performed based on the traffic flow prediction data, the real-time traffic flow data and accident information to obtain target path planning information, and the mobile robot is controlled to perform a traffic diversion task on the lane of the highway based on the target path planning information.

[0012] Optionally, performing temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and performing spatial feature convolution processing on the acquired highway section data to obtain spatial feature information, including:

[0013] Taking the target traffic flow characteristic data as input, according to Δ d Q(t)=Q(t)-Q(td) is used for differential processing to eliminate non-stationarity, and the preset autoregressive model is input. Perform convolution processing to extract temporal feature information;

[0014] Analyze the monitoring points and the connection relationship between sections of the highway, perform abstract processing, and construct the abstract graph of the section G = (V, E);

[0015] Take the nodes V and edges E in the road segment abstract graph as input, according to Perform graph convolution operation to obtain spatial feature information;

[0016] Among them, Q(t) is the target traffic flow characteristic data after preprocessing, α p Autoregressive coefficient, β p is the moving average coefficient, ∈(t) is white noise, is an adjacency matrix with self-loops, is the degree matrix, W (l) is the weight of the first layer, X (l) Node features.

[0017] Optionally, performing accident identification and accident classification assessment based on the traffic flow prediction data and the real-time traffic flow data obtained in real time to determine accident information includes:

[0018] Based on the traffic flow prediction data, combined with real-time traffic flow data, traffic flow mutation detection is performed to obtain mutation analysis results;

[0019] When the mutation analysis result is a determined mutation result, obtaining video data of each lane of the highway to perform visual feature analysis to determine the accident analysis result;

[0020] When the accident analysis result is to determine the occurrence of the accident, an impact range analysis and an accident level assessment are performed based on the traffic flow prediction data and the real-time traffic flow data to obtain the accident impact range and the accident level.

[0021] Optionally, traffic flow mutation detection is performed based on the traffic flow prediction data in combination with real-time traffic flow data to obtain mutation analysis results, including:

[0022] Based on the traffic flow prediction data and the real-time traffic flow data, the average speed v(t) of each vehicle in the lane at time t and the density of vehicles in the lane are analyzed as ρ(t), and the mutation judgment formula is used. Compute mutation analysis results;

[0023] Among them, μ is the historical mean, σ is the standard deviation, w1, w2 and w3 are weight coefficients. When the score is greater than the preset threshold, it is determined that an accident may have occurred and a certain mutation result is obtained.

[0024] Optionally, obtain video data from each lane of the highway to perform visual feature analysis and determine accident analysis results, including:

[0025] The acquired video data is detected by the target detection algorithm to obtain the vehicle bounding box set B = {b1,…,b n};

[0026] Based on the vehicle bounding box set B={b1,…,b n} to perform trajectory analysis and obtain the vehicle attitude angle set S = {s1,…,s n};

[0027] The vehicle attitude angle set S = {s1,…,s n} is input, according to Analyze key abnormal features and obtain collision angle variance

[0028] According to the vehicle bounding box set B={b1,…,b n Analyze the pixels of the vehicle bounding box to obtain the total pixels b, and mark the pixels covered by the vehicle bounding box as the vehicle area, and mark the uncovered pixels as the non-vehicle area, and obtain the number of pixels a in the non-vehicle area;

[0029] Take the total pixels b and the number of pixels a as input, according to R debris =a / b Analyze key abnormal characteristics and obtain the debris area ratio R debris ;

[0030] Based on collision angle variance and the debris area ratio R debris , according to the decision function Conduct accident assessment and obtain accident information;

[0031] Among them, P accident When ≥0.8, it is determined that an accident has occurred and the accident result is obtained. k1 and k2 are weight coefficients, b i Represents the coordinates of the rectangular bounding box of the vehicle in the i-th frame of the image, s i Indicates the driving direction or body posture angle of the i-th vehicle in the image.

[0032] Optionally, performing an impact range analysis and an accident level assessment based on the traffic flow prediction data and the real-time traffic flow data to obtain the accident impact range and accident level includes:

[0033] Analyzing the upstream traffic flow q1 and the downstream traffic flow q2 in the highway according to the traffic flow prediction data and the real-time traffic flow data;

[0034] Based on the upstream traffic flow q1 and downstream traffic flow q2, according to vw =q2-q1 / ρ2-ρ1, calculate the congestion propagation speed v w ;

[0035] Based on the congestion propagation speed v w , according to L max =v w ·T response Calculate the maximum impact distance as the accident impact range;

[0036] Based on the vehicle bounding box set, analyze the covered lane area, determine the number of blocked lanes, and obtain the probability of personal injury;

[0037] Based on the traffic flow prediction data and the real-time traffic flow data, the driving speed v of each vehicle in the congested section of the highway is analyzed. i , driving distance l i and the free section flow velocity v free ;

[0038] Based on the driving speed v i , driving distance l i and the free section flow velocity v free , according to V delay =∑ i (v free -v i )·l i Calculate the total delayed vehicle kilometers;

[0039] Based on the total delayed vehicle kilometers, the accident level is determined by fuzzy comprehensive evaluation method.

[0040] Optionally, dynamic path planning is performed based on the traffic flow prediction data, the real-time traffic flow data, and accident information using a dynamic path planning algorithm, including:

[0041] Performing road network modeling based on the traffic flow prediction data, the real-time traffic flow data, the highway road data, and the accident information, and performing path optimization using a cost function to obtain rescue path optimization information, wherein the rescue path optimization information is used to plan an optimal path for the robot from the starting point to the accident point;

[0042] Based on the accident point corresponding to the accident information, a diversion path is planned by using a gradient descent method of potential field theory to generate a target diversion path;

[0043] Based on the optimal path and the target diversion path, target path planning information is generated.

[0044] Optionally, a road network model is performed based on the traffic flow prediction data, the real-time traffic flow data, the highway road data, and the accident information, and a cost function is used to perform path optimization to obtain rescue path optimization information, including:

[0045] Based on the traffic flow prediction data and the real-time traffic flow data, the real-time traffic flow q is determined. ij , and, determining the road capacity C based on the highway road data ij ;

[0046] Real-time traffic q ij and road capacity C ij For input, according to Carry out road network modeling and use the cost function of the improved algorithm f(n)=g(n)+h(n)+γVar(c ij ) perform path optimization to obtain rescue path optimization information;

[0047] Based on the accident point corresponding to the accident information, the diversion path planning is performed by the gradient descent method of the potential field theory to generate the target diversion path, including: using the gradient descent method of the potential field theory to analyze the accident point, according to Perform diversion path planning and generate target diversion paths;

[0048] Among them, the starting point i and the end point j of the rescue section are identified by the accident information, d ij is the distance of the road section, v ij (t) is the real-time average speed of the road section at time t, λ is the weight coefficient, g(n) represents the actual cost from the starting point to the current node, h(n) represents the heuristic estimated cost from the current node to the end point, γ represents the path stability factor, Var(c ij ) represents the variance of the cost of each section in the path, is the potential field gradient, Q i represents the repulsive force strength at the accident point, r represents the current vehicle position coordinate, r i represents the coordinates of the accident point, represents the unit vector pointing from the accident point to the vehicle.

[0049] Optionally, after obtaining the target path planning information, the following steps are also included:

[0050] When dispatching the mobile robot to the accident point corresponding to the accident information, the dispatch information x is obtained. ij and scheduling time t ij , define the decision variables

[0051] According to the decision variables, construct the objective function And based on the decision variables, construct constraints

[0052] Optimize resource scheduling of mobile robots based on objective functions and constraints;

[0053] After dispatching the mobile robot to perform the evacuation task, obtaining handling information, wherein the handling information includes handling time and accident impact range;

[0054] Based on the disposal information, Calculate the recovery efficiency η;

[0055] The reward value r is determined based on the recovery efficiency η, according to Construct a reward function and continuously iterate the learning mechanism to update the mobile robot's execution strategy when an accident occurs;

[0056] Among them, w j is the weight determined based on the accident level, T actual is the actual processing time, T predict is the estimated disposal time, L actual is the actual impact range of the accident, L actual is the estimated impact range of the accident, α is the learning rate, γ is the discount factor, (s, a) represents the current state and action, and s' represents the new state after the action is executed.

