Edge calculation traffic light emergency control method and device for sudden traffic event

Abstract

The invention provides an edge calculation traffic light emergency control method and device for a sudden traffic event. The method comprises the following steps: acquiring real-time traffic flow information of a target intersection; determining an emergency response area of the target intersection based on the position identifier of the sudden traffic event, and extracting a dynamic offset between a historical traffic flow feature and a current traffic flow feature in the emergency response area through an edge computing node; inputting the dynamic offset into a pre-trained traffic signal adaptive model, generating a traffic signal lamp control parameter set corresponding to the emergency response area, and sending a control instruction sequence to the traffic signal lamp of the target intersection according to the traffic signal lamp control parameter set, and the execution priority of the control instruction sequence is dynamically adjusted based on the real-time change trend of the vehicle density distribution data and the pedestrian movement track data. According to the invention, the intersection passing efficiency and the safety guarantee capability of traffic participants under the sudden traffic event can be improved.

Description

Technical Field

[0001] The present invention relates to the fields of data processing and machine learning, and particularly to an edge computing traffic light emergency control method and device for sudden traffic events. Background Art

[0002] In the field of traffic signal control, fixed timing schemes or adaptive control methods based on simple traffic flow detection are commonly used in the prior art. For example, some systems collect the vehicle queue lengths of a single lane through geomagnetic sensors or cameras and adjust the green light duration based on preset thresholds; other systems rely on the statistical laws of historical traffic flow data to generate periodic signal timing schemes. However, in the event of sudden traffic events (such as traffic accidents, road obstacles), such methods often struggle to effectively handle abnormal fluctuations in traffic flow. Specifically, fixed timing schemes are unable to sense local traffic paralysis caused by sudden events, and adaptive control based on single-lane traffic flow data lacks multi-dimensional correlation analysis of pedestrian activities, sudden changes in vehicle density, and the scope of event impact, which is likely to cause the following problems: First, the signal switching frequency and phase duration cannot match the vehicle evacuation requirements in the sudden event area, resulting in congestion quickly spreading to adjacent intersections; second, the competition for pedestrian passage requirements and vehicle right of way intensifies, posing safety hazards; third, models driven by historical data are difficult to capture dynamic traffic states in a timely manner, with significant response delays. In addition, signal control decisions under traditional centralized cloud computing architectures are limited by data transmission delays and are difficult to provide real-time decision support for emergency responses to sudden events. Therefore, there is an urgent need for a traffic light emergency control method that can integrate multi-source traffic data, quickly identify the scope of event impact, and dynamically optimize signal control parameters to improve intersection traffic efficiency and the safety guarantee ability of traffic participants under sudden traffic events. Summary of the Invention

[0003] The present invention provides an edge computing traffic light emergency control method and device for sudden traffic events.

[0004] According to one aspect of the present invention, there is provided an edge computing traffic light emergency control method for sudden traffic events, the method comprising:

[0005] Obtain real-time traffic flow information of a target intersection, the real-time traffic flow information including vehicle density distribution data, pedestrian activity trajectory data, and the location identifier of a sudden traffic event;

[0006] Based on the location identifier of the sudden traffic event, determine the emergency response area of the target intersection, and extract the dynamic offset between the historical traffic flow characteristics and the current traffic flow characteristics within the emergency response area through an edge computing node;

[0007] Input the dynamic offset into a pre-trained traffic signal adaptive model to generate a set of traffic signal control parameters corresponding to the emergency response area, where the set of traffic signal control parameters includes a phase switching frequency parameter, a green light duration parameter, and a direction passing priority parameter;

[0008] According to the set of traffic signal control parameters, send a control instruction sequence to the traffic signals at the target intersection, and dynamically adjust the execution priority of the control instruction sequence based on the real-time change trends of the vehicle density distribution data and the pedestrian activity trajectory data.

[0009] According to another aspect of the present invention, there is provided a control device, including:

[0010] At least one processor;

[0011] And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method described above.

[0012] The present invention has at least the following beneficial effects:

[0013] The edge - computing traffic - light emergency control method for sudden traffic events provided by the present invention can comprehensively capture the traffic - state changes at the target intersection and accurately locate the direct - impact range of sudden events by obtaining the vehicle - density distribution data, pedestrian - activity trajectory data, and the location identifier of sudden traffic events in the real - time traffic - flow information of the target intersection. Based on the extraction of the dynamic offset between the historical traffic - flow characteristics and the current traffic - flow characteristics, it can effectively quantify the real - time impact degree of sudden events, and quickly generate response strategies in combination with the local - processing ability of edge - computing nodes. By means of a pre - trained traffic - signal adaptive model, the dynamic offset is converted into a set of parameters including phase - switching frequency parameters, green - light duration parameters, and direction - passing priority parameters, ensuring the scientific nature and multi - objective optimization ability of signal - control parameters. According to the real - time change trends of vehicle density and pedestrian trajectories, the execution priority of control instructions is dynamically adjusted, enabling the signal control to adapt to the dynamic evolution characteristics of sudden events. In this way, the vehicle - density distribution data can reflect the occupancy of road resources, the pedestrian - activity trajectory data can capture the pedestrian - passing demands, the location identifier of sudden events can clarify the core area of emergency response, and the collaborative analysis of multi - dimensional data can enhance the global perception ability of traffic situations. The calculation of the dynamic offset between historical and current traffic - flow characteristics can identify the abnormal - fluctuation patterns of traffic flow caused by sudden events, providing key inputs for model decision - making. The traffic - signal adaptive model maps complex traffic states into executable signal - control instructions through parameterized control strategies, taking into account both vehicle - passing efficiency and pedestrian - safety guarantee. The dynamic - priority adjustment mechanism ensures the flexibility and timeliness of control strategies by real - time feedback on the changes in the trends of vehicle and pedestrian flows, thereby achieving a rapid recovery of intersection - passing capacity and an effective resolution of conflict risks in sudden traffic events.

[0014] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings exemplarily show embodiments and form a part of the specification, and are used together with the written description of the specification to explain the exemplary embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0016] Figure 1 FIG. shows a schematic diagram of the application scenario of the edge - computing traffic - light emergency control method for sudden traffic events according to an embodiment of the present invention.

[0017] Figure 2The flowchart of an edge - computing traffic - light emergency control method for sudden traffic events according to an embodiment of the present invention is shown.

[0018] Figure 3 The schematic composition diagram of a control device according to an embodiment of the present invention is shown. Detailed implementation manners

[0019] The following describes exemplary embodiments of the present invention with reference to the accompanying drawings. Various details of the embodiments of the present invention are included to facilitate understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present invention. Similarly, for the sake of clarity and conciseness, descriptions of well - known functions and structures are omitted below.

[0020] Figure 1 The schematic diagram of an application scenario provided according to an embodiment of the present invention is shown. The application scenario includes one or more traffic sensing devices 101, a control device 120, and one or more communication networks 110 that couple the one or more traffic sensing devices 101 to the control device 120. The traffic sensing device 101 can be configured to execute one or more application programs.

[0021] In Figure 1 the configuration shown, the control device 120 can include one or more components that implement the functions executed by the control device 120. These components can include software components, hardware components, or a combination thereof that can be executed by one or more processors. The user operating the traffic sensing device 101 can sequentially utilize one or more application programs to interact with the control device 120 to utilize the services provided by these components. It should be understood that various different system configurations are possible, which can be different from the application scenario. Therefore, Figure 1 is an example of a system for implementing the various methods described herein and is not intended to be limiting.

[0022] The traffic sensing device 101 can be various types of sensors or a combination thereof. For example, millimeter - wave radar, inductive loop sensor, image sensor, infrared sensor, thermal imager, temperature sensor, etc.

[0023] The control device 120 may include one or more general-purpose computers, dedicated server computers (e.g., PC (personal computer) servers, UNIX servers, midrange servers), blade servers, mainframes, server clusters, or any other suitable arrangement and / or combination. The control device 120 may include one or more virtual machines running a virtual operating system, or other computing architectures involving virtualization (e.g., one or more flexible pools of logical storage devices that can be virtualized to maintain virtual storage devices of the server). In various embodiments, the control device 120 may run one or more services or software applications that provide the functions described below.

[0024] The application scenario of the present invention may also include one or more databases 130. In certain embodiments, these databases may be used to store data and other information. For example, one or more of the databases 130 may be used to store such as real-time traffic flow information, historical traffic flow characteristics. The databases 130 may reside in various locations. For example, the databases used by the control device 120 may be local to the control device 120, or may be remote from the control device 120 and may communicate with the control device 120 via a network-based or dedicated connection. The databases 130 may be of different types. In certain embodiments, the databases used by the control device 120, for example, may be relational databases. One or more of these databases may store, update, and retrieve data to and from the databases in response to commands.

[0025] Please refer to Figure 2 , which is a flowchart of the edge computing traffic light emergency control method for sudden traffic events provided by an embodiment of the present invention, including the following steps S100 to S400:

[0026] Step S100: Obtain the real-time traffic flow information of the target intersection, where the real-time traffic flow information includes vehicle density distribution data, pedestrian activity trajectory data, and the location identifier of the sudden traffic event.

[0027] Specifically, real-time traffic flow information refers to a set of traffic status data collected in real time by sensor networks, camera equipment, and on-board communication units deployed at target intersections. Vehicle density distribution data is specifically manifested as the number of vehicles in each lane per unit time, vehicle type classification, and the degree of aggregation in different areas of the intersection. For example, the dwell time and moving speed of vehicles on the lane are detected by geomagnetic sensors, and combined with the traffic images captured by the camera, it is calculated that 12 small passenger cars and 3 large trucks pass through the north-south main lane every minute, and the right turn lane on the west side is congested, resulting in a queue length of 80 meters. Pedestrian activity trajectory data tracks the spatiotemporal distribution characteristics of pedestrian crossing behavior through thermal imaging cameras and pedestrian re-identification algorithms. For example, 25 pedestrians are detected moving from east to west at an average speed of 1.2 meters per second per minute at the southeast corner crosswalk, and pedestrians are stranded on the northwest safety island. The location identification of sudden traffic incidents is the coordinate positioning information generated by multi-source data fusion technology. When the traffic accident detection system identifies abnormal parking behavior or airbag triggering signals, it will combine the on-board GPS data with the timestamp of the roadside unit to determine the precise geographic coordinates of the incident. For example, the longitude and latitude (116.403°E, 39.914°N) is marked as the point of a three-car rear-end collision and mapped to the three-dimensional spatial model of the intersection through the digital twin model. The implementation of this step relies on the edge computing gateway to pre-process the raw data at the millisecond level, including noise filtering, data alignment, and format standardization, to ensure that the subsequent analysis module obtains a high-confidence input data source.

[0028] As an implementation manner, the location identifier of the traffic emergency may be obtained by the following steps:

[0029] Step S110: receiving real-time video stream data uploaded by the monitoring device at the target intersection, and performing moving target detection on each frame image in the real-time video stream data.

[0030] Among them, the real-time video stream data comes from high-definition network cameras deployed on the four-way poles at the target intersection. For example, it transmits an H.264 encoded video stream with a resolution of 1920×1080 at a rate of 25 frames per second. The moving object detection can adopt an improved YOLOv5 algorithm to achieve real-time processing on the edge computing node. For example, in the video frame at the T-th second, three moving objects are detected on the north-south lane: a white sedan with a license plate number of Shanghai A12345 (bounding box coordinates [450, 320, 580, 400]), an unlicensed new energy logistics vehicle (bounding box coordinates [620, 280, 750, 380]), and a group of pedestrians at the southeast crosswalk (bounding box coordinates [200, 600, 300, 720]). The detection process synchronously outputs the class confidence of the moving object. For example, the average confidence of the vehicle object reaches 98.7%, and the detection confidence of the pedestrian group is 94.2%. The detection results are stored in a structured data format, including timestamp, target ID, coordinate information, and estimated motion vector.

[0031] Step S120: Extract the position change sequence of the detected moving object and calculate the position offset between adjacent frames of the moving object.

[0032] Exemplarily, the position change sequence can be generated by a target tracking algorithm. For example, Kalman filter tracking is implemented for the white sedan, and the centroid coordinate sequence of five consecutive frames is (512, 360) → (515, 362) → (518, 365) → (522, 368) → (525, 370). The Euclidean distance offsets between adjacent frames are calculated as 3.6 pixels, 4.2 pixels, 4.5 pixels, and 3.6 pixels respectively. The position offset, for example, can be calculated by introducing perspective transformation correction, and the pixel offset is converted into an actual physical displacement according to the camera calibration parameters. For example, it is detected that the displacement of the new energy logistics vehicle between frames at the (T + 0.4)-th second reaches 12 pixels, and the corresponding actual displacement is 4.8 meters after coordinate transformation. The abnormal displacement threshold is set considering the road speed limit and object type. For example, the reasonable displacement threshold for a car in a congested section is 0 - 3 meters per second, and an alarm is triggered when it is detected that a certain target has a displacement of 5.2 meters per second for three consecutive frames.

[0033] Step S130: If the position offset continuously exceeds the preset abnormal displacement threshold, it is determined that the area where the moving object is located is a suspected sudden traffic event area.

[0034] Exemplarily, the preset abnormal displacement threshold can adopt a dynamic adjustment mechanism. For example, under rainy and foggy weather conditions, the vehicle threshold is lowered to 3.5 m / s, and the pedestrian threshold is set to 1.8 m / s. When it is detected that a white car has three displacement overlimit situations (4.8 m, 5.1 m, 5.3 m) within 0.8 seconds, the system automatically demarcates an area with a radius of 15 m centered on this car as a suspected area, and the geographical coordinate range is marked as X: 450 - 580, Y: 300 - 420. The determination process implements multi-source verification. For example, when analyzing the sonar sensor data in this area synchronously and detecting an abnormal sound of 106 dB, the suspected level is raised to level one.

[0035] Step S140: Send a high-resolution capture instruction to the monitoring device through the edge computing node to obtain multi-angle detailed images of the suspected sudden traffic event area.

[0036] The high-resolution capture instruction can trigger the camera to switch to the 3840×2160 resolution mode, and perform three consecutive captures on the suspected area at a frame rate of 60 frames per second, while activating the collaborative shooting mechanism of adjacent poles. For example, the main camera obtains a front-view image (azimuth angle 0°), auxiliary camera 1 shoots a 45° side-view image, and auxiliary camera 2 captures a top-view image. The multi-angle images are aligned through the time synchronization protocol. For example, three-view images are synchronously captured at 18:15:23.455 to form a stereoscopic observation data set. The image enhancement algorithm automatically adjusts the exposure parameters. For example, in a backlight scenario, the gain value is increased to 18 dB to ensure that the details of scattered objects are clearly visible.

[0037] Step s150: Classify the event type of the multi-angle detailed images. If the classification result is vehicle collision or road obstacle, mark the suspected sudden traffic event area as the location identifier of the sudden traffic event.

[0038] Exemplarily, the event type classification model can adopt a multi-modal fusion architecture. For example, the front-view image is input into ResNet-50 to extract global features, the side-view image is segmented into vehicle components through the Point Rend algorithm, and the top-view image is used to measure the distribution area of scattered objects. The cross-view feature alignment is implemented during the classification process. For example, in the front-view image, fragments of the front bumper are identified (confidence level 92%), in the top-view image, the scattered area is measured to be 3.2 square meters, and in the side-view image, the length of the braking marks of the adjacent vehicle is 8.7 meters. It is comprehensively determined as a vehicle collision event. When the classification confidence level exceeds 85%, taking the geometric center point (X: 515, Y: 370) of the suspected area as the reference, a geographical fence with a radius of 20 m is generated as the location identifier, and the corresponding WGS84 longitude and latitude (121.4732°E, 31.2315°N) after coordinate system conversion, with the accuracy error controlled within ±0.3 m.

[0039] As an implementation manner, in step s150, classifying the event types of the multi-angle detail images may specifically include:

[0040] Step s151: Input the multi-angle detail images into a pre-trained event classification model to extract the key object contour features and scene context features in the images.

[0041] Exemplarily, the pre-trained event classification model can be improved based on the EfficientNet-B7 architecture, and its multi-scale feature pyramid can process inputs of different resolutions synchronously. The key object contour features are extracted through a deformable convolutional layer. For example, the contour of glass fragments distributed radially is recognized in the front view, and the dimension of the feature vector is 1024 dimensions. The scene context features are captured by a spatial attention module. For example, indirect evidences such as the long-on state of the brake lights of adjacent vehicles (lasting for more than 0.8 seconds) and the irregular shaking pattern of roadside trees are detected. The feature fusion layer stitches the contour features of the front view with the topological features of the top view to form a 2560-dimensional joint feature vector, which is input into the classification decision layer.

[0042] Step S152: Identify the visual patterns of vehicle component debris, road surface cracks, or pedestrian falling actions based on the key object contour features.

[0043] Exemplarily, the identification of vehicle component debris can be implemented by a component-level segmentation network. For example, 16 irregular metal fragments are detected, among which 3 fragments have brand logo features (confidence level 89%), and 5 fragments show high-temperature deformation features. The detection of road surface cracks combines U-Net segmentation and Inception-v3 classification to identify a cracked damage with a length of 2.3 meters and a width of 0.15 meters, and the included angle between the crack direction and the vehicle sliding trajectory is 32°. The identification of pedestrian falling actions is realized through a pose estimation model, and it is detected that the spatial distribution of human key points conforms to the falling pose template (the height difference between the hip joint and the ankle joint < 30 cm) and lasts for more than 5 seconds.

[0044] Step S153: Analyze the sudden speed change of adjacent vehicles or the collective turning behavior of the movement direction of the pedestrian group based on the scene context features.

[0045] Exemplarily, the analysis of the sudden speed change of adjacent vehicles can adopt the dense optical flow method. For example, it is detected that the speed of the rear vehicle drops suddenly from 36 km / h to 8 km / h within 0.4 seconds, and the deceleration reaches 7 m / s 2, exceeding the normal braking threshold. The collective pedestrian turning behavior is identified through a trajectory clustering algorithm. For example, 83% of the pedestrian trajectory directions are detected to turn from due west to southwest within 10 seconds on the east side of the accident point, with an average turning angle of 45°, forming a statistically significant group behavior pattern. The spatio-temporal correlation analysis of context features and target features uses a graph neural network to establish a causal relationship chain among elements such as debris distribution, vehicle displacement, and pedestrian turning.

[0046] Step S154: If the vehicle component debris is identified and there is a speed mutation, then determine that the event type is a vehicle collision.

[0047] Exemplarily, the sufficient conditions for a vehicle collision, for example, include: the number of debris ≥ 5 pieces, and the distribution area ≥ 2m 2 , while the average deceleration of adjacent vehicles ≥ 4m / s 2 . For example, 12 pieces of debris are detected to be fan-shaped distributed (area 3.8m 2 ), and the decelerations of the three vehicles behind are 6.2m / s 2 , 5.8m / s 2 , 7.1m / s 2 , and the estimated collision energy reaches 82kJ, meeting the standard of a medium-sized collision event. The determination process implements multi-evidence chain verification. When the video analysis shows the airbag deployment signal of the involved vehicle, the classification confidence is increased to 98%.

[0048] Step S155: If the road crack is identified and there is a collective turning behavior in the direction of the pedestrian group movement, then determine that the event type is a road obstacle.

[0049] Exemplarily, the determination standard for a road obstacle, for example, is that the crack width ≥ 5cm, and the group turning ratio ≥ 60%. For example, a transverse crack with a width of 8cm and a depth of 12cm is detected, and 82% of the pedestrian trajectories in the west-to-east direction deviate by an average of 38° at a distance of 5 meters from the obstacle, accompanied by a 23% reduction in the walking frequency. The control device 120 verifies in combination with the municipal facility database. When there is no registered construction plan in this area, it is immediately marked as a sudden road obstacle event. The classification result triggers a work order for the road maintenance system, and the obstacle coordinate information is updated synchronously on the navigation map.

[0050] Step S200: Based on the location identifier of the sudden traffic event, determine the emergency response area of the target intersection, and extract the dynamic offset between the historical traffic flow characteristics and the current traffic flow characteristics within the emergency response area through the edge computing node.

[0051] Exemplarily, the emergency response area is a dynamic influence range delimited with the location identifier of the sudden traffic incident as the center, comprehensively considering the road topology, traffic control rules, and vehicle diffusion model. For example, when a traffic accident occurs 50 meters north of an intersection, the emergency response area will cover a range of 200 meters of the northbound through lane, the adjacent left-turn lane, and the associated crosswalk area. The historical traffic flow characteristics refer to the typical traffic patterns in this area stored in the edge database during the same time period and under the same weather conditions, including parameters such as vehicle passing speed, pedestrian crossing frequency, signal light cycle, etc. For example, the average traffic flow in this area during the evening rush hour on Friday is 45 vehicles per minute, and the peak pedestrian crossing reaches 60 person-times per minute. The current traffic flow characteristics are the immediate data captured by the real-time sensor network. For example, the accident causes a 300-meter congestion zone where vehicles in the northbound lane are stagnant, and the number of vehicles detouring eastward surges to 65 vehicles per minute. The dynamic offset is calculated by comparing the difference between the historical characteristics and the current characteristics, and the dynamic time warping algorithm is used to quantify the change in the traffic flow pattern. For example, the traffic flow in the northbound lane decreases by 82% year-on-year, the traffic flow in the eastbound right-turn lane increases by 140% year-on-year, and the number of unplanned pedestrian detour routes increases by 3. The edge computing node performs spatio-temporal correlation analysis during this process. For example, the standard deviation of the vehicle speed within 50 meters around the accident point is increased from the historical value of 8 km / h to the current value of 32 km / h, which is used as a quantitative indicator of the traffic flow disorder degree.

[0052] As an implementation manner, in step S200, the edge computing node extracts the dynamic offset between the historical traffic flow characteristics and the current traffic flow characteristics within the emergency response area, which may specifically include:

[0053] Step S210: Retrieve the historical traffic flow characteristics within the emergency response area within a preset time window from the local database of the edge computing node, where the historical traffic flow characteristics include the historical average vehicle speed, the historical pedestrian waiting duration, and the historical congestion index.

[0054] Historical traffic flow characteristics refer to the set of traffic state statistical indicators continuously stored by edge computing nodes, which are derived from the aggregated analysis results of sensor data at the target intersection during past periodic time periods. The preset time window is set according to the periodic law of traffic patterns. Exemplarily, for example, for sudden traffic events during the evening rush hour on weekdays, the historical data set from 17:00 to 19:00 every day in the past four weeks in this area is retrieved. The historical average vehicle speed is calculated through the time difference of vehicle passing times collected by geomagnetic sensors and license plate recognition cameras, and specifically manifests as the weighted average of the vehicle movement speeds in each lane during this period. For example, the historical average vehicle speed of the north-south straight lane is 32 km / h, and that of the east-west left-turn lane is 18 km / h. The historical pedestrian waiting time is generated by synchronizing data statistics from pedestrian detection cameras and signal light states, and is reflected as the time interval from when a pedestrian triggers the request button to when they obtain the green light passing permission. For example, the average historical pedestrian waiting time at the southeast crosswalk is 45 seconds, with a standard deviation of 8 seconds. The historical congestion index integrates multi-dimensional data such as lane occupancy rate, vehicle queue length, and number of parking times, and adopts a normalized scoring system of 0-10. For example, when it is detected that the vehicle queue at the north entrance exceeds 200 meters continuously for 5 minutes, its historical congestion index is marked as 8.7. The local database uses time series compression storage technology to ensure millisecond-level response to query requests. For example, when a traffic accident occurs at 18:15, the edge computing node immediately retrieves the historical feature data set at the same minute-level time slice (18:14:30 to 18:15:30) in the past four weeks at this coordinate point.