[0057] In a second aspect, the present application provides a highway traffic flow prediction and accident identification system, comprising:

[0058] A traffic flow characteristic data processing module is used to perform normalization processing on the acquired highway traffic flow characteristic data to obtain pre-processed target traffic flow characteristic data. The traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow on the highway in real time during dynamic patrol. The mobile robot can be equipped with an anti-interference multimodal sensor array, including a rain and fog resistant camera, millimeter wave radar, lidar, and an inertial navigation system. The mobile robot fuses multi-source data through an adaptive Kalman filter algorithm to suppress environmental noise (such as rain, snow, and strong light interference), ensuring stable collection of traffic flow data in extreme environments such as the temperature range of -40°C to 70°C and heavy rainfall, and determining traffic flow characteristic data.

[0059] a feature convolution extraction module, configured to perform temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and to perform spatial feature convolution processing on the acquired highway section data to obtain spatial feature information;

[0060] A feature splicing and fusion module, configured to perform splicing and fusion based on the temporal feature information and the spatial feature information to obtain traffic flow prediction data for a specified time period;

[0061] An accident identification and classification module is used to identify and classify accidents based on the traffic flow prediction data and the real-time traffic flow data obtained in real time, and determine accident information, wherein the accident information includes accident analysis results, accident impact range, and accident level;

[0062] The path planning module is used to perform dynamic path planning based on the traffic flow prediction data, the real-time traffic flow data and accident information through a dynamic path planning algorithm to obtain target path planning information, and control the mobile robot to perform the diversion task on the lane of the highway based on the target path planning information; it can support multi-robot collaborative operation, and through a distributed collaborative scheduling algorithm, combine real-time traffic flow data with robot position information to dynamically allocate diversion tasks and optimize the paths of each robot to avoid path conflicts, ensure that multiple robots arrive at the accident point quickly and orderly, and minimize interference with normal traffic flow.

[0063] In summary, the embodiment of the present application first uses a mobile robot to monitor the highway in real time, obtains traffic flow feature data, and obtains traffic flow forecast data for a specified time period through feature extraction, analysis and fusion. Then, based on the traffic flow forecast data and the real-time traffic flow data, accident identification and accident classification are performed. When an accident on the highway is identified, an algorithm is used to combine the real-time identification data to perform dynamic path planning, and based on the planned path, the mobile robot is controlled to perform traffic diversion tasks on the lanes of the highway. On the one hand, the present application uses a mobile monitoring and diversion robot, combined with multi-source sensors (such as radar and video) and dynamic deployment strategies, to achieve dynamic coverage and real-time monitoring of the entire highway section, eliminating the monitoring blind spots of fixed equipment and improving the comprehensiveness of road condition information collection. On the other hand, the robot performs automated monitoring based on the collected video data and combines it with intelligent algorithms to achieve accurate traffic flow prediction, automated accident identification and handling of the entire process, reducing manual intervention, improving monitoring efficiency and staff safety, and reducing operating costs. The existing technology for automated traffic flow monitoring on highways has low efficiency and cannot automatically identify / handle accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0065] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0066] Figure 1 A flowchart of a method for predicting highway traffic flow and identifying accidents provided in an embodiment of the present application;

[0067] Figure 2 This is a schematic flow chart of the steps of a method for predicting highway traffic flow and identifying accidents provided by an optional embodiment of the present application;

[0068] Figure 3 This is a structural block diagram of a highway traffic flow prediction and accident identification system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0070] To facilitate understanding of the embodiments of the present application, further explanation will be given below in conjunction with the drawings and specific embodiments. The embodiments do not constitute a limitation on the embodiments of the present application.

[0071] Figure 1 This is a flow chart of a highway traffic flow prediction and accident identification method provided in an embodiment of the present application. Figure 1 As shown, the highway traffic flow prediction and accident identification method provided by the embodiment of the present application may specifically include the following steps:

[0072] Step 110 , performing normalization processing on the acquired highway traffic flow characteristic data to obtain pre-processed target traffic flow characteristic data.

[0073] The traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow on the highway in real time during dynamic patrol.

[0074] In this embodiment, the mobile robot is typically deployed on a highway. This mobile robot platform is equipped with multiple sensors, including rain- and fog-resistant cameras, radar, video, and lidar. It integrates multimodal sensing equipment such as high-precision cameras, millimeter-wave radar, infrared sensors, and weather detectors to form a fusion architecture. The mobile robot continuously monitors the highway in real time, providing comprehensive coverage of the entire highway. By adjusting the robot's patrol path and sensor coordination strategies in real time, it eliminates the blind spots of traditional fixed cameras and significantly improves the integrity and timeliness of road condition data collection.

[0075] In this embodiment, a mobile robot monitors highways to collect relevant data, including but not limited to historical traffic flow, real-time traffic flow, time characteristics (such as hour / day / holiday), weather indices, and traffic flow on adjacent sections. This data is then preprocessed, including normalization, to uniformly scale the features of different data to a closed interval and eliminate differences in feature scales. This results in target traffic flow feature data.

[0076] Step 120 : performing temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and performing spatial feature convolution processing on the acquired highway section data to obtain spatial feature information.

[0077] Step 130 : performing splicing and fusion based on the temporal feature information and the spatial feature information to obtain traffic flow prediction data for a specified time period.

[0078] A unified description of steps 120 to 130 is provided:

[0079] In this embodiment, the road section data includes but is not limited to: each road section, monitoring point, the road section connection relationship between monitoring points, lanes, length, etc.

[0080] When this embodiment performs time feature convolution on the target traffic flow characteristic data, the target traffic flow data is subjected to differential processing to eliminate non-stationarity. Then, it is input into the model (such as the autoregressive model, ARIMA) to analyze the time correlation, including capturing the trend of traffic flow over time (such as weekday / weekend differences) and periodicity (such as daily peaks) through historical traffic flow, time characteristics and other data, and extracting time series features through graph convolution operations to obtain time feature information. Among them, ARIMA eliminates the non-stationarity of traffic flow data by differential, and combines autoregressive and sliding average modeling time correlation, thereby providing reliable time dimension analysis support for traffic flow prediction and accident warning for the entire road section. This technology works in conjunction with the subsequent spatial modeling capabilities of GCN to jointly build the "time and space perception" foundation of the intelligent monitoring system for highways.

[0081] When performing spatial feature convolution on road segment data, the data is abstracted into a graph and input into the graph convolutional network (GCN) for convolution operations. This captures local (adjacent road segments) and global (cross-regional road segments) spatial correlations and extracts spatial feature information. Graph convolutional GCN abstracts highways into graph networks and utilizes normalization operations on the adjacency matrix and degree matrix to efficiently model the spatial correlation between monitoring points. Its core value lies in: geometric invariance: regardless of the road segment layout, spatial features can be adaptively learned through the graph structure; end-to-end training: combined with ARIMA, a joint spatiotemporal model is formed, providing key support for traffic flow prediction and accident identification for the entire road segment; engineering interpretability: the physical meaning of nodes and edges is clear, making it easy to optimize the model in combination with actual highway scenarios (such as adjusting the adjacency matrix to reflect real-time construction diversions).

[0082] This embodiment uses ARIMA to process trends and cycles in the time dimension and GCN to process neighbor influences in the spatial dimension. It then combines the two to perform feature fusion to obtain traffic flow forecast data for a specified time period (e.g., the next h hours), implements "time-space dual flow" modeling, and improves traffic flow forecast accuracy.

[0083] Step 140 , performing accident identification and accident classification assessment based on the traffic flow prediction data and the real-time traffic flow data obtained in real time, and determining accident information.

[0084] The accident information includes the accident analysis results, the accident impact range and the accident level.

[0085] In this embodiment, the accident analysis results are divided into accident occurrence results and accident non-occurrence results.

[0086] In its implementation, this embodiment primarily determines whether an accident has occurred by analyzing changes in traffic flow and video data. Specifically, it combines predicted and real-time traffic flow data to analyze the average speed of each lane, among other factors, to determine if a sudden change in traffic flow exists. For example, changes in vehicle speeds upstream and downstream of a lane, as well as changes in vehicle density within the lane, are compared to analyze the sudden change in traffic flow to determine if a traffic accident has occurred. Furthermore, when a sudden change in traffic flow is determined, visual feature analysis can be performed in conjunction with the corresponding video data. By analyzing the changes in vehicle traffic based on the video data and the sudden change in traffic flow, a determination is made as to whether a traffic accident has occurred. If a traffic accident is determined to have occurred, an accident occurrence result is generated; otherwise, an accident non-occurrence result is generated. This combined analysis of multi-source data helps avoid misjudgments by a single sensor. For example, sudden changes in traffic flow may be caused by weather, construction, or other factors (e.g., a decrease in vehicle speed due to heavy rain), requiring visual evidence to eliminate interference. Visual features may be affected by image blur (e.g., at night), requiring the assistance of a sudden change in traffic flow signal trigger.