[0055] Step S220: Perform time dimension alignment processing on the current traffic flow characteristics so that the time stamp of the current traffic flow characteristics is consistent with the time window range of the historical traffic flow characteristics.

[0056] The time - dimension alignment process refers to matching the traffic flow data collected in real - time with the time granularity of historical data on the time axis to eliminate the deviation caused by inconsistent data sampling intervals. For example, the historical traffic flow characteristics store the average value with an aggregation granularity of 5 minutes, while the original data of the current traffic flow characteristics is updated once per second. The current data needs to be resampled into statistics at 5 - minute intervals through the moving - window averaging method. During this process, if a current traffic accident occurs at 18:15:03, the complete historical time - window data from 18:10:00 to 18:15:00 needs to be extracted, and the current time - slice is extended to 18:15:00 to 18:20:00 for alignment. For incomplete time - slice data, the linear interpolation method is used to fill in the missing values. For example, when only the first 3 minutes of data are collected in the time - slice from 18:15:00 to 18:20:00, the values for the last 2 minutes are predicted based on the changing trend of the average vehicle speed in the first 3 minutes. The dynamic time warping algorithm is applied to the alignment of non - uniformly sampled data. For example, the discrete event timestamps of pedestrian waiting durations are mapped to the fixed - time grid of historical data to ensure that the number of samples in each statistical period is the same. After data alignment, the current traffic flow characteristics and historical traffic flow characteristics have a comparable time - reference. For example, the accident trigger moment at 18:15:03 is aligned to the starting point of the 18:15:00 time - slice in historical data.

[0057] Step S230: Calculate the speed - difference coefficient between the historical average vehicle speed and the current average vehicle speed, and the waiting - duration difference coefficient between the historical pedestrian waiting duration and the current pedestrian waiting duration.

[0058] The speed difference coefficient is an indicator that reflects the traffic flow obstruction by quantifying the deviation degree between the historical benchmark and the real-time state. For example, its calculation method is the Euclidean distance between the historical average vehicle speed and the current average vehicle speed divided by the historical standard deviation. For example, when the historical average vehicle speed is 32 km / h and the standard deviation is 4.5 km / h, and the current average vehicle speed suddenly drops to 9 km / h, the speed difference coefficient is |32 - 9| / 4.5 ≈ 5.11, indicating a serious abnormal congestion in this lane. The waiting time difference coefficient adopts the relative difference percentage algorithm. For example, its calculation method is (the current pedestrian waiting time - the historical pedestrian waiting time) / the historical pedestrian waiting time × 100%. For example, when the historical pedestrian waiting time is 45 seconds and the current waiting time is extended to 120 seconds due to traffic control, the waiting time difference coefficient is (120 - 45) / 45 × 100% ≈ 166.7%, indicating a significant reduction in the pedestrian passing efficiency. During this process, the edge computing node synchronously calculates the difference coefficient matrix of each lane. For example, the speed difference coefficient of the east-west straight lane is 2.3, the left-turn lane is 5.1, and the pedestrian waiting difference coefficient is 82%, forming a multi-dimensional traffic state change map. The calculation results of the difference coefficients need to be filtered by thresholds. For example, when the speed difference coefficient exceeds 3.0, a secondary alarm is triggered, and when it exceeds 5.0, an emergency response protocol is activated.

[0059] Step S240: Construct a multi-dimensional feature vector of the dynamic offset according to the speed difference coefficient and the waiting time difference coefficient, and perform scale unification processing on the multi-dimensional feature vector through the normalization layer of the edge computing node to obtain a normalized multi-dimensional feature vector, where the multi-dimensional feature vector is used to represent the dynamic offset.

[0060] A multi-dimensional feature vector is a data structure that maps difference coefficients with different dimensions to a unified mathematical space. For example, the speed difference coefficient 5.11, the waiting duration difference coefficient 166.7%, and the change in congestion index 8.2 are combined into a three-dimensional vector [5.11, 1.667, 8.2]. The normalization layer uses the min-max scaling algorithm to linearly transform the data of each dimension into the interval [0, 1]. For example, if the maximum threshold of the speed difference coefficient is set to 10, then 5.11 is converted to 0.511; the waiting duration difference coefficient is capped at 200%, and 1.667 is converted to 0.833; the congestion index is directly obtained by dividing the original value by 10, resulting in 0.82. The finally generated normalized multi-dimensional feature vector is [0.511, 0.833, 0.82], and its geometric modulus reflects the degree of deviation of the overall traffic state, and the vector direction identifies the main influencing dimension. Exemplarily, the edge computing node can achieve pattern matching through the calculation of feature vector similarity. For example, when the cosine similarity between the current feature vector and the feature vector in the historical accident database exceeds 0.95, it is determined as the same type of event and the preset disposal plan is loaded. The normalization process eliminates the interference of dimension differences on the model input, ensuring that the subsequent traffic signal adaptive model can accurately analyze the weight relationship of each dimension. For example, in the vector [0.511, 0.833, 0.82], the waiting duration difference coefficient has the largest proportion, and the model will give priority to adjusting the pedestrian phase control parameters.

[0061] Step S300: Input the dynamic offset into the pre-trained traffic signal adaptive model to generate a set of traffic signal control parameters corresponding to the emergency response area, where the set of traffic signal control parameters includes a phase switching frequency parameter, a green light duration parameter, and a direction passing priority parameter.

[0062] Exemplarily, the pre-trained traffic signal adaptive model is a multi-objective optimization system constructed based on a deep reinforcement learning framework, which can be obtained through a large number of simulations of intersection states in a simulation environment. The dynamic offset data received by the model is converted into a 128-dimensional tensor through a feature encoding layer and input into a neural network containing long short-term memory units for spatio-temporal feature extraction. The phase switching frequency parameter is calculated based on the dynamic distribution of traffic flow pressure. For example, the straight phase in the east-west direction is adjusted from a fixed 120-second cycle to a 90-second rapid rotation, while the phase time of the damaged lane in the north-south direction is compressed to 30 seconds. The green light duration parameter is jointly optimized through a congestion propagation model and queuing theory. For example, to divert the vehicles detouring eastward, the left-turn green light is extended from 40 seconds to 65 seconds, and the minimum safe passing time is set to 15 seconds. The direction passing priority parameter is dynamically adjusted according to the emergency rescue requirements. When the ambulance RFID signal is detected, the passing weight coefficient in the relevant direction is forced to be increased to the highest level, the current phase is interrupted, and a dedicated green channel is activated. The output layer of the model uses the Softmax function to generate a multi-dimensional parameter combination, and a constraint satisfaction module is used to ensure that the parameter set complies with the road traffic safety specifications. For example, the yellow light time between adjacent phases is always not less than 3 seconds, and the minimum guarantee period for pedestrians to cross the street does not exceed two signal rotations.

[0063] As an implementation manner, in step S300, inputting the dynamic offset into the pre-trained traffic signal adaptive model to generate a set of traffic signal control parameters corresponding to the emergency response area may specifically include:

[0064] Step S310: Invoke the feature fusion layer of the traffic signal adaptive model to perform spatial association encoding on the dynamic offset and the location identifier of the sudden traffic event to generate an event perception feature map.

[0065] Among them, the feature fusion layer is a network component in the traffic signal adaptive model responsible for the spatial integration of multi-source data. Exemplarily, it realizes the deep binding of traffic state features and geographical locations through geometric topology mapping and attention mechanism. The location identifier of the sudden traffic event is input in the form of longitude and latitude coordinates. For example, the accident point is located in the target intersection coordinate system (120.35°E, 31.78°N). This coordinate system establishes a plane rectangular coordinate system with the intersection center as the origin, the east-west direction as the X-axis, and the north-south direction as the Y-axis. The normalized multi-dimensional feature vector carried by the dynamic offset and the location identifier are jointly input into the spatial association coding module. Among them, each dimensional feature of the dynamic offset is mapped to the corresponding geographical space unit. For example, when the emergency response area is divided into grids of 10 meters × 10 meters, the grid cell G23 (coordinate range X: 50 - 60m, Y: 30 - 40m) corresponding to the speed difference coefficient of 0.511 in the dynamic offset is marked as the core area affected by the event. Exemplarily, the event perception feature map stores the event influence intensity of each spatial unit through a three-dimensional tensor data structure. For example, the feature value generated at the grid cell G23 is 0.87, while the feature value of the adjacent grid cell G24 decays to 0.52, reflecting the propagation law that the accident influence decreases with distance. This feature map also encodes direction dependence. For example, the propagation coefficient of the northbound lane of the accident point is 30% higher than that of the southbound lane, reflecting the asymmetry of the traffic flow blockage direction.

[0066] As an implementation manner, in step S310, performing spatial association coding on the dynamic offset and the location identifier of the sudden traffic event to generate an event perception feature map may specifically include:

[0067] Step S311: Converting the location identifier of the sudden traffic event into a spatial coordinate grid in the target intersection coordinate system, and the division granularity of the spatial coordinate grid is positively correlated with the lane distribution density of the emergency response area.

[0068] Exemplarily, the construction of the spatial coordinate grid is based on the lane line detection result and the digital elevation model. For example, in the core area with a lane distribution density of 3 lanes / 10 meters, a high-precision grid of 1 meter × 1 meter is used, while in the auxiliary road area with a density of 1 lane / 10 meters, a coarse-grained grid of 5 meters × 5 meters is used. The location identifier of the sudden traffic event is mapped to this grid system through a coordinate transformation matrix. For example, the GPS coordinates (116.403°E, 39.914°N) of the accident point are located in grid G1015 (X: 150 - 151m, Y: 15 - 16m) after projection transformation. The grid division algorithm is dynamically adjusted according to the lane topology structure. For example, in a roundabout, a polar coordinate system grid is used, and each sector area corresponds to an angular resolution of 10 degrees and a radial resolution of 2 meters. This process retains the road geometric attributes. For example, the grid cells of the bus-only lane are marked as a special type, and the weight distribution for prohibiting ordinary vehicle passage is assigned.

[0069] Step S312: According to the coverage range of the historical traffic flow characteristics corresponding to each grid cell in the spatial coordinate grid and the dynamic offset, spatially weight the dynamic offset by grid cell to generate a dynamic offset distribution map.

[0070] Exemplarily, the spatial weight assignment of the dynamic offset can adopt the inverse distance weighted interpolation algorithm. For example, the initial weight of the grid G1015 where the accident point is located is 1.0, and the weights of adjacent grids decay inversely with the square of the distance. The weight of grid G1016 (distance 1 meter) is 0.8, and the weight of G1115 (distance 1.41 meters) is 0.5. The coverage range of the historical traffic flow characteristics is determined by kernel density estimation. For example, the influence radius of the historical traffic flow in the northbound through lane is calculated as 50 meters, then the dynamic offset of all grid cells within this range needs to be superimposed with the lane characteristic correction coefficient of 0.6. The dynamic offset distribution map stores the composite offset values of each grid cell in matrix form. For example, the normalized multi-dimensional feature vector [0.511, 0.833, 0.82] of grid G1015 is converted to [0.511×1.0, 0.833×0.9, 0.82×0.8] after spatial weight assignment, reflecting the attenuation effect of the geographical location on each offset. This distribution map also records the time decay factor. For example, the offset weight at the 5th minute after the accident automatically drops to 70% of the initial value.

[0071] Step S313: Perform multi-scale spatial convolution processing on the offset data in each grid cell of the dynamic offset distribution map, extract the radiation influence characteristics of the sudden traffic event on adjacent lanes, and superimpose the geometric constraint conditions of the spatial coordinate grid.

[0072] Exemplarily, the multi-scale spatial convolution uses three convolution kernels of 3×3, 5×5, and 7×7 in parallel. For example, use the 3×3 convolution kernel to extract the influence characteristics between local lanes and capture the detail that the northbound vehicle speed drops by 23% within 10 meters around the accident point; the 5×5 convolution kernel identifies the regional-level propagation mode and discovers that the radius of the increased pressure in the eastward detour lane reaches 50 meters; the 7×7 convolution kernel reveals the global road network-level chain reaction and detects an abnormal fluctuation of 10% in the traffic flow at an intersection 1 kilometer away. The geometric constraint conditions are implemented through the road topology mask. For example, the upper limit of the radiation influence coefficient is forcibly set to 0.3 in the sidewalk grid cell to prevent the wrong propagation of vehicle congestion characteristics to the pedestrian control area. The feature map after convolution processing retains the road connectivity information. For example, only the influence is allowed to propagate along the lane direction, and the invalid diffusion across the central isolation belt is prohibited.

[0073] Step S314: Based on the superposition result of the radiation influence feature and the geometric constraint condition, generate an initial spatial association feature map, and input the initial spatial association feature map into a bidirectional attention mechanism layer for cross-region dependency modeling.

[0074] Exemplarily, the initial spatial association feature map fuses radiation intensity and road structure information. For example, in the grid cell of the northbound straight lane, the convolution output value of 0.75 is multiplied by the lane connectivity coefficient of 0.9 to obtain the final feature value of 0.675. The bidirectional attention mechanism layer includes a direction perception module and a region association module: The direction perception module calculates the attention weights of each grid cell in the east-west and north-south axes. For example, the attention score of the grid on the east side of the accident point in the east-west direction is 0.92, which is significantly higher than 0.35 on the west side; The region association module establishes an influence transmission path across lanes. For example, the correlation coefficient for identifying that the traffic pressure of the eastward detour traffic flow is conducted to the southward branch road through the feeder road is 0.68. The attention weight matrix is normalized by the softmax function to ensure the additivity of the influence coefficients in each direction.

[0075] Step S315: In the bidirectional attention mechanism layer, calculate the direction-sensitive association degree between each grid cell in the initial spatial association feature map and the location identifier of the sudden traffic event, and generate an event-aware feature map containing the event influence propagation path through adaptive weighted fusion.

[0076] Exemplarily, the direction-sensitive association degree is calculated using the direction cosine similarity algorithm. For example, the connection direction between grid cell G2015 and accident point G1015 is 30 degrees east-south, and the cosine value of its angle with the northward reference direction is 0.866. This value is used as a direction attenuation factor to participate in the association degree calculation. Adaptive weighted fusion dynamically adjusts the weights of each propagation path according to the real-time traffic state. For example, during the morning rush hour, the propagation weight in the main commuting direction (north-south) is increased to 1.2, and the non-main direction is reduced to 0.8. The event influence propagation path is visualized through the heatmap gradient. For example, the influence intensity in the north direction from the accident point drops to 0.4 at 50 meters, and the influence value in the east direction remains 0.6 at 100 meters due to the detour demand. The final event-aware feature map stores the comprehensive features of each grid cell in three dimensions: space, direction, and intensity in the form of a three-dimensional tensor.

[0077] Step s316: Mark the grid cells in the event-aware feature map whose association degree exceeds the preset threshold as high-priority regulation regions, and map the boundary coordinates of the high-priority regulation regions to the direction passing priority parameters in the traffic signal control parameter set.

[0078] Exemplarily, the preset threshold is dynamically set according to the road grade. For example, in the urban expressway scenario, the correlation threshold is set to 0.7 and the branch road is set to 0.5. When the correlation of grid unit G1015 reaches 0.87, its lane and the adjacent 50-meter range are designated as a high-priority control area. The boundary coordinates are extracted by the convex hull algorithm. For example, the grid unit coordinates with a correlation of ≥ 0.7 are input into the Graham scanning algorithm to generate a minimum enclosing polygon vertex sequence. The directional traffic priority parameter is set according to the lane direction covered by the area. For example, when the high-priority control area includes an east-south left-turn lane and a north-bound straight lane, its directional traffic priority parameters are set to Level 3 and Level 2 respectively. The parameter mapping process retains geometric topological relationships. For example, the high-priority control area of ​​the roundabout area is partitioned by polar angles, and each sector area corresponds to a specific phase priority number. In the final generated set of traffic light control parameters, the directional traffic priority parameter carries geographic fence information to ensure accurate matching of signal control with the physical road structure.

[0079] Step S320: predicting the vehicle arrival rate change curve and the pedestrian gathering density change curve of the emergency response area within a future time interval based on the event perception feature graph through the time series prediction layer of the traffic signal adaptive model.

[0080] Exemplarily, the time series prediction layer can adopt a gated recurrent unit network architecture, whose input is the time evolution sequence of the event perception feature map, and the output is the traffic state prediction value of the granularity of the next 5 minutes. The vehicle arrival rate change curve is generated by analyzing the vehicle accumulation rate of each grid unit in the event perception feature map. For example, it is predicted that the vehicle arrival rate within 200 meters north of the accident point will increase linearly from the current 12 vehicles per minute to 28 vehicles in the next 10 minutes, while the arrival rate of the eastbound detour lane will show a nonlinear trend of first increasing to a peak of 35 vehicles and then falling back to 22 vehicles. The pedestrian gathering density change curve is calculated based on the heat map diffusion model. For example, the current gathering density of the sidewalk on the west side of the accident point is 0.8 people / square meter. It is predicted that it will increase to 1.2 people / square meter in the next 5 minutes due to pedestrian detours, and the evacuation warning will be triggered after exceeding the safety threshold. Multimodal data fusion technology is used in the prediction process. For example, the pedestrian growth rate during the evening peak period in the historical data of the same period is used as a benchmark, and the current event impact coefficient is superimposed with a weight of 1.3 times, and the slope parameter of the prediction curve is corrected. The output results of the time series prediction layer are annotated with reliability through confidence intervals. For example, the 95% confidence interval of the vehicle arrival rate prediction value is ±3 vehicles / minute. When the interval width exceeds the preset threshold, the manual review mechanism is activated.

[0081] As an implementation manner, the step S320, predicting the vehicle arrival rate change curve and the pedestrian gathering density change curve of the emergency response area in the future time interval based on the event perception feature map, may specifically include:

[0082] Step S321: Divide the event perception feature map into a continuous feature map time series in the time dimension, and extract the spatial influence intensity distribution corresponding to each time slice in the feature map time series.

[0083] The feature map time series refers to a four-dimensional data structure formed by slicing a three-dimensional event perception feature map along the time axis, and its time resolution matches the decision-making cycle of the traffic signal control system. For example, for the prediction requirement of the next 15 minutes, the event perception feature map is divided into 30 time slices at intervals of 30 seconds, and each time slice stores the three-dimensional feature values of each grid cell at that moment. The spatial influence intensity distribution is generated by aggregating the feature values of each grid cell in the feature map time series. For example, at time slice T5 (150 seconds after the accident), the radiation influence feature value of grid cell G1015 is 0.87, and the geometric constraint coefficient is 0.9. After superposition calculation, the spatial influence intensity value of this cell is 0.783. This distribution is visually presented through a two-dimensional matrix. For example, in the area with coordinates X: 100 - 150m, Y: 0 - 50m, a red high-influence area (intensity value ≥ 0.7) and a yellow medium-influence area (0.4 ≤ intensity value < 0.7) are marked, clearly showing that the influence range of the accident on the eastward detour lane extends to 120 meters.

[0084] Step S322: Based on the spatial influence intensity distribution, identify the spatial overlapping area between the main path of the vehicle passing direction and the pedestrian activity hot spot in the emergency response area, and generate a path-hot spot coupling feature vector.

[0085] Exemplarily, the main path can be determined through vehicle trajectory clustering analysis. For example, the east-west straight lane forms a dense traffic belt with a width of 3 lanes due to detour requirements, and its path centerline is composed of continuous grid cells G2015 - G2018 - G2021. The pedestrian activity hot spot is identified using the kernel density estimation algorithm. For example, a northwest crosswalk forms an elliptical hot spot with a diameter of 15 meters during the evening rush hour, and the density peak reaches 1.5 people per square meter. The spatial overlapping area detection is achieved through the intersection operation of the geofence. For example, the intersection area of the main path G2015 - G2021 and the pedestrian hot spot HZ03 is the grid cells G2017 - G2019, and this area is marked as a high-risk conflict area. The path-hot spot coupling feature vector consists of three dimensions: the path traffic pressure index, the hot spot density index, and the overlapping area ratio. For example, when the main path pressure index is 0.82, the hot spot density index is 1.3, and the overlapping ratio is 35%, a feature vector [0.82, 1.3, 0.35] is generated to quantify the potential intensity of the vehicle-pedestrian conflict.

[0086] Step S323: Retrieve historical event data similar to the path-hotspot coupling feature vector from the historical database of the edge computing node, and extract the association rules between the vehicle arrival rate fluctuation pattern and the pedestrian density growth pattern in the historical events.

[0087] Exemplarily, the historical event similarity matching can adopt the cosine similarity algorithm. For example, the similarity between the current feature vector [0.82, 1.3, 0.35] and the feature vector [0.79, 1.28, 0.33] of historical record E029 is 0.98, and it is determined as the same type of event. The association rule mining is realized by the Apriori algorithm. For example, it is found that when the third dimension (overlap ratio) of the path-hotspot coupling feature vector exceeds 30%, the probability of the vehicle arrival rate showing a first increasing and then decreasing fluctuation within the subsequent 10 minutes reaches 87%, and at the same time, the probability of the pedestrian density showing a stepwise growth is 73%. The historical event data carries the traffic parameter change sequence marked with time stamps. For example, in the similar event E029, the vehicle arrival rate reaches the peak of 48 vehicles per minute at the 5th minute after the accident, and the pedestrian density breaks through the safety threshold of 1.2 people per square meter at the 8th minute.

[0088] Step S324: Align the association rules with the spatial influence intensity distribution of the current event perception feature map in space-time to construct a space-time fusion feature including vehicle motion inertia and pedestrian group behavior tendency.

[0089] Exemplarily, the space-time alignment process can adopt the dynamic time warping algorithm. For example, the time axis of the vehicle arrival rate curve of historical event E029 is compressed by 15% to adapt to the earlier peak period of the current event. The vehicle motion inertia feature is calculated by the standard deviation of acceleration in historical data. For example, the average deceleration amplitude of the eastward detouring vehicles when avoiding the accident point reaches 40%, and this feature is encoded as an inertia coefficient of 0.6. The pedestrian group behavior tendency is obtained by mining smartphone signaling data. For example, it is detected that the pedestrian flow at the subway station exit 500 meters west of the accident point increases by 20%, and it is deduced that the probability of pedestrians choosing the detour path increases to 65%. The space-time fusion feature finally forms a multi-dimensional tensor structure. For example, at time slice T10, the eastward lane inertia coefficient of 0.6, the pedestrian detour probability of 0.65, and the real-time spatial influence intensity of 0.78 are fused to form a feature unit [0.6, 0.65, 0.78].