[0087] In actual implementation, traffic flow prediction data and real-time traffic flow data can be used to analyze changes in traffic flow, including analyzing traffic flow upstream and downstream of lanes to determine the scope of accident impact.

[0088] When an accident is determined, this embodiment can classify the accident. Accident classification is primarily based on the severity of the accident (which can be divided into grades I-IV). Accident classification can be performed by determining the degree of congestion based on changes in traffic flow, or by analyzing the degree of congestion in conjunction with video data. Furthermore, the probability of personal injury can be analyzed (typically based on predictions from onboard sensors (such as collision accelerometers), visual image analysis (personnel fall detection), or historical accident models). The total vehicle-kilometer delay (VK) is then further analyzed, i.e., the congestion caused by the accident, and the resulting total number of kilometers of delay for each vehicle, to distinguish the accident's severity.

[0089] Furthermore, dynamic route planning can be implemented based on the accident level for traffic diversion and rescue. In addition, resource scheduling optimization can be achieved based on the accident level, that is, different levels of resources are allocated according to the different accident levels.

[0090] Step 150, through a dynamic path planning algorithm, dynamic path planning is performed based on the traffic flow prediction data, the real-time traffic flow data and accident information to obtain target path planning information, and the mobile robot is controlled to perform a traffic diversion task on the lane of the highway based on the target path planning information.

[0091] In this embodiment, the target path planning information includes a target planned path, which includes two optimized planned paths, namely a rescue path and a diversion path.

[0092] The dynamic path planning algorithm in this embodiment is an improved algorithm, which is mainly composed of two algorithms: ①, the improved A* algorithm (heuristic path search algorithm), which uses the accident point and traffic flow data corresponding to the accident information as a benchmark to perform efficient optimal path search, thereby planning an efficient optimal path for rescue path optimization; ②, the vehicle detour path generation algorithm (implemented based on the gradient descent method of potential field theory), which mainly focuses on accident point analysis and generates a detour path based on the reverse direction of the accident point as a diversion path, allowing vehicles to pass along the diversion path.

[0093] In a specific implementation, after the target planning path is determined in this embodiment, the robot can move on the lane of the highway (such as the emergency lane) according to the target planning path to perform the traffic diversion task.

[0094] As can be seen, the existing technology has limited monitoring range: fixed cameras can only shoot from a single direction, resulting in blind spots. The existing technology relies on manual patrols, which consumes a lot of manpower and time, and in severe weather or complex road conditions, patrol efficiency and safety are affected. This embodiment first uses mobile robots to conduct dynamic patrols and real-time monitoring on highways to address the limitations of the monitoring range. Then, it uses multi-source data obtained from real-time monitoring to perform feature extraction and fusion processing to accurately predict future traffic flow, thus achieving accurate prediction of traffic flow. To address the existing technology's lack of real-time performance and response speed in accident detection and accident handling (specifically, from accident detection to information transmission and then to personnel arriving at the scene to handle the situation), the mobile robot uses real-time analysis data to perform accident identification and classification, determine the accident situation, and then uses an improved dynamic path planning algorithm to dynamically plan rescue and diversion paths. Finally, the mobile robot performs the traffic diversion task according to the planned path. Thus, this embodiment, through robotic automated monitoring and intelligent algorithms, achieves unmanned operation of the entire process of traffic flow prediction, accident identification, and handling, reducing manual intervention, improving monitoring efficiency and staff safety, and lowering operating costs.

[0095] Reference Figure 2 , shows a schematic flow chart of the steps of a method for predicting highway traffic flow and identifying accidents provided by an optional embodiment of the present application. The method may specifically include the following steps:

[0096] Step 210 , performing normalization processing on the acquired highway traffic flow characteristic data to obtain pre-processed target traffic flow characteristic data.

[0097] The traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow on the highway in real time during dynamic patrol.

[0098] In response to the shortcomings of insufficient monitoring coverage and manual patrol methods, some existing technologies use drone-based monitoring systems. These drones are equipped with cameras and other equipment and take off according to preset routes or when needed to conduct aerial monitoring of road conditions. This existing technology has the following main shortcomings: 1. Limited endurance: Drone batteries have short battery life, which limits the duration of a single monitoring session, making it difficult to continuously and stably monitor highways for long periods of time. 2. Weak anti-interference capabilities: In complex electromagnetic environments or severe weather conditions, drones' flight stability and data transmission reliability may be affected, resulting in loss of monitoring data or failure to transmit properly. 3. Flight safety risks: When flying over highways with heavy traffic and complex environments, drones pose certain safety risks, such as collisions with other objects or falls, which may threaten the normal operation of the highway and the safety of personnel.

[0099] This embodiment primarily utilizes intelligent mobile robots. Based on an integrated fusion architecture, combined with spatiotemporal registration and data fusion algorithms, it enables comprehensive, real-time monitoring of factors such as road damage, ice and water accumulation, traffic density, vehicle speed, and abnormal events. For example, LiDAR point cloud data modeling enables accurate identification of road obstacles, and combined with visual image analysis technology, dynamic classification of vehicle type and driving status is achieved, providing multi-dimensional data support for subsequent decision-making.

[0100] In the specific implementation, this embodiment uses a mobile robot to monitor the highway to obtain historical traffic flow Q(t), time characteristics (hour / week / holiday T(t)), weather index W(t), and traffic flow of adjacent sections Q adj (t), etc. Then, the acquired traffic flow characteristic data is normalized to obtain the target traffic flow characteristic data.

[0101] For example, the normalization process can be achieved by the formula: Where X represents the input data. For example, when the input normalized data is the historical traffic flow Q(t), then By normalizing the data to the range [0,1], we ensure that all input features meet the model's unified data format requirements.

[0102] This embodiment uses normalization processing to eliminate significant differences in the original units and numerical ranges of different feature data, thereby preventing the model from being dominated by high-value features; secondly, all features are mapped to the same interval, and the data scale is unified, which facilitates the model (such as ARIMA, graph convolutional network) to learn the weight relationship between features, thereby improving training efficiency and prediction accuracy; finally, the data distribution is maintained, that is, the relative distribution characteristics of the original data are retained, and only the numerical range is adjusted without changing the inherent laws of the data.

[0103] In actual implementation, mobile robots have high adaptability to complex environments, can ensure reliable and stable data, use the high-precision equipment on board to achieve large-scale continuous monitoring, obtain a variety of data within the monitoring range, and combine algorithms to solve the problem of detection failure of a single device in bad weather or complex terrain.

[0104] In step 220 , a temporal feature convolution process is performed on the target vehicle flow characteristic data to obtain temporal feature information, and a spatial feature convolution process is performed on the acquired highway section data to obtain spatial feature information.

[0105] In the implementation, considering that highway traffic flow data often exhibits non-stationary characteristics (such as trends and seasonality), direct modeling can lead to forecasting bias. Therefore, differencing is introduced. Differencing eliminates long-term trends and seasonality in the data by calculating the differences between adjacent values ​​in the series, transforming it into a stationary series that meets the prerequisites for the ARIMA model. The model then extracts time series features, capturing the temporal dependencies within the stationary series and enabling modeling of traffic flow time series characteristics.

[0106] As for spatial characteristics, the main focus is on capturing the spatial dependencies between highway sections, such as the impact of upstream congestion on downstream flow.

[0107] In an optional embodiment, the above-mentioned time feature convolution processing is performed on the target traffic flow characteristic data to obtain time feature information, and the spatial feature convolution processing is performed on the acquired highway section data to obtain spatial feature information, which may include: taking the target traffic flow characteristic data as input, and calculating the time feature information according to Δ d Q(t)=Q(t)-Q(td) is used for differential processing to eliminate non-stationarity, and the preset autoregressive model is input. Perform convolution processing to extract time feature information; analyze the monitoring points and section connection relationship of the highway section, and perform abstract processing to construct the section abstract graph G = (V, E); take the node V and edge E in the section abstract graph as input, according to Perform graph convolution operation to obtain spatial feature information; among them, Q(t) is the target traffic flow feature data after preprocessing, α p Autoregressive coefficient, β p is the moving average coefficient, ∈(t) is white noise, is an adjacency matrix with self-loops, is the degree matrix, W (l) is the weight of the first layer, X (l) Node features.

[0108] The modeling and extraction of time features are described in detail:

[0109] In this embodiment, first, in the differential processing formula: d is the differential order, Δ d Q(t) represents the stationary sequence after first-order differencing of the original sequence Q(t). For example, the first-order difference d = 1 is Q(t) - Q(td), which is used to eliminate linear trends. By using differencing, the non-stationary traffic flow series is converted into a stationary series, making it easier to capture temporal correlations.

[0110] Then, the time features are extracted from the input model. The model formula consists of the following components: autoregressive term (AR part): Mainly use the difference sequence value to predict the current value; moving average item (MA part): The error correlation is modeled using the past white noise (forecast error) sequence. The time dependency in the stationary sequence is captured through the linear combination of AR and MA terms, thus achieving the modeling of the time series characteristics of traffic flow.

[0111] Among them, the parameter order is determined: the moving average term Q is determined by observing the autocorrelation function (ACF), the autoregressive term order P is determined by observing the partial autocorrelation function (PACF) graph, and the difference order d is confirmed by the ADF test.

[0112] The modeling and extraction of spatial features are described in detail:

[0113] In the abstract highway graph, nodes V represent monitoring points (e.g., sensor nodes installed every kilometer), each containing feature data. Edges E represent road segment connections, namely, the physical connectivity between adjacent monitoring points (e.g., upstream and downstream lanes). In the self-loop adjacency matrix, A is the original adjacency matrix (node ​​connectivity matrix), and I is the identity matrix (allowing nodes to retain their own characteristics). The node feature matrix at layer l contains multidimensional data such as historical traffic flow, time, and weather; the weight matrix at layer l is primarily determined by training and learning the transformation of node features; σ is the activation function. By standardizing the adjacency matrix and degree matrix, weighted aggregation of adjacent node features is achieved, thereby capturing the spatial dependencies between road segments.

[0114] Step 230 : performing splicing and fusion based on the temporal feature information and the spatial feature information to obtain traffic flow prediction data for a specified time period.

[0115] In this embodiment, the time features and spatial features are spliced ​​in the dimension to form a joint feature vector containing time and space information. The joint features are nonlinearly transformed through a fully connected neural network to learn the interactive relationship between time and space features, and finally the traffic flow prediction data for the next h period is output.

[0116] Therefore, this embodiment utilizes joint spatiotemporal modeling to simultaneously capture the temporal patterns of traffic flow (such as morning and evening rush hours) and spatial propagation characteristics (such as congestion diffusion), breaking through the limitations of traditional models that only consider time or space. By integrating multidimensional information, the bias of a single model is reduced, improving prediction accuracy and making it suitable for complex traffic scenarios.

[0117] This embodiment uses the Mean Square Error (MSE) loss function to combine and fuse the two features and output the predicted data, taking into account that traffic flow prediction is a continuous numerical prediction (regression task). This loss function measures the mean squared deviation between the predicted value and the true value, with smaller values ​​indicating higher prediction accuracy. The squaring operation amplifies the weight of larger errors, forcing the model to prioritize reducing predictions with significant deviations.

[0118] For example, the loss function can be

[0119] Step 240 : Based on the traffic flow prediction data and in combination with the real-time traffic flow data, traffic flow mutation detection is performed to obtain mutation analysis results.

[0120] In this embodiment, traffic flow mutation detection and visual feature analysis are used to determine whether a traffic accident has occurred. Traffic flow mutation detection uses vehicle flow data as a benchmark and analyzes three key parameters in real time to capture abnormal dynamics within the lane. Key parameters include: lane average speed (a sudden drop in the average speed of all vehicles in a lane at the same time is a direct signal of traffic congestion or an accident), lane density (the number of vehicles within a certain length of lane at the same time; a surge in density reflects the converging or stagnant flow of traffic), and acceleration (according to speed calculation, a sudden increase in deceleration may indicate emergency braking of the vehicle, a precursor to an accident). This embodiment uses real-time monitoring and analysis of key parameters to identify the vehicle's situation in the lane and capture abnormal dynamics, thereby analyzing mutations. When it is determined that no mutation has occurred, the non-mutation result is used as the mutation analysis result; when it is determined that a mutation has occurred, the mutation result is used as the mutation analysis result.

[0121] Optionally, based on the traffic flow prediction data, combined with the real-time traffic flow data, traffic flow mutation detection is performed to obtain mutation analysis results, including: based on the traffic flow prediction data and the real-time traffic flow data, analyzing the average speed v(t) of each vehicle in the lane at time t and the density of vehicles in the lane as ρ(t), and according to the mutation judgment formula Calculate the mutation analysis results; where μ is the historical mean, σ is the standard deviation, w1, w2, and w3 are weight coefficients. When the score is greater than the preset threshold, it is determined that an accident may have occurred and the mutation result is determined.

[0122] In this embodiment, μ and σ are updated in real time based on historical data of a sliding time window (e.g., the past 30 minutes) to adapt to the periodic changes in traffic flow.

[0123] The weight coefficient satisfies w1+w2+w3=1, and the weight can be adjusted according to non-passing scenarios such as urban highways or mountain highways.

[0124] When the score exceeds the preset threshold (judged by empirical analysis), the system determines that a sudden change in traffic flow has occurred, triggering the subsequent accident identification process, namely the visual feature analysis process. This embodiment calculates the score value in real time based on edge computing, which can achieve anomaly recognition in seconds, buying time for emergency response.

[0125] Therefore, this embodiment combines data of different dimensions to perform parameter fusion and realize traffic flow mutation analysis. This embodiment performs traffic flow mutation analysis to realize accident precursor detection and congestion warning.

[0126] Step 250 , when the mutation analysis result is a determined mutation result, obtain video data of each lane of the highway to perform visual feature analysis to determine the accident analysis result.

[0127] In this embodiment, when a sudden change in traffic flow is detected, a visual feature analysis process is executed, such as calling the acquired video data to analyze whether there are vehicle collision debris, abnormal vehicle posture, etc. Specifically, the video stream is collected by the camera deployed on the robot, and the image is analyzed in real time using a target detection algorithm (such as YOLO and Faster R-CNN). Two types of key information are output: vehicle positioning and vehicle posture. This is used to analyze and calculate basic parameters such as vehicle density and spacing, as well as whether the vehicle has experienced abnormal deviation (such as yaw or rollover after a collision), so as to determine whether an accident has occurred. If an accident is determined to have occurred, the result of determining the occurrence of the accident is used as the accident analysis result; if it is determined that an accident has not occurred, the result of not having occurred is used as the accident analysis result.

[0128] In an optional embodiment, the embodiment of the present application obtains video data of each lane of the highway to perform visual feature analysis and determine the accident analysis result, which may specifically include: detecting the obtained video data through a target detection algorithm to obtain a vehicle bounding box set B = {b1, ..., b n}; Based on the vehicle bounding box set B={b1,…,b n} to perform trajectory analysis and obtain the vehicle attitude angle set S = {s1,…,s n}; With the vehicle attitude angle set S = {s1,…,s n} is input, according to Analyze key abnormal features and obtain collision angle variance According to the vehicle bounding box set B={b1,…,b n}Analyze the pixels of the vehicle bounding box to obtain the total pixels b, and mark the pixels covered by the vehicle bounding box as the vehicle area, and mark the uncovered pixels as the non-vehicle area, and obtain the number of pixels a in the non-vehicle area; take the total pixels b and the number of pixels a as input, according to R debris =a / b Analyze key abnormal characteristics and obtain the debris area ratio R debris ; Based on collision angle variance and the debris area ratio R debris , according to the decision function Perform accident judgment and obtain accident information; among them, P accident When ≥0.8, it is determined that an accident has occurred and the accident result is obtained. k1 and k2 are weight coefficients, b i Represents the coordinates of the rectangular bounding box of the vehicle in the i-th frame of the image, s iIndicates the driving direction or body posture angle of the i-th vehicle in the image.

[0129] In this embodiment, video data analysis yields two key pieces of information: a vehicle bounding box set and a vehicle attitude angle set. Each bounding box in the vehicle bounding box set represents the position and size of the corresponding vehicle in the image, useful for locating the vehicle and calculating its spatial distribution. The attitude angle represents the angle between the vehicle's direction of travel and the lane centerline, estimated from the bounding box's tilt angle or the vehicle's tire orientation.

[0130] This embodiment analyzes key abnormal features through visual analysis, mainly analyzing: collision angle variance and the debris area ratio R debris Two types of abnormal features.

[0131] The collision angle variance is mainly based on Analyze the collision angle, It represents the mean of all vehicle attitude angles in the current lane, and its variance reflects the degree of dispersion of vehicle attitude angles. When the vehicle is driving normally, the vehicle attitude angles are concentrated around 0°. Approaching 0; when an accident occurs, the vehicle may deviate at multiple angles due to the collision. Significantly increased.