[0090] Step S325: Invoke the pre-trained time series prediction model, use the space-time fusion feature as the input, and predict the vehicle arrival rate increment ratio and the gradient change direction of the pedestrian aggregation density in each lane within the future time interval step by step.

[0091] Exemplarily, the time series prediction model can adopt a long short-term memory network architecture with an attention mechanism, and its input layer unfolds the spatio-temporal fusion feature tensor into a time step sequence. The vehicle arrival rate increment ratio is calculated as the relative change rate of the number of vehicles in adjacent time slices. For example, it is predicted that the arrival rate on the eastbound straight lane increases from 28 vehicles per minute to 32 vehicles per minute in the time slice from T5 to T6, and the increment ratio is 14.3%. The change direction of the pedestrian aggregation density gradient is derived through a thermodynamic diffusion model. For example, after the density in the hot area on the northwest side reaches 1.2 people per square meter at time slice T8, the gradient direction vector points to the southeast, indicating that pedestrians start to disperse to the side road. The prediction process performs multiple rounds of Monte Carlo simulations. For example, 100 sampling predictions are made for the arrival rate on the eastbound lane, and the 90% confidence interval is taken as 29 - 35 vehicles per minute to ensure the robustness of the results.

[0092] Step S326: Generate a vehicle arrival rate change curve at consecutive time points according to the vehicle arrival rate increment ratio, and at the same time fit a pedestrian aggregation density change curve based on the gradient change direction, and mark the intersection time point of the two curves as the phase switching trigger signal.

[0093] Exemplarily, the vehicle arrival rate change curve can be smoothed by using the cubic spline interpolation method. For example, the discrete prediction values T5: 28 vehicles, T6: 32 vehicles, T7: 37 vehicles are connected into a continuous curve, and 34 vehicles per minute are interpolated at the time point T6.5. The pedestrian aggregation density change curve is generated by integrating the gradient direction field. For example, the southeast gradient at time slice T8 causes the density to decrease by 0.1 person per square meter every 30 seconds, and an exponential decay curve is fitted. The intersection time point detection adopts a numerical approximation algorithm. For example, when the vehicle arrival rate curve reaches 35 vehicles per minute at the time point T7.2 and the pedestrian density curve synchronously drops to 1.0 person per square meter, it is determined as the demand balance point, and the phase switching trigger signal SIG 07 is generated. This signal carries a timestamp and a position encoding. For example, it is marked as "2023-09-15T17:23:45@G2017" to ensure the spatio-temporal accuracy of the control instruction.

[0094] Step S330: Determine the initial value of the phase switching frequency parameter according to the intersection point of the vehicle arrival rate change curve and the pedestrian aggregation density change curve.

[0095] Exemplarily, the calculation of the initial value of the phase switching frequency parameter is based on the principle of traffic pressure balance. When the vehicle arrival rate curve and the pedestrian aggregation density curve form an intersection on the time axis, it indicates that the vehicle passing demand and the pedestrian crossing demand reach a critical balance state at that moment. For example, the prediction shows that at the 8th minute after the accident, the vehicle arrival rate of the eastbound left-turn lane will rise to 32 vehicles per minute, and at the same time, the aggregation density of the northwest sidewalk reaches 1.1 people per square meter. At this time, the control device 120 determines that the east-west straight-through phase switching frequency needs to be adjusted from a fixed cycle of 90 seconds to a dynamic range of 60 - 120 seconds. The initial value generation algorithm uses differential equations to solve for the optimal switching point. For example, an optimization model with the minimum delay time as the objective function is established, and the initial value of the phase switching frequency is calculated to be 78 seconds through the Lagrange multiplier method. This process considers the multi-directional competition relationship. For example, when the demand for the north-south ambulance lane surges, a dedicated phase is forcibly inserted, resulting in the baseline value of the east-west phase switching frequency being reduced to 65 seconds. After the initial value is set, conflict detection needs to be performed. For example, verify whether the phase switching causes the green light time of adjacent intersections to overlap by more than the safety threshold. If a conflict is detected, recalculate until the constraint conditions are met.

[0096] As an implementation manner, in step S330, according to the intersection point of the vehicle arrival rate change curve and the pedestrian aggregation density change curve, to determine the initial value of the phase switching frequency parameter, it can be specifically implemented as the following steps:

[0097] Step S331: Detect the intersection point of the vehicle arrival rate change curve and the pedestrian aggregation density change curve on the time axis, and extract the vehicle arrival rate rising slope and the pedestrian density growth rate within a preset time window before and after the intersection point.

[0098] Exemplarily, the preset time window is set according to the traffic flow response delay. For example, an asymmetric observation interval composed of 3 minutes before and 2 minutes after the intersection point is selected. The vehicle arrival rate rising slope is calculated by linear regression. For example, within the time window of T4 - T7, the vehicle arrival rate of the eastbound lane increases from 22 vehicles per minute to 38 vehicles per minute, and the slope is calculated as (38 - 22) / (3×60) = 0.089 vehicles per second. The pedestrian density growth rate is calculated using the instantaneous derivative. For example, at the intersection point T7.2, the derivative of the density curve is -0.015 people per square meter per second, indicating that the density begins to decrease. This step synchronously calculates the direction-specific parameters. For example, the arrival rate slope of the north-south bus dedicated lane is separately calculated as 0.021 vehicles per second and is processed separately from the ordinary lane.

[0099] Step S332: Based on the ratio of the vehicle arrival rate rising slope to the pedestrian density growth rate, calculate the conflict intensity coefficient of the vehicle and pedestrian passing demands at the intersection point, and map the conflict intensity coefficient to a preset phase switching frequency level interval.

[0100] Exemplarily, the conflict intensity coefficient is calculated as |vehicle arrival rate slope / pedestrian density slope|. For example, when the vehicle slope is 0.089 vehicles per second and the pedestrian slope is -0.015 persons per square meter per second, the coefficient is 5.93. The preset grade intervals are dynamically divided according to the road type. For example, for the urban arterial road, four-level intervals are set: coefficient < 3 corresponds to grade 1 (frequency 90 - 120 seconds), 3 ≤ coefficient < 6 corresponds to grade 2 (60 - 90 seconds), 6 ≤ coefficient < 9 corresponds to grade 3 (40 - 60 seconds), and coefficient ≥ 9 corresponds to grade 4 (emergency mode). The current coefficient 5.93 is mapped to grade 2, generating the initial frequency range of 60 - 90 seconds. The mapping process takes into account the direction weight. For example, the weight coefficient of the eastward arterial road is 1.2, and the actual grade is adjusted to 2.4, corresponding to the frequency range being corrected to 50 - 75 seconds.

[0101] Step S333: Determine the dynamic adjustment step size of the phase switching frequency parameter in the initial stage according to the peak holding duration of the vehicle arrival rate change curve and the position of the inflection point of the decline of the pedestrian aggregation density change curve after the intersection.

[0102] Exemplarily, the peak holding duration is detected by the zero point of the second derivative of the curve. For example, the vehicle arrival rate remains above 38 vehicles per minute for 110 seconds after T7.2, and this duration is used as the frequency adjustment benchmark. The inflection point of the pedestrian density decline is detected by the extreme value of the curvature. For example, the curvature reaches the maximum value at the time point T7.8, indicating the highest evacuation efficiency point. The dynamic adjustment step size is calculated as the ratio of the peak duration to the interval of the inflection point. For example, the peak duration of 110 seconds and the interval of 60 seconds from T7.2 to T7.8 result in a step size coefficient of 1.83. This coefficient is converted into the phase switching step size rule. For example, the basic adjustment step size is set to 10 seconds, and the actual step size is 10 × 1.83 ≈ 18 seconds, allowing the frequency to be evaluated and adjusted every 18 seconds based on the initial value of 78 seconds.

[0103] Step S334: Extract the association rule between the phase switching frequency and the conflict intensity coefficient within the historical time window before the intersection, and generate the initial candidate value set of the phase switching frequency parameter in combination with the dynamic adjustment step size.

[0104] Exemplarily, the association rule can be mined and generated by a decision tree. For example, in the historical data, the frequencies of successful regulation cases corresponding to the conflict intensity coefficient in the range of 5.0 - 6.0 are mostly concentrated in the range of 70 - 85 seconds. The candidate value set is generated by an arithmetic sequence. For example, with the reference value of 78 seconds as the center, a candidate set {73, 78, 83} is generated at intervals of ±5 seconds, and then combined with the periodic evaluation requirement of the dynamic step size of 18 seconds, it is expanded to a candidate value set of {73, 78, 83, 91}. The candidate values need to be filtered by physical constraints. For example, values lower than the minimum safety period of 45 seconds are excluded, and the final valid set is {73, 78, 83}.

[0105] Step S335: Input the initial candidate value set into a pre-trained phase decision model, evaluate the comprehensive influence weights of each candidate value on the vehicle passing delay rate and pedestrian waiting time shortening rate in the emergency response area, and select the candidate value with the highest comprehensive influence weight as the initial value of the phase switching frequency parameter.

[0106] Exemplarily, the phase decision model can adopt a multi-objective optimization architecture. The vehicle delay rate is calculated as ∑(actual passing time - free flow time), and the pedestrian waiting time shortening rate is calculated as (historical waiting time - predicted waiting time) / historical waiting time. For example, the vehicle delay rate corresponding to the candidate value of 78 seconds is 35 vehicles per hour, and the pedestrian waiting time shortening rate is 18%; the delay rate corresponding to the candidate value of 83 seconds is 42 vehicles per hour, and the shortening rate is 23%. The comprehensive influence weight is calculated by the entropy weight method. Assuming the delay rate weight is 0.6 and the shortening rate weight is 0.4, the score for 78 seconds is 35×0.6 + 18×0.4 = 29.4, and the score for 83 seconds is 42×0.6 + 23×0.4 = 35.6. Finally, 83 seconds is selected as the initial value. A real-time correction factor is introduced in the evaluation process. For example, when an emergency vehicle is detected approaching, the vehicle delay weight is temporarily increased to 0.8.

[0107] Step S336: Generate a phase switching instruction test sequence based on the initial value, and verify the mitigation effect of the test sequence on vehicle-pedestrian conflicts after the intersection in the simulation environment of the edge computing node. If the verification passes, lock the initial value; otherwise, re-trigger the associated rule extraction and candidate value generation steps.

[0108] Exemplarily, the phase switching instruction test sequence includes phase order, duration, and transition rules. For example, "Phase B lasts for 83 seconds → yellow light for 3 seconds → Phase C lasts for 45 seconds". The simulation environment loads a real-time road network state snapshot, including vehicle positions, speeds, and pedestrian distribution data. The verification metrics include the conflict point throughput (for example, the number of conflicts in the eastbound lane during the predicted T7.2 - T8.0 period drops from 28 to 9), the average delay change (the vehicle delay is reduced by 12%), etc. When the simulation results show that the pedestrian density drops to 0.9 people per square meter at the T8.5 time point and there are no new conflicts, it is determined that the verification passes, and 83 seconds is written into the control parameter set. If secondary congestion (queue length exceeds 150 meters) is detected in the westbound lane, then roll back to the candidate value of 78 seconds for re-verification until a feasible solution is found or an artificial intervention protocol is triggered.

[0109] Step S340: Based on the initial value, iteratively adjust the direction weight of the green light duration parameter through the parameter optimization layer of the traffic signal adaptive model until the direction passing priority parameter meets the preset conflict avoidance constraint conditions.