[0132] Debris area ratio R debris It is mainly determined based on pixel analysis, including analyzing the number of non-vehicle pixels a and the total number of pixels b in the image. The number of non-vehicle pixels can be determined by image segmentation algorithms (such as MaskR-CNN) to identify pixels other than tires and vehicle bodies (such as scattered debris, liquid, mud, etc.). debris Close to 0; after the accident, there will be debris, oil stains, etc. on the road surface, R debris Significantly increased.

[0133] This embodiment will R debris As input, calculate the probability P of the accident accident First, R debris Convert to linear combination Among them, the weight coefficients k1 and k2 are mainly determined by training historical accident data, which are used to adjust the importance of the two types of features. Then, the Sigmoid function is used to map the linear combination results to the [0,1] interval and convert them into the accident probability P accident .P accident The closer it is to 1, the higher the probability of an accident.

[0134] Therefore, this embodiment realizes multi-dimensional data fusion, combining vehicle posture anomalies (collision angle variance) and environmental anomalies (scattered debris) to avoid misjudgment of a single feature (for example, temporary obstacles in the construction area will not trigger two types of features at the same time).

[0135] It should be noted that this embodiment utilizes localized edge computing nodes for all processes, from traffic flow prediction to accident identification and handling. This integrated multimodal data real-time processing and decision-making framework, using lightweight convolutional neural networks and spatiotemporal sequence analysis models, reduces accident detection response time to seconds, forming a low-latency closed loop of "perception-decision-handling."

[0136] Step 260: When the accident analysis result is a determination of the occurrence of an accident, an impact range analysis and an accident level assessment are performed based on the traffic flow prediction data and the real-time traffic flow data to obtain the accident impact range and accident level.

[0137] In specific implementation, after determining that an accident has occurred, the scope of the accident impact can be further analyzed, and the level of the accident can be assessed, which serves as an important basis for subsequent path planning, resource scheduling / decision-making.

[0138] Specifically, the impact range analysis mainly includes analyzing the upstream and downstream congestion conditions, as well as the rescue impact and congestion propagation caused by the accident.

[0139] The accident level assessment is mainly to determine the severity of the accident, which can be determined based on three core indicators, including: personal injury rate (i.e., estimated personal injuries), number of blocked lanes (the number of lanes that are impassable due to the accident) and total delayed vehicle kilometers (i.e., the cumulative delay time of affected vehicles).

[0140] In an optional embodiment, the above-mentioned analysis of the impact range and assessment of the accident level based on the traffic flow prediction data and the real-time traffic flow data to obtain the accident impact range and accident level may include: analyzing the upstream traffic flow q1 and the downstream traffic flow q2 on the highway based on the traffic flow prediction data and the real-time traffic flow data; based on the upstream traffic flow q1 and the downstream traffic flow q2, according to v w =q2-q1 / ρ2-ρ1, calculate the congestion propagation speed v w ; Based on the congestion propagation speed v w , according to L max =v w ·T response Calculate the maximum impact distance as the accident impact range; analyze the covered lane area based on the vehicle bounding box set, determine the number of blocked lanes, and obtain the probability of personal injury; analyze the driving speed v of each vehicle in the congested section of the highway based on the traffic flow prediction data and the real-time traffic flow data i, driving distance l i and the free section flow velocity v free ; Based on driving speed v i , driving distance l i and the free section flow velocity v free , according to V delay =∑ i (v free -v i )·l i Calculate the total delayed vehicle kilometers; based on the total delayed vehicle kilometers, determine the accident level through fuzzy comprehensive evaluation method.

[0141] To analyze the scope of an accident's impact, the congestion propagation speed in this embodiment reflects the rate at which traffic anomalies spread between upstream and downstream road sections. When an accident occurs, downstream vehicles decelerate, leading to increased density and decreased traffic flow. The ratio of the upstream and downstream traffic flow difference to the density difference quantifies the speed at which congestion spreads upstream. This embodiment combines traffic flow data to calculate the traffic flow in upstream and downstream sections using the average lane speed v and the upstream lane densities ρ1 and ρ2, using the formula q = v·ρ.

[0142] After determining the congestion propagation speed, it can be used as one of the parameters, combined with the processing time T response (Refers to the time from the occurrence of the accident to the arrival of the robot or rescue force at the scene and the start of evacuation), calculate the maximum upstream distance that the accident may affect, and thus obtain the scope of the accident impact.

[0143] For accident level assessment, three core indicators are mainly defined to assess the severity of the accident: probability of injury P inj , number of blocked lanes L block , total delayed vehicle kilometers V delay , construct the core indicators into a three-dimensional indicator space (P inj ,L block ,V delay ) and determine the accident level (I-IV) through fuzzy comprehensive evaluation method.

[0144] Among them, the probability of injury P inj Through camera image analysis, sensor data estimation, or passive detection driven by V2X data, by receiving OBU (on-board unit) data (such as speed, braking signal, steering angle) actively sent by the vehicle, combined with the communication coverage of the roadside RSU (roadside unit), a vehicle driving status network is built to determine the vehicle and personnel status when the accident occurs.

[0145] Number of blocked lanes L block , which can be determined by directly counting the number of lanes that are impassable due to accidents.

[0146] Total delayed vehicle kilometers Vdelay , reflecting the cumulative delay time of all affected vehicles. The larger the value, the more serious the impact of traffic disruption on traffic efficiency.

[0147] Since the three indicators have different dimensions (probability, number of lanes, and vehicle kilometers), fuzzy mathematics methods are used to transform them into a unified membership function. The comprehensive scores are calculated through weighted calculation and finally divided into grades I-IV accident levels.

[0148] This embodiment thus standardizes the accident classification process, providing a quantitative basis for resource scheduling and mitigation strategies. Furthermore, LoRa low-power wide-area network nodes can be deployed on both sides of the highway. Robots can transmit real-time monitoring data and receive cloud-based commands via LoRa communication modules. Fuzzy comprehensive evaluation results can be automatically synchronized to a cloud-based management platform, triggering emergency response plans corresponding to the accident level.

[0149] Step 270 , performing dynamic path planning based on the traffic flow prediction data, the real-time traffic flow data, and the accident information through a dynamic path planning algorithm.

[0150] For the description of step 270 , reference may be made to step 150 , which will not be described in detail in this embodiment.

[0151] Optionally, the dynamic path planning algorithm may include the following sub-steps based on the traffic flow prediction data, the real-time traffic flow data, and the accident information:

[0152] Sub-step 2701, based on the traffic flow prediction data, the real-time traffic flow data, the highway road data and the accident information, a road network model is performed, and a cost function is used to perform path optimization to obtain rescue path optimization information.

[0153] The rescue path optimization information is used to plan the optimal path from the starting point to the accident point for the robot.

[0154] In its implementation, this embodiment employs an improved dynamic path planning algorithm, using the previously acquired data as input to plan a rescue route. Specifically, when planning a rescue route, a corresponding road network model is first constructed using highway and real-time data as a benchmark. A cost function, implemented based on a modified A* algorithm, is then introduced into the road network model to generate an optimized rescue route.

[0155] In an optional embodiment, a road network model is performed based on the traffic flow prediction data, the real-time traffic flow data, the highway road data and the accident information, and a cost function is used to perform path optimization to obtain rescue path optimization information, which may specifically include: determining the real-time traffic flow q based on the traffic flow prediction data and the real-time traffic flow data.ij , and, determining the road capacity C based on the highway road data ij ; Real-time traffic q ij and road capacity C ij For input, according to Carry out road network modeling and use the cost function of the improved algorithm f(n)=g(n)+h(n)+γVar(c ij ) to optimize the path and obtain the rescue path optimization information.

[0156] Among them, the starting point i and the end point j of the rescue section are identified by the accident information, d ij is the distance of the road section, v ij (t) is the real-time average speed of the road section at time t, λ is the weight coefficient, g(n) represents the actual cost from the starting point to the current node, h(n) represents the heuristic estimated cost from the current node to the end point, γ represents the path stability factor, Var(c ij ) represents the variance of the cost of each section in the path, is the potential field gradient, Q i represents the repulsive force strength at the accident point, r represents the current vehicle position coordinate, r i represents the coordinates of the accident point, represents the unit vector pointing from the accident point to the vehicle.

[0157] In the process of optimizing the rescue path: first, the road network modeling is carried out, the starting point i and the end point j of the rescue are identified through the accident information, and the road network modeling is carried out based on the time-dependent cost function. The parameters input to the road network modeling include: real-time traffic q ij and road capacity C ij , road section distance d ij (representing the physical distance from road section i to j), the real-time average speed v of the road section at time t ij (t) and weight coefficient λ (used to balance the effects of speed and flow).

[0158] In addition, in road network modeling, the time cost term d is constructed ij / v ij (t) and the congestion cost term q ij (t) / C ij The time cost term represents the travel time of a road segment; lower speeds result in higher time costs. The congestion cost term represents the ratio of flow to capacity, reflecting the degree of congestion. A higher ratio (closer to 1) indicates more congested road segments and a higher cost. This allows for dynamic updates of road segment costs using real-time data to avoid choosing routes with severe congestion or slow traffic.

[0159] Then, based on the improved A* algorithm, the actual cost (accumulated segment cost) and estimated cost (such as Euclidean distance) are analyzed, and the path stability factor γVar(cij ), where Var(c ij ) represents the variance of the cost of each section in the path, which measures the fluctuation of the path cost; γ represents the adjustment coefficient, which is used to control the preference for stability.

[0160] Therefore, by improving the A* algorithm, the algorithm tends to choose paths with low total cost and small fluctuations in the cost of each road section, reducing the risk of path failure due to real-time traffic changes (such as temporary accidents) and solving the problem that the traditional A* algorithm ignores the volatility of path costs (such as a path containing a road section with a high risk of sudden congestion). This embodiment is based on a hybrid path planning method based on the improved A* algorithm and potential field theory, combined with real-time traffic flow prediction data, to dynamically generate the optimal diversion path. Through a distributed collaborative scheduling algorithm, multi-robot task allocation and path collision avoidance are achieved, ensuring rapid arrival at the accident point and minimizing interference with normal traffic flow.

[0161] Sub-step 2702 , based on the accident point corresponding to the accident information, diversion path planning is performed using the gradient descent method of potential field theory to generate a target diversion path.

[0162] In a specific implementation, this embodiment introduces a potential field gradient descent method to analyze the accident point, and then plans the diversion path to generate a target diversion path.

[0163] In an optional embodiment, the above-mentioned method of performing diversion path planning based on the accident point corresponding to the accident information by the gradient descent method of potential field theory to generate the target diversion path may include: analyzing the accident point by the gradient descent method of potential field theory, and Perform diversion path planning and generate target diversion path.

[0164] in, is the potential field gradient, Q i represents the repulsive force strength at the accident point, r represents the current vehicle position coordinate, r i represents the coordinates of the accident point, represents the unit vector pointing from the accident point to the vehicle.

[0165] In the process of optimizing the diversion path, the accident point is first regarded as the source point with "repulsion" based on the potential field theory gradient descent method, and the normal vehicles are regarded as the particles affected by the repulsion. The detour path is generated by the gradient descent method, and the potential field gradient is determined by analyzing the direction of reduced repulsion (i.e. guiding the vehicle away from the accident point). To avoid entering the accident zone, the robot iteratively calculates the gradient direction and dynamically adjusts the driving path to form a detour trajectory. This enables real-time traffic diversion and guidance. When the robot detects an accident, it transmits the detour route calculated by the potential field model to vehicles via roadside variable information signs or V2X technology, preventing traffic from converging into the accident area.

[0166] Sub-step 2703: generating target path planning information based on the optimal path and the target diversion path.

[0167] In this embodiment, the mobile robot generates a target planned path based on the generated rescue and diversion paths. When an accident occurs, the robot generates an optimized rescue path to ensure a rapid response. Furthermore, it generates a diversion path to simultaneously divert surrounding traffic, reducing the risk of secondary accidents. This creates a closed loop of "perception-action-diversion." This ensures that rescue vehicles arrive at the scene as quickly as possible while also preventing delays caused by sudden traffic jams during the rescue process.

[0168] Step 280: Control the mobile robot to perform a traffic diversion task on the lane of the highway based on the target path planning information.

[0169] The existing technology has a time delay from accident discovery to information transmission and disposal execution, making it difficult to divert traffic in a timely manner. This embodiment uses edge computing and real-time data processing technology to shorten the traffic flow prediction cycle and accident identification response time. Through localized decision-making and coordinated scheduling of mobile robots, a low-latency closed loop of "detection-analysis-disposal" is achieved, thereby improving the efficiency of highway emergency response. This embodiment uses mobile robots to achieve unmanned intelligent operation, reducing manpower dependence and safety risks: through automated robot detection and autonomous algorithm decision-making, the entire process from traffic flow prediction, accident identification to diversion and disposal is unmanned, replacing manual patrols and on-site intervention. It avoids manual work in high-risk scenarios such as bad weather and high traffic, significantly reducing manpower input; at the same time, it shortens the time period from accident discovery to disposal, improves monitoring efficiency and emergency response speed, and reduces the safety hazards of manual inspections at the source.

[0170] In an optional embodiment, after obtaining the target path planning information, the method may further include: when dispatching the mobile robot to the accident point corresponding to the accident information, obtaining the dispatch information x ij and scheduling time t ij , define the decision variables According to the decision variables, construct the objective function And based on the decision variables, construct constraints Optimize resource scheduling of mobile robots based on objective functions and constraints; after scheduling mobile robots to perform evacuation tasks, obtain disposal information, which includes disposal time and accident impact range; based on the disposal information, Calculate the recovery efficiency η; determine the reward value r based on the recovery efficiency η, according to Construct a reward function and continuously iterate and update the learning mechanism to update the execution strategy of the mobile robot when an accident occurs; T actualis the actual processing time, T predict is the estimated disposal time, L actual is the actual impact range of the accident, L actual is the estimated impact range of the accident, α is the learning rate, γ is the discount factor, (s, a) represents the current state and action, and s' represents the new state after the action is executed.

[0171] Regarding resource scheduling optimization, this embodiment can select the best combination from multiple available resources when an accident is detected. For example, a multi-objective resource scheduling model can be used to schedule a designated mobile robot to handle the accident. This model is used for resource scheduling optimization of the highway intelligent monitoring and evacuation robot system. It aims to solve the problem of how to reasonably allocate resources such as robots (such as patrol cars and rescue equipment) when an accident occurs, so as to achieve the dual goals of the fastest response speed and the lowest scheduling cost. ij Indicates whether resource i is dispatched to incident j. The parameter x ij The “select-or-no-select” logic of resource allocation can be simplified by binary variables, which is convenient for mathematical modeling. For example, x ij =1 Resource i is dispatched to incident j; x ij = 0 means resource i is not scheduled to incident j; scheduling time t ij It represents the time it takes for resource i to reach accident point j. It can be calculated by calculating the shortest time it takes for the resource to travel from its current location to the accident point during dynamic path planning.

[0172] The objective function is constructed using decision variables to obtain the objective function model. This model realizes the scientific and intelligent dispatch of highway emergency resources by quantifying decision variables and constraints. It is the core component of the system's "perception-decision-execution" closed loop. Among them, the model contains two minimization optimization objectives, namely weighted response time min∑w j t ij x ij and the total scheduling cost min∑c ij x ij The weighted response time represents the time it takes for all scheduled resources to arrive at each accident point, and the weight w is determined according to the accident priority. j , and perform weighted summation to obtain; the total scheduling cost is the sum of the total costs of scheduling all resources, c ij Represents the operating cost of resource i. By balancing response time and scheduling cost, the optimal scheduling solution is generated.

[0173] Then, the resource scheduling constraint is constructed, which can meet the minimum resource demand ∑x ij ≥N j and reaches the time limit t ij ≤T max, thereby ensuring that resources are available to handle accidents while ensuring the timely response to emergency accidents and avoiding secondary accidents or increased congestion due to slow scheduling.

[0174] In this embodiment, after each accident identification and detection, the efficiency of accident handling can be analyzed, and the effectiveness of the scheduling plan / strategy can be evaluated through the recovery efficiency η to provide feedback for updating the parameters of the resource scheduling model. actual / T predict It is used to measure the deviation between the actual processing time and the expected time, reflecting whether the system response efficiency meets the standard; L actual / L predict This measure measures the deviation between the actual impact range and the expected range, reflecting the system's ability to control the spread of an incident. Recovery efficiency is determined by comprehensively evaluating the system's performance in terms of both time efficiency and impact control. Based on this recovery efficiency, we can dynamically optimize the update path algorithm, adjust the incident level assessment algorithm, and optimize the resource scheduling model.

[0175] Furthermore, this embodiment can adopt a learning mechanism (such as Q-learning algorithm) to dynamically optimize the handling strategy of the highway intelligent monitoring and evacuation robot system. In the reward function, the Q value (Q(s,a)) obtained by the previous scheduling is combined with the immediate reward r and the discounted future reward as input to update the Q value. Among them, Q(s,a) represents the long-term expected value of performing action a in state s. The higher the Q value, the better the action; the reward r can be calculated based on the recovery efficiency index η, as direct feedback after the action is executed, to guide the strategy to optimize the short-term goal; the discounted future reward It represents the maximum expected value of all possible actions under the new state s', which is multiplied by the discount factor as the long-term reward to guide the strategy towards the long-term optimal development.