[0110] Exemplarily, the parameter optimization layer may adopt a hybrid optimization strategy combining genetic algorithms and gradient descent, which models the green light duration parameter as a multi-dimensional decision variable, and each dimension corresponds to the weight of the passing time allocation in a specific direction. For example, the initial value of the straight green light duration in the east-west direction is 55 seconds, and the weight coefficient is 0.7; the initial value of the damaged lane in the north-south direction is 30 seconds, and the weight is 0.3. During the iteration process, the direction weights are dynamically adjusted according to the real-time traffic pressure: when it is detected that the number of detoured vehicles in the east direction increases by 150 per hour, the weight coefficient increases to 0.8 according to the gradient ascent rule, and the corresponding green light time is extended to 62 seconds; at the same time, to ensure the safe passage of pedestrians, the weight coefficient of the north-side crosswalk is increased to 0.4, and a 15-second dedicated phase is forced to be inserted. The conflict avoidance constraint conditions are implemented through a system of linear inequalities. For example, it is ensured that the yellow light time ≥ 3 seconds when the phase sequence switches, and the green light interval between motor vehicles and pedestrians in the same direction ≥ 5 seconds. After each round of iteration, the satisfaction degree of the constraints is verified. When it is detected that there is a 2-second overlap between the west-turning right and the pedestrian phase, the weight rollback mechanism is triggered and the direction passing priority parameters are recalculated. The optimization termination condition is that the change amount of the direction weights is less than 0.01 in three consecutive iterations. At this time, the final set of traffic signal control parameters is generated. For example, the phase switching frequency is 78 seconds, the east-west straight green light is 62 seconds, and the priority level of the north-side pedestrian phase is increased to Level 2.

[0111] As an implementation manner, the traffic signal adaptive model can be trained by the following steps:

[0112] Step s301: Collect training data corresponding to multiple historical sudden traffic events, where the training data includes real-time traffic flow data at the time of the event, traffic signal adjustment records after the event is processed, and traffic recovery efficiency indicators after the event ends.

[0113] Historical sudden traffic events refer to event cases with complete disposal records screened through the accident database of the traffic management department. For example, a four-vehicle rear-end collision accident that occurred at the intersection of the urban arterial road on May 15, 2023. The data collection scope covers the real-time traffic flow data during the accident duration (17:23 - 18:45), specifically including the vehicle speed matrix collected by geomagnetic sensors (updated per second), the crosswalk frequency captured by pedestrian detection cameras (counted per minute), and the signal light status change log recorded by roadside units. The traffic signal adjustment records after event processing are stored in the form of a control instruction sequence marked with timestamps. For example, the instruction of "extending the green light for the east-west straight lane to 65 seconds" issued at the 8th minute after the accident and the configuration parameter of "upgrading the priority of the north turn phase to Level 2" triggered at the 12th minute. The traffic recovery efficiency index is quantitatively calculated by comparing the traffic states before and after event disposal. For example, the average vehicle delay time is reduced from 82 seconds at the accident peak to 37 seconds after disposal, the standard deviation of pedestrian crosswalk waiting time is optimized from 45 seconds to 18 seconds, and the road capacity recovery rate is increased to 92%. The data collection process follows the integrity verification rules. For example, only event cases with a disposal cycle of at least 30 minutes and a sensor coverage rate exceeding 95% are selected and included in the training set.

[0114] Step S302: Perform spatio-temporal slicing on the real-time traffic flow data to generate training samples that match the dimensions of the emergency response area, and convert the traffic signal adjustment records into a set of label parameters for the training samples.

[0115] Exemplarily, the spatio-temporal slicing process can adopt a sliding window mechanism to divide the continuous time series data into spatio-temporal units of a fixed length. For example, with a 5-minute time slice length and a 50m×50m spatial grid unit, resample the vehicle speed matrix generated during the accident. Each training sample contains a multi-modal feature tensor within the emergency response area. For example, in the 17:25 - 17:30 time slice, the average vehicle speed in grid G1015 is 18 km / h, the pedestrian density is 0.7 people per square meter, and the signal light status is phase C activated. The set of label parameters is generated by parsing the control instructions in the traffic signal adjustment records. For example, map the instruction of "extending the green light to 65 seconds" to the green light duration parameter 65, and encode the phase priority Level 2 as the direction passing priority parameter 0.8. During the data dimension alignment process, the bilinear interpolation algorithm is used to match the original sensor data to the standard grid coordinate system of the emergency response area. For example, aggregate the lane-level vehicle speed data to grid units with a 5m accuracy to ensure the consistency of the training sample spatial resolution.

[0116] Step S303: Construct the neural network structure of the initial traffic signal adaptive model. The neural network structure includes a graph convolutional branch for extracting spatial dependencies and a recurrent neural network branch for capturing temporal dependencies.

[0117] Exemplarily, the graph convolutional branch constructs a graph attention network based on the road network topology. The nodes represent the real-time traffic state features of each lane, and the edge weights are jointly determined by the lane connection relationship and the historical traffic flow correlation. For example, between the connection nodes of the north-south straight lane and the east-west left-turn lane, an initial edge weight of 0.75 is set according to historical data statistics, reflecting the traffic flow competition intensity between the two during the peak period. The recurrent neural network branch adopts a gated recurrent unit architecture, and its time step is synchronized with the spatio-temporal slice interval. For example, each 5-minute time slice corresponds to an input of one time step, and the memory unit retains the traffic state evolution features of the previous three time steps. The outputs of the spatial and temporal branches are fused through a feature concatenation layer. For example, a 32-dimensional spatial feature vector extracted by graph convolution and a 64-dimensional temporal feature vector output by the recurrent neural network are concatenated into a 96-dimensional joint feature and input into a fully connected layer for dimensionality reduction processing. The residual connection module embedded in the network structure ensures the effective transmission of deep features. For example, a cross-layer connection is introduced after the fifth layer of convolution, and the original input feature and the convolution output are superimposed in a ratio of 0.3:0.7.

[0118] Step S304: Input the training samples into the initial traffic signal adaptive model, calculate the mean square error loss between the label parameter set and the model output parameters, and update the connection weights of the neural network structure through the gradient descent algorithm.

[0119] Exemplarily, the mean square error loss function simultaneously considers the collaborative optimization of multi-objective parameters. For example, the error values are calculated separately for the phase switching frequency parameter, the green light duration parameter, and the direction passing priority parameter, and weighted and summed according to the weight ratio of 0.4:0.4:0.2. The adaptive moment estimation optimization algorithm is adopted in the gradient descent process, and the initial learning rate is set to 0.001. When the validation set loss does not decrease for 5 consecutive training cycles, the learning rate decay mechanism is triggered. For example, the learning rate is reduced to one-tenth of the original value. A gradient clipping strategy is implemented during the weight update process to limit the parameter adjustment amplitude not to exceed 0.1 to prevent model oscillation. Each training batch contains 32 spatio-temporal slice samples, and the forward propagation and backward propagation calculations are completed in a GPU-accelerated environment. For example, processing a complete training set containing 2000 slices requires 63 iterations of batches, and the processing time for a single batch is 850 milliseconds. The model takes a snapshot every 10 training cycles and records the current best weight combination for subsequent recovery.

[0120] Step S305: When the improvement rate of the traffic recovery efficiency index on the validation dataset reaches the preset threshold, stop the training and deploy the model parameters to the edge computing node.

[0121] The validation dataset contains historical event cases that did not participate in the training. For example, the bridge closure event that occurred on June 10, 2023, and its traffic recovery efficiency index is calculated by comparing the model prediction results with the actual disposal effects. The preset threshold is set to a comprehensive index improvement rate of not less than 15%, specifically including the geometric mean of the vehicle delay reduction rate, the pedestrian waiting time optimization rate, and the road capacity recovery rate. When the model reaches an improvement rate of 18.7% in three consecutive validation cycles, the early stopping mechanism is triggered to terminate the training process. Before deploying the model parameters, quantization compression processing is required. For example, convert 32-bit floating-point weights to 8-bit integer representation, and reduce the model size to one-fourth of the original size while maintaining 98.3% prediction accuracy. During the deployment process, A / B testing verification is implemented. For example, run the new and old versions of the model simultaneously on the edge computing node, and compare the conflict resolution efficiency of the disposal plan in the simulation environment. When the new model achieves a winning rate of 83% in 1000 tests, the final replacement is completed. The model update mechanism sets a rollback protection policy. If the prediction error exceeds the safety threshold for 5 consecutive times after deployment, it will automatically revert to the previous stable version and trigger an alarm notification.

[0122] Step S400: According to the traffic signal control parameter set, send a control instruction sequence to the traffic signals at the target intersection, and dynamically adjust the execution priority of the control instruction sequence based on the real-time change trends of the vehicle density distribution data and the pedestrian activity trajectory data.

[0123] The control instruction sequence is a standard control instruction set that conforms to the NTCIP protocol, including phase activation commands, timing parameter update instructions, and priority override instructions. For example, first send a forced switching instruction of "Phase B terminate immediately", then issue a parameter modification instruction of "Extend the green light of Phase C to 55 seconds", and finally write a configuration update instruction of "Direction priority table version 3.2". The dynamic adjustment mechanism is implemented through an online learning algorithm, and the sensor data is incrementally analyzed every 500 milliseconds: when the vehicle density distribution data shows that the eastward detour traffic flow suddenly decreases by 20%, automatically reduce the execution priority of the green light delay in that direction; when the general pedestrian detection system detects that the number of people staying in the northwest exceeds the safety threshold, immediately insert an emergency pedestrian phase instruction to the head of the queue. The execution engine adopts a double-buffer mechanism to ensure the continuity of instruction switching, and at the same time ensures the reliable transmission of control instructions through the CRC check and retransmission protocol. The priority adjustment strategy comprehensively considers multi-modal conflicts. For example, when the priority request of a fire truck to pass first overlaps with the pedestrian crossing demand, an arbitration decision is made based on the preset emergency response hierarchy system to ensure the absolute priority of the key rescue channel.

[0124] As an implementation manner, in step S400, based on the real-time change trends of the vehicle density distribution data and the pedestrian activity trajectory data, the execution priorities of the control instruction sequences are dynamically adjusted, which may specifically include the following steps:

[0125] Step S410: Monitor the position coordinates and movement directions of the newly added vehicles within the emergency response area, and calculate the shortest path distance between the newly added vehicles and the position identifier of the sudden traffic event.

[0126] Exemplarily, the position coordinates of the newly added vehicles are obtained through the collaborative positioning of the in-vehicle global positioning system unit and the roadside multi-target tracking radar. For example, a white van is detected at 31.23 degrees north latitude and 121.47 degrees east longitude, traveling eastward along the Yan'an Elevated Road at a speed of 8.5 meters per second. The calculation of the shortest path distance is based on the real-time road network topology map, and the dynamic Dijkstra algorithm is used to determine the optimal path under the current traffic conditions. For example, this van needs to bypass through the Yan'an East Road Flyover to the accident point, and the total path length is calculated as 1.2 kilometers. After excluding the Tibet Middle Road ramp closed due to road control, the effective travel distance is corrected to 1.5 kilometers. The dynamic update frequency of the path distance is once per second. When it is detected that the emergency channel on Xinxing Road is temporarily opened, the path is recalculated immediately and updated to 1.0 kilometer. The calculation process integrates the real-time traffic flow data. For example, when the queuing length of 80 meters is detected at the Henan Middle Road intersection, the path weight coefficient automatically increases by 0.3 to reflect the actual travel difficulty.

[0127] Step S420: Predict the time interval for the newly added vehicles to reach the sudden traffic event area according to the shortest path distance and the current average vehicle speed.