[0176] Therefore, this embodiment builds a self-learning framework for intelligent traffic diversion strategies to adapt to complex highway environments, such as high vehicle speeds, variable weather conditions (such as heavy rain, fog, ice and snow), and complex road conditions. This addresses the difficulty of traditional fixed signal control or manual intervention models in adapting to real-time traffic demands.

[0177] This embodiment utilizes an intelligent decision-making system based on deep learning and reinforcement learning algorithms, capable of processing massive amounts of sensory data in real time. Using traffic flow prediction models, it predicts congestion nodes and trends, and generates personalized traffic flow management solutions by integrating strategies such as shortest path planning, dynamic lane control, and emergency channel development. Furthermore, the robot supports real-time communication with cloud-based traffic management platforms, surrounding intelligent devices (such as variable information boards and traffic lights), and passing vehicles (via V2X technology), enabling vehicle-road-cloud-pedestrian collaborative control and forming a fully interconnected intelligent traffic flow management network.

[0178] This demonstrates that this embodiment implements a closed-loop system for treatment effect evaluation and strategy optimization based on reinforcement learning. By constructing a reward function based on historical treatment data and real-time feedback, the decision matrix is ​​continuously iteratively updated, enabling autonomous evolution and scenario-adaptive adaptation of the grooming strategy. Specifically, the reward value r is determined using the recovery efficiency η, forming an "evaluation-learning-optimization" closed loop that improves system efficiency. In complex environments, the learning mechanism can improve data reliability by adjusting sensor fusion weights (e.g., the priority of rain and fog cameras).

[0179] The advantages include: A. Autonomous strategy evolution without human intervention. The system automatically optimizes its handling strategy through continuous interaction with the environment to adapt to dynamic changes in highway scenarios (such as traffic flow fluctuations and sudden weather changes); B. Improved generalization ability. By accumulating experience in different scenarios, it can generate more universal diversion solutions and reduce dependence on specific conditions; C. Enhanced robustness. When encountering new types of accidents or sudden interference, the learning mechanism can gradually find reliable solutions by exploring new actions (such as trying to combine different diversion strategies).

[0180] Furthermore, building upon mobile robots, a collaborative drone-ground sensor solution can be constructed: low-cost sensor nodes (magnetic coils, RFID tags) deployed on the ground monitor basic traffic data, while drones periodically launch into the air for high-altitude video inspections. A LoRa wireless sensor network involves deploying LoRa low-power wide-area network nodes on both sides of the highway. Robots use LoRa communication modules to transmit real-time monitoring data and receive cloud-based commands.

[0181] In summary, the embodiment of the present application first uses a mobile robot to monitor the highway in real time, obtains traffic flow feature data, and obtains traffic flow forecast data for a specified time period through feature extraction, analysis and fusion. Then, based on the traffic flow forecast data and the real-time traffic flow data, accident identification and accident classification are performed. When an accident on the highway is identified, an algorithm is used to combine the real-time identification data to perform dynamic path planning, and based on the planned path, the mobile robot is controlled to perform traffic diversion tasks on the lanes of the highway. On the one hand, the present application uses a mobile monitoring and diversion robot, combined with multi-source sensors (such as radar and video) and dynamic deployment strategies, to achieve dynamic coverage and real-time monitoring of the entire highway section, eliminating the monitoring blind spots of fixed equipment and improving the comprehensiveness of road condition information collection. On the other hand, the robot performs automated monitoring based on the collected video data and combines it with intelligent algorithms to achieve accurate traffic flow prediction, automated accident identification and handling of the entire process, reducing manual intervention, improving monitoring efficiency and staff safety, and reducing operating costs. The existing technology for automated traffic flow monitoring on highways has low efficiency and cannot automatically identify / handle accidents.

[0182] It should be noted that, for the purpose of simple description, the method embodiments are expressed as a series of action combinations, but those skilled in the art should know that the embodiments of the present application are not limited to the described order of actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously.

[0183] like Figure 3 As shown, the embodiment of the present application further provides a highway traffic flow prediction and accident identification system 300, comprising:

[0184] The traffic flow characteristic data processing module 310 is configured to perform normalization processing on the acquired highway traffic flow characteristic data to obtain pre-processed target traffic flow characteristic data. The traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow on the highway in real time during dynamic patrol.

[0185] A feature convolution extraction module 320 is configured to perform temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and to perform spatial feature convolution processing on the acquired highway section data to obtain spatial feature information;

[0186] A feature splicing and fusion module 330 is configured to perform splicing and fusion based on the temporal feature information and the spatial feature information to obtain traffic flow prediction data for a specified time period;

[0187] An accident identification and classification module 340 is configured to perform accident identification and accident classification assessment based on the traffic flow prediction data and the real-time traffic flow data acquired in real time, and determine accident information, wherein the accident information includes accident analysis results, accident impact range, and accident level;

[0188] The path planning module 350 is used to perform dynamic path planning based on the traffic flow prediction data, the real-time traffic flow data and accident information through a dynamic path planning algorithm to obtain target path planning information, and control the mobile robot to perform the traffic diversion task on the lane of the highway based on the target path planning information.

[0189] Optionally, the accident identification and classification module is specifically used to implement accident identification and classification. During the accident identification process, it can integrate V2X data (such as braking signals and steering angles actively sent by vehicles) and receive vehicle-mounted unit data through roadside RSUs to build a vehicle driving status network, assist in determining abnormal driving behavior, and improve the accuracy of accident detection. The accident identification and classification module specifically includes:

[0190] A mutation detection submodule is used to perform traffic flow mutation detection based on the traffic flow prediction data in combination with real-time traffic flow data to obtain mutation analysis results;

[0191] A visual analysis submodule, configured to obtain video data of each lane of the highway for visual feature analysis to determine an accident analysis result when the mutation analysis result is a confirmed mutation result;

[0192] The accident determination submodule is used to perform an impact range analysis and an accident level assessment based on the traffic flow prediction data and the real-time traffic flow data when the accident analysis result is a determination of the occurrence of an accident, so as to obtain the accident impact range and the accident level.

[0193] Optional, path planning module, including:

[0194] a first path planning submodule, configured to perform road network modeling based on the traffic flow prediction data, the real-time traffic flow data, the highway road data, and the accident information, and to perform path optimization using a cost function to obtain rescue path optimization information, wherein the rescue path optimization information is used to plan an optimal path for the robot from the starting point to the accident site;

[0195] The second path planning submodule is used to perform diversion path planning based on the accident point corresponding to the accident information by using the gradient descent method of potential field theory to generate a target diversion path;

[0196] The third path planning submodule is configured to generate target path planning information based on the optimal path and the target diversion path.

[0197] It should be noted that the highway traffic flow prediction and accident identification system provided in the embodiments of the present application can execute the highway traffic flow prediction and accident identification method provided in any embodiment of the present application, and has the corresponding functions and beneficial effects of the execution method.

[0198] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0199] The foregoing is merely a list of specific embodiments of the present application, intended to enable those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the broadest scope consistent with the principles and novel features of the present application.

Claims

1. A method for predicting traffic flow and identifying accidents on a highway, characterized in that: include: Normalizing the acquired traffic flow characteristic data of the highway to obtain pre-processed target traffic flow characteristic data, wherein the traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow of the highway in real time during dynamic patrol; Performing temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and performing spatial feature convolution processing on the acquired highway section data to obtain spatial feature information; Based on the time feature information and the space feature information, the traffic flow prediction data for the specified time period is obtained by splicing and fusing; Perform accident identification and accident classification assessment based on the traffic flow prediction data and the real-time traffic flow data obtained in real time, and determine accident information, wherein the accident information includes accident analysis results, accident impact range, and accident level; Through a dynamic path planning algorithm, dynamic path planning is performed based on the traffic flow prediction data, the real-time traffic flow data and accident information to obtain target path planning information, and the mobile robot is controlled to perform a traffic diversion task on the lane of the highway based on the target path planning information.

2. The method according to claim 1, characterized in that Performing temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and performing spatial feature convolution processing on the acquired highway section data to obtain spatial feature information, including: Taking the target traffic flow characteristic data as input, according to Δ d Q(t)=Q(t)-Q(td) is used for differential processing to eliminate non-stationarity, and the preset autoregressive model is input. Perform convolution processing to extract temporal feature information; Analyze the monitoring points and the connection relationship between sections of the highway, perform abstract processing, and construct the abstract graph of the section G = (V, E); Take the nodes V and edges E in the road segment abstract graph as input, according to Perform graph convolution operation to obtain spatial feature information; Among them, Q(t) is the target traffic flow characteristic data after preprocessing, α p Autoregressive coefficient, β p is the moving average coefficient, ∈(t) is white noise, is an adjacency matrix with self-loops, is the degree matrix, W (l) is the weight of the lth layer, X (l) Node features.

3. The method according to claim 1, characterized in that Accident identification and accident classification assessment are performed based on the traffic flow prediction data and the real-time traffic flow data obtained in real time to determine accident information, including: Based on the traffic flow prediction data, combined with real-time traffic flow data, traffic flow mutation detection is performed to obtain mutation analysis results; When the mutation analysis result is a determined mutation result, obtaining video data of each lane of the highway to perform visual feature analysis to determine the accident analysis result; When the accident analysis result is to determine the occurrence of the accident, an impact range analysis and an accident level assessment are performed based on the traffic flow prediction data and the real-time traffic flow data to obtain the accident impact range and the accident level.

4. The method according to claim 3, characterized in that Based on the traffic flow prediction data, combined with real-time traffic flow data, traffic flow mutation detection is performed to obtain mutation analysis results, including: Based on the traffic flow prediction data and the real-time traffic flow data, the average speed v(t) of each vehicle in the lane at time t and the density of vehicles in the lane are analyzed as ρ(t), and the mutation judgment formula is used. Compute mutation analysis results; Among them, μ is the historical mean, σ is the standard deviation, w1, w2 and w3 are weight coefficients. When the score is greater than the preset threshold, it is determined that an accident may have occurred and a certain mutation result is obtained.

5. The method according to claim 3, characterized in that Obtain video data from each lane of the highway for visual feature analysis to determine the accident analysis results, including: The acquired video data is detected by the target detection algorithm to obtain the vehicle bounding box set B = {b1,…,b n }; Based on the vehicle bounding box set B={b1,…,b n } to perform trajectory analysis and obtain the vehicle attitude angle set S = {s1,…,s n }; The vehicle attitude angle set S = {s1,…,s n } is input, according to Analyze key abnormal features and obtain collision angle variance According to the vehicle bounding box set B={b1,…,b n Analyze the pixels of the vehicle bounding box to obtain the total pixels b, and mark the pixels covered by the vehicle bounding box as the vehicle area, and mark the uncovered pixels as the non-vehicle area, and obtain the number of pixels a in the non-vehicle area; Take the total pixels b and the number of pixels a as input, according to R debris =ab analyzes the key abnormal characteristics and obtains the fragment area ratio R debris ; Based on collision angle variance and the debris area ratio R debris , according to the decision function Conduct accident assessment and obtain accident information; Among them, P accident When ≥0.8, it is determined that an accident has occurred and the accident result is obtained. k1 and k2 are weight coefficients, b i Represents the coordinates of the rectangular bounding box of the vehicle in the i-th frame of the image, s i Indicates the driving direction or body posture angle of the i-th vehicle in the image.

6. The method according to claim 5, characterized in that An impact range analysis and accident level assessment are performed based on the traffic flow prediction data and the real-time traffic flow data to obtain the accident impact range and accident level, including: Analyzing the upstream traffic flow q1 and the downstream traffic flow q2 in the highway according to the traffic flow prediction data and the real-time traffic flow data; Based on the upstream traffic flow q1 and downstream traffic flow q2, according to v w =q2-q1 / ρ2-ρ1, calculate the congestion propagation speed v w ; Based on the congestion propagation speed v w , according to L max =v w ·T response Calculate the maximum impact distance as the accident impact range; Based on the vehicle bounding box set, analyze the covered lane area, determine the number of blocked lanes, and obtain the probability of personal injury; Based on the traffic flow prediction data and the real-time traffic flow data, the driving speed v of each vehicle in the congested section of the highway is analyzed. i , driving distance l i and the free section flow velocity v free ; Based on the driving speed v i , driving distance l i and the free section flow velocity v free , according to V delay =∑i(v free -v i )·l i Calculate the total delayed vehicle kilometers; Based on the total delayed vehicle kilometers, the accident level is determined by fuzzy comprehensive evaluation method.

7. The method according to claim 1, characterized in that Dynamic path planning is performed based on the traffic flow prediction data, the real-time traffic flow data, and accident information using a dynamic path planning algorithm, including: Performing road network modeling based on the traffic flow prediction data, the real-time traffic flow data, the highway road data, and the accident information, and performing path optimization using a cost function to obtain rescue path optimization information, wherein the rescue path optimization information is used to plan an optimal path for the robot from the starting point to the accident point; Based on the accident point corresponding to the accident information, diversion path planning is performed using the gradient descent method of potential field theory to generate a target diversion path; Based on the optimal path and the target diversion path, target path planning information is generated.

8. The method according to claim 7, characterized in that A road network model is performed based on the traffic flow prediction data, the real-time traffic flow data, the highway road data, and the accident information, and a cost function is used to perform path optimization to obtain rescue path optimization information, including: Based on the traffic flow prediction data and the real-time traffic flow data, the real-time traffic flow q is determined. ij , and, determining the road capacity C based on the highway road data ij ; Real-time traffic q ij and road capacity C ij For input, according to Carry out road network modeling and use the cost function of the improved algorithm f(n)=g(n)+h(n)+γVar(c ij ) perform path optimization to obtain rescue path optimization information; Based on the accident point corresponding to the accident information, the diversion path planning is performed by the gradient descent method of potential field theory to generate the target diversion path, including: using the gradient descent method of potential field theory to analyze the accident point, according to Perform diversion path planning and generate target diversion paths; Among them, the starting point i and the end point j of the rescue section are identified by the accident information, d ij is the distance of the road section, v ij (t) is the real-time average speed of the road section at time t, λ is the weight coefficient, g(n) represents the actual cost from the starting point to the current node, h(n) represents the heuristic estimated cost from the current node to the end point, γ represents the path stability factor, Var(c ij ) represents the variance of the cost of each section in the path, ▽φ is the potential field gradient, Q i represents the repulsive force strength at the accident point, r represents the current vehicle position coordinate, r i represents the coordinates of the accident point, represents the unit vector pointing from the accident point to the vehicle.

9. The method according to any one of claims 1 to 8, characterized in that After obtaining the target path planning information, it also includes: When dispatching the mobile robot to the accident point corresponding to the accident information, the dispatch information x is obtained. ij and scheduling time t ij , define the decision variables According to the decision variables, construct the objective function And construct constraints based on decision variables Optimize resource scheduling of mobile robots based on objective functions and constraints; After dispatching the mobile robot to perform the evacuation task, obtaining handling information, wherein the handling information includes handling time and accident impact range; Based on the disposal information, Calculate the recovery efficiency η; The reward value r is determined based on the recovery efficiency η, according to Construct a reward function and continuously iterate the learning mechanism to update the mobile robot's execution strategy when an accident occurs; Among them, w j is the weight determined based on the accident level, T actual is the actual processing time, T predict is the estimated disposal time, L actual is the actual impact range of the accident, L actual is the estimated impact range of the accident, α is the learning rate, γ is the discount factor, (s, a) represents the current state and action, and s' represents the new state after the action is executed.

10. A highway traffic flow prediction and accident identification system, characterized in that: include: a traffic flow characteristic data processing module for performing normalization processing on the acquired traffic flow characteristic data of the highway to obtain pre-processed target traffic flow characteristic data, wherein the traffic flow characteristic data is data obtained by the mobile robot monitoring the traffic flow of the highway in real time during dynamic patrol; a feature convolution extraction module, configured to perform temporal feature convolution processing on the target traffic flow feature data to obtain temporal feature information, and to perform spatial feature convolution processing on the acquired highway section data to obtain spatial feature information; A feature splicing and fusion module, configured to perform splicing and fusion based on the temporal feature information and the spatial feature information to obtain traffic flow prediction data for a specified time period; An accident identification and classification module is used to identify and classify accidents based on the traffic flow prediction data and the real-time traffic flow data obtained in real time, and determine accident information, wherein the accident information includes accident analysis results, accident impact range, and accident level; The path planning module is used to perform dynamic path planning based on the traffic flow prediction data, the real-time traffic flow data and accident information through a dynamic path planning algorithm to obtain target path planning information, and control the mobile robot to perform a traffic diversion task on the lane of the highway based on the target path planning information.

Images