[0128] Exemplarily, the current average vehicle speed is calculated through the fusion of floating car data and fixed detector data. For example, on the section from East Nanjing Road to The Bund, the average instantaneous speed uploaded by the in-taxi terminal is 18 kilometers per hour, and the sectional speed statistically measured by the geomagnetic sensor is 15 kilometers per hour. After weighted averaging, the current average vehicle speed is 16.2 kilometers per hour. The time interval prediction model uses the sliding window regression algorithm. For example, at a path distance of 1.5 kilometers, the initial predicted arrival time is 5 minutes and 33 seconds. When it is detected that the vehicle speed suddenly drops to 5 kilometers per hour at a distance of 200 meters ahead, the model dynamically corrects the predicted value to 8 minutes and 17 seconds. The prediction result is marked with reliability through the confidence interval. For example, at a confidence level of 95%, the time interval range is from 7 minutes and 45 seconds to 8 minutes and 49 seconds, and the width of this interval is used as the risk assessment basis for the emergency response decision.

[0129] Step S430: If the time interval is less than a preset emergency response time threshold, then increase the green light duration parameter of the lane where the new vehicle is located to the first priority level.

[0130] Exemplarily, the preset emergency response time threshold is dynamically configured according to the severity level of the event. For example, in the second-level emergency response state, the threshold is set to 10 minutes. When the detected estimated arrival time interval of the ambulance is 7 minutes and 22 seconds, the priority adjustment mechanism is triggered. The increase amplitude of the green light duration parameter follows a hierarchical control strategy. The first priority level corresponds to an adjustment upper limit of 150% of the reference value. For example, the original straight-ahead green light duration in the east-west direction of 60 seconds is increased to 90 seconds, and at the same time, the yellow light transition time of the adjacent phase is compressed to 2 seconds. The priority adjustment instruction is sent to the signal controller through a dedicated short-range communication protocol. For example, at the intersection of East Yan'an Road - Middle Shandong Road No. 1, after receiving the priority instruction, the signal controller immediately interrupts the current phase cycle and forcibly inserts an instruction to extend the straight-ahead green light in the east-west direction. During the adjustment process, a conflict detection mechanism is implemented. When it is detected that the remaining time for pedestrians to cross the street is less than 5 seconds, a 3-second all-red phase is automatically inserted to ensure pedestrian safety.

[0131] Step S440: Synchronously detect the abnormal aggregation area in the pedestrian activity trajectory data. If the abnormal aggregation area overlaps with the location identifier of the sudden traffic event, then allocate a dedicated pedestrian passing time window in the direction passing priority parameter.

[0132] Exemplarily, the abnormal aggregation area is identified through the fusion of thermal imaging camera and millimeter-wave radar data. For example, at the crosswalk on the south side of the accident point, a cross-street flow of 45 people per minute is detected, exceeding 120% of the historical peak, and the median crowd retention time reaches 85 seconds. The position overlap determination uses a geographic fence intersection algorithm. When the overlapping area between the geographic coordinate boundary of the pedestrian aggregation area and the 50-meter buffer zone of the accident point exceeds 60%, the pedestrian priority response mechanism is activated. The allocation of the dedicated pedestrian passing time window follows the principle of minimum interference. For example, a 30-second dedicated green light is inserted during the low-peak period of north-south vehicle traffic, and at the same time, the east-west green light time is shortened to 40 seconds. When writing the time window parameter into the direction passing priority parameter table, it is necessary to verify the conflict with the emergency rescue channel. When it is detected that a fire truck needs to cross the dedicated pedestrian phase area, the pedestrian priority is automatically downgraded and a detour guidance message is sent to the variable message board.

[0133] As an implementation manner, in step S440, allocating a dedicated pedestrian passing time window in the direction passing priority parameter includes:

[0134] Step S441: Obtain the real-time moving direction distribution of pedestrians in the abnormal aggregation area, and extract the target moving direction with the highest frequency of occurrence in the real-time moving direction distribution as the pedestrian dominant flow direction.

[0135] Exemplarily, the real-time moving direction distribution is constructed by a binocular stereo vision sensor and a pedestrian re-identification algorithm. For example, at the east entrance of Nanjing East Road Pedestrian Street, it is detected that 62% of the pedestrians move towards the due west direction, 28% diverge towards the northwest direction, and the remaining 10% stay and observe. The extraction of the target moving direction adopts a method combining kernel density estimation and direction clustering. For example, a density peak is detected in the direction interval of 0° - 30°, and the dominant flow direction of the pedestrians is determined to be 10° north of west. The data sampling interval is 500 milliseconds. When the standard deviation of the detected dominant flow direction exceeds 15°, the direction stability check is triggered to exclude the interference of temporary direction fluctuations. The vector representation of the dominant flow direction includes two dimensions: angle and intensity. For example, the intensity coefficient in the direction of 10° north of west is 0.78, which is used to quantify the significance of the moving trend of the pedestrian group.

[0136] Step S442: Based on the dominant flow direction of the pedestrians and the vehicle passing directions of each lane within the emergency response area, determine a set of target conflict lanes that have cross conflicts with the dominant flow direction of the pedestrians.

[0137] Exemplarily, the lane passing direction data can be sourced from a pre-deployed general road network digital twin model. For example, the direction angle of the north-south straight lane on the east side of the Middle East Road is 175°, and the intersection angle formed with the dominant flow direction of the pedestrians, which is 10° north of west, is 15°, and it is determined as a potential conflict point. Conflict detection uses geometric topology analysis. When the angle between the lane center line and the extension line of the dominant flow direction of the pedestrians is less than 45° and the projection distance is less than 5 meters, it is marked as a first-level conflict lane. For example, it is detected that the west-east left-turn lane on Jiangxi Middle Road (direction angle 85°) forms a 75° intersection angle with the dominant flow direction of the pedestrians. Since the projection distance is only 2.3 meters, it is included in the set of target conflict lanes. The conflict level division considers the traffic flow intensity. When the hourly traffic volume of the conflict lane exceeds 800 vehicles, it is automatically upgraded to a special-level conflict lane, triggering more stringent control measures.

[0138] Step S443: According to the real-time vehicle flow and pedestrian density of each lane in the set of target conflict lanes, predict the vehicle arrival peak moment and the pedestrian waiting overrun moment in a preset future time interval.

[0139] Exemplarily, the prediction of the vehicle arrival peak can be achieved by using a general autoregressive integrated moving average model. For example, in the conflict lane of Jiangxi Middle Road, based on the current arrival rate of 28 vehicles per minute, the model predicts that the peak will reach 42 vehicles per minute after 6 minutes and 15 seconds. The moment when the pedestrian waiting time exceeds the limit is calculated by a survival analysis model. When the pedestrian density exceeds 1.2 people per square meter and the average waiting time exceeds 90 seconds, an over-limit warning is triggered. For example, the current average pedestrian waiting time is 68 seconds, and the model predicts that the threshold will be reached after 4 minutes and 50 seconds. The prediction process implements a dynamic correction mechanism, and the latest observed data is input into the model every 30 seconds to recalibrate the parameters. For example, when a sudden passenger flow is detected at the subway station, the growth rate parameter of the pedestrian waiting time is increased from 5% per minute to 8%.

[0140] Step S444: Based on the overlapping time period between the vehicle arrival peak moment and the pedestrian waiting over-limit moment, generate multiple candidate sets of pedestrian passing time windows, and each candidate set of pedestrian passing time windows includes a start time and a duration.

[0141] Exemplarily, the generation of the candidate set of time windows follows the principle of phase coordination. For example, it is detected that the vehicle peak moment is T + 6:15 and the pedestrian over-limit moment is T + 4:50. Take the overlapping time period from T + 4:50 to T + 6:15 to generate three candidate sets: candidate set A (from T + 5:00 to T + 5:30), candidate set B (from T + 5:30 to T + 6:00), and candidate set C (from T + 5:15 to T + 6:00). The duration setting meets the requirements of the minimum executable unit. For example, it is not less than 20 seconds and not more than 90 seconds to avoid signal cycle fragmentation. The start time of the candidate set is aligned with the existing signal phase cycle. For example, when the remaining time of the current phase is 12 seconds, the start time of candidate set A is set to be activated immediately after the end of the current cycle.

[0142] Step S445: Through the edge computing node, evaluate the influence weight of each candidate set of pedestrian passing time windows on the overall traffic efficiency of the emergency response area, and select the candidate set with the highest influence weight as the final dedicated pedestrian passing time window.

[0143] Exemplarily, the influence weight evaluation model integrates a multi-objective optimization function to calculate the quantitative influence of each candidate set on indicators such as vehicle delay, pedestrian safety, and emergency rescue. For example, candidate set B causes an increase in the total vehicle delay of 85 vehicle-hours, but reduces the probability of pedestrian injury risk by 23%. After weighted calculation, the comprehensive score is 72 points; candidate set C causes an increase in delay of 120 vehicle-hours but a risk reduction of 35%, and the score is 68 points. The real-time constraint conditions are introduced into the evaluation process. For example, when it is detected that the ambulance is expected to pass at T + 5:45, candidate set B is excluded due to time overlap. Finally, the candidate set A with the highest score is selected, and its time window parameters are written into the signal control queue.

[0144] Step S446: Assign an independent control instruction to the final pedestrian-only passing time window in the direction passing priority parameter, and simultaneously prohibit the green light signal of the target conflict lane set within the final pedestrian-only passing time window.

[0145] Exemplarily, the independent control instruction follows the NTCIP protocol standard format. For example, generate the instruction "Phase PedestrianEW Start = 17:23:45 Duration = 30", specifying that the east-west pedestrian phase starts at the specified moment and lasts for 30 seconds. The green light prohibition of the target conflict lane is achieved through a phase mask. For example, in the west-to-east left-turn phase parameter table of Jiangxi Middle Road, mark the period from T+5:00 to T+5:30 as the red light forced period. After the instruction is issued, a double-check mechanism is implemented. The status of the signal controller is confirmed through the signal transmitted back by the roadside unit. When an abnormal phase switch is detected, a backup control scheme is automatically triggered. After the time window ends, the original phase parameters are restored, and an effect evaluation process is started to count the actual reduced pedestrian waiting time and the caused vehicle delay data for optimizing the subsequent decision-making model.

[0146] As an optional implementation manner, after step S400 dynamically adjusts the execution priority of the control instruction sequence, the method may further include:

[0147] Step S500: Continuously collect the traffic flow recovery indicators of the emergency response area. The traffic flow recovery indicators include the vehicle queue length decrease rate, the pedestrian average waiting time shortening rate, and the lane passing capacity recovery coefficient.

[0148] The traffic flow recovery indicators are a set of comprehensive evaluation parameters that quantify the degree of the traffic system's return from an abnormal state to normal operation. Their data collection is realized relying on the detection device network deployed at the boundaries and core positions of the emergency response area. The vehicle queue length decrease rate is calculated in real time through a video recognition algorithm. For example, the number of queuing vehicles 200 meters north of the accident point has decreased from a peak of 58 to the current 32. The decrease rate per minute is calculated as (58 - 32) / (58×5) = 8.97%. The pedestrian average waiting time shortening rate is statistically calculated based on the time stamp difference of the pedestrian crossing request signal controller. For example, the average waiting time of pedestrians at the southeast crosswalk has been optimized from 85 seconds at the initial stage of the disposal to the current 63 seconds. The shortening rate is calculated as (85 - 63) / 85×100% = 25.88%. The lane passing capacity recovery coefficient is determined by the ratio of the cross-sectional flow to the free flow speed. For example, the current passing capacity of the eastward detour lane is 1200 vehicles per hour, and the recovery coefficient compared to the reference value of 1600 vehicles is 0.75. The data collection frequency is synchronized with the signal control cycle, and a recovery status snapshot containing three-dimensional indicators is generated every 30 seconds and stored in the circular buffer of the edge computing node for real-time analysis.

[0149] Step S600: Compare the traffic flow recovery index with a preset recovery benchmark curve, and calculate the deviation between the actual recovery progress and the expected recovery progress.

[0150] Exemplarily, the preset recovery benchmark curve can be derived from a weighted average model of the disposal effects of historical similar events. For example, in the scenario of a secondary traffic accident, the benchmark curve stipulates that the vehicle queue length decline rate should reach 35% within 30 minutes after the accident, and the pedestrian waiting time shortening rate should not be less than 20%. The deviation calculation uses the dynamic time warping algorithm to align the time axis of the actual measurement sequence with the benchmark curve. For example, the actual vehicle queue decline rate reaches 28% at the 10th minute, while the corresponding time of the benchmark curve is 32%. After time alignment, the root mean square error is calculated to be 4.2 percentage points. The comprehensive deviation of multi-dimensional indicators is aggregated by the entropy weight method, with a weight of 0.5 for vehicle indicators, 0.3 for pedestrian indicators, and 0.2 for traffic capacity indicators. When the actual vehicle queue decline rate error is 8%, the pedestrian waiting shortening rate error is -5% (reverse growth), and the traffic capacity error is 12%, the comprehensive deviation is calculated as 8×0.5 + 5×0.3 + 12×0.2 = 7.9. The deviation threshold is dynamically set according to the event level. For example, the upper limit of the allowable deviation for a secondary response is 10, and an alarm is triggered when the detected current value reaches 12.3.

[0151] Step S700: If the deviation exceeds the allowable error range, trigger the parameter recalibration mechanism of the traffic signal adaptive model, and the parameter recalibration mechanism includes the following steps:

[0152] Step S701: Freeze the weight parameters of the convolutional branch of the traffic signal adaptive model, and unlock the hidden layer nodes of the recurrent neural network branch.

[0153] Exemplarily, the weight freezing operation is achieved by setting a gradient mask, and only the parameters of the recurrent neural network branch are updated during the backpropagation process. For example, the weight matrix of the 128-dimensional spatial feature extraction layer of the convolutional branch is locked, while the input gate and forget gate parameters of the 64 gated recurrent unit nodes of the recurrent neural network branch remain trainable. Unlocking the hidden layer nodes involves adjusting the slope of the activation function. For example, the output gate tanh function of the long short-term memory unit is replaced with the LeakyReLU function to enhance the short-term memory ability. During parameter recalibration, the model retains the ability to extract spatial features but strengthens the adaptive learning of the temporal evolution pattern. For example, when it is detected that the change in night lighting conditions leads to an increase in video detection errors, the robustness of the time series prediction module is focused on optimization.

[0154] Step S702: Inject the current traffic state feature vector into the recurrent neural network branch, and recalculate the direction weight allocation ratio of the green light duration parameter.

[0155] Exemplarily, the current traffic state feature vector is generated by encoding real-time sensor data. For example, the queue length of the eastbound lane is 32 vehicles, the pedestrian waiting time is 63 seconds, and the traffic capacity is 0.75. The indicators are normalized to [0.4, 0.63, 0.75] input feature vectors. The directional weight distribution ratio is dynamically adjusted through the attention mechanism. For example, the initial weight of the northbound bypass lane at the accident point is 0.7. After injecting the new feature vector, the model detects the increase in eastbound pressure and increases its weight to 0.8, while reducing the weight of the westbound secondary lane to 0.2. The recalculation process performs five rounds of iterative optimization, and each round injects the latest traffic state snapshot. For example, after the third round of iteration, the green light duration of the north-south main road increases from 55 seconds to 60 seconds, and the east-west auxiliary road decreases from 40 seconds to 35 seconds, and the weight ratio is adjusted to [0.6, 0.4].

[0156] Step S703: When the deviation drops to within the allowable error range, the parameter recalibration mechanism is released and the model full parameter update mode is restored.

[0157] Exemplarily, the deviation monitoring adopts a sliding window mechanism. For example, when the comprehensive deviation is lower than the threshold value of 10 for three consecutive control cycles (90 seconds), the recovery is judged to be effective. The release mechanism is implemented by gradient mask removal, the weights of the graph convolution branches are re-added to the training parameter set, and the activation function of the recurrent neural network branch is restored to the original configuration. The full parameter update mode adopts an elastic weight consolidation algorithm to retain the temporal pattern features learned during recalibration. For example, an elastic coefficient of 0.3 is applied to the weight changes of the recurrent neural network branches in subsequent training to prevent catastrophic forgetting. An integrity check is performed immediately after the model is restored. For example, it verifies whether the recognition accuracy of the spatial feature extraction module for the newly added construction fence area meets the standard. If it fails, the supplementary training process is triggered.

[0158] The activation conditions of the parameter recalibration mechanism include:

[0159] Condition 1: It is detected that the vehicle queue length reduction rate does not reach the expected threshold value for three consecutive control cycles.

[0160] Exemplarily, the control cycle is synchronized with the phase switching cycle of the traffic light. For example, in the 90-second fixed cycle mode, the mechanism is activated when the vehicle queue decline rate is lower than 80% of the corresponding value of the baseline curve in three consecutive cycles (4 minutes and 30 seconds). The expected threshold is dynamically adjusted according to the disposal stage. For example, the threshold for 0-15 minutes after the accident is set to a decline rate of 5% per minute, and it is increased to 8% for 15-30 minutes. When the actual monitoring value is measured at a decline rate of 3.2%, 4.1%, and 3.8% respectively during the period of 18:15-18:19:30, it is determined that three consecutive cycles are not up to standard, triggering parameter recalibration.

[0161] Or, Condition 2: The reduction rate of the average pedestrian waiting time shows a reverse growth trend within a preset time period.

[0162] Exemplarily, the reverse growth trend is determined by the sign of the linear regression slope. For example, within a 10-minute monitoring window, the pedestrian waiting time increases from 62 seconds to 68 seconds, and the reduction rate is calculated as -8.9%, with a positive slope value. The preset time period is set according to the intensity of pedestrian crossing demand. For example, it is set to 5 minutes during the evening peak period and 10 minutes during the off-peak period. When it is detected that the waiting time continues to increase and the slope exceeds 0.5 seconds / minute during the period from 19:00 to 19:05, the calibration mechanism is immediately activated.

[0163] Or, Condition 3: The difference between the lane capacity recovery coefficient and the cooperative control index of the adjacent intersection exceeds the cooperative tolerance.

[0164] Exemplarily, the cooperative tolerance is set based on the principle of regional traffic balance. For example, within a 500-meter radius linkage control range, the maximum allowable difference is 15%. When the recovery coefficient of the eastward detour lane at the accident point is 0.75, while the corresponding coefficient at the adjacent Middle East First Road intersection reaches 0.95, and the difference of 20% exceeds the threshold, cross-intersection parameter calibration is triggered. The cooperative control index is obtained in real time through V2X communication. For example, the phase state and traffic flow data of the signal machine at the adjacent intersection are received, and the spectral radius difference of the coordination coefficient matrix is calculated.

[0165] Please refer to Figure 3 , which is the structural block diagram of the control device 120 of the present invention. The control device 120 includes a calculation unit 1001, which can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 1002 or the computer program loaded from the storage unit 1008 into the random access memory (RAM) 1003. In the RAM 1003, various programs and data required for the operation of the control device 120 can also be stored. The calculation unit 1001, the ROM 1002, and the RAM 1003 are connected to each other through a bus 1004. The input / output (I / O) interface 1005 is also connected to the bus 1004. In other words, the control device 120 provided in the embodiment of the present application includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned edge computing traffic light emergency control method for sudden traffic events.

[0166] A plurality of components in the control device 120 are connected to the I / O interface 1005, including: an input unit 1006, an output unit 1007, a storage unit 1008, and a communication unit 1009. The input unit 1006 can be any type of device capable of inputting information to the control device 120. The input unit 1006 can receive input digital or character information, and generate key signal inputs related to user settings and / or function controls of the server, and can include, but are not limited to, a mouse, a keyboard, a touch screen, a trackpad, a trackball, a joystick, a microphone, and / or a remote control. The output unit 1007 can be any type of device capable of presenting information, and can include, but are not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1008 can include, but are not limited to, magnetic disks and optical discs. The communication unit 1009 allows the control device 120 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include, but are not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, for example, a BluetoothTM device, an 802.11 device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0167] The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 executes the various methods and processes described above, for example, the edge computing traffic light emergency control method for sudden traffic events. For example, in some embodiments, the edge computing traffic light emergency control method for sudden traffic events can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, for example, the storage unit 1008. In some embodiments, part or all of the computer program can be loaded and / or installed onto the control device 120 via the ROM 1002 and / or the communication unit 1009. When the computer program is loaded into the RAM 1003 and executed by the computing unit 1001, one or more steps of the edge computing traffic light emergency control method for sudden traffic events described above can be executed. Alternatively, in other embodiments, the computing unit 1001 can be configured to execute the edge computing traffic light emergency control method for sudden traffic events in any other suitable manner (for example, by means of firmware).

[0168] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps recited in the present invention can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved, and no limitation is imposed herein.

[0169] Although embodiments or examples of the present invention have been described with reference to the accompanying drawings, it should be understood that the above methods, systems, and devices are merely exemplary embodiments or examples, and the scope of the present invention is not limited by these embodiments or examples, but is only defined by the authorized claims and their equivalent scope. Various elements in the embodiments or examples can be omitted or replaced by their equivalent elements. In addition, the steps can be executed in an order different from that described in the present invention. Further, the various elements in the embodiments or examples can be combined in various ways. Importantly, with the evolution of technology, many of the elements described herein can be replaced by equivalent elements that emerge after the present invention.

Claims

1. An edge computing traffic light emergency control method for sudden traffic events, characterized in that: The method comprises: Acquire real-time traffic flow information of the target intersection, wherein the real-time traffic flow information includes vehicle density distribution data, pedestrian activity trajectory data, and location identification of sudden traffic events; Based on the location identifier of the sudden traffic event, the emergency response area of ​​the target intersection is determined, and the dynamic offset between the historical traffic flow characteristics and the current traffic flow characteristics in the emergency response area is extracted through the edge computing node; Inputting the dynamic offset into a pre-trained traffic signal adaptive model to generate a set of traffic signal light control parameters corresponding to the emergency response area, wherein the set of traffic signal light control parameters includes a phase switching frequency parameter, a green light duration parameter, and a directional traffic priority parameter; According to the traffic light control parameter set, a control instruction sequence is sent to the traffic light at the target intersection, and based on the real-time change trend of the vehicle density distribution data and the pedestrian activity trajectory data, the execution priority of the control instruction sequence is dynamically adjusted.

2. The method according to claim 1, characterized in that The extracting of the dynamic offset between the historical traffic flow characteristics and the current traffic flow characteristics in the emergency response area through the edge computing node includes: Retrieving historical traffic flow characteristics of the emergency response area within a preset time window from a local database of the edge computing node, the historical traffic flow characteristics including historical average vehicle speed, historical pedestrian waiting time, and historical congestion index; Performing time dimension alignment processing on the current traffic flow feature so that the timestamp of the current traffic flow feature is consistent with the time window range of the historical traffic flow feature; Calculating a speed difference coefficient between the historical average vehicle speed and the current average vehicle speed, and a waiting time difference coefficient between the historical pedestrian waiting time and the current pedestrian waiting time; According to the speed difference coefficient and the waiting time difference coefficient, a multidimensional feature vector of the dynamic offset is constructed, and the multidimensional feature vector is scaled uniformly through the normalization layer of the edge computing node to obtain a normalized multidimensional feature vector, and the multidimensional feature vector is used to characterize the dynamic offset.

3. The method according to claim 2, characterized in that The step of inputting the dynamic offset into a pre-trained traffic signal adaptive model to generate a set of traffic signal light control parameters corresponding to the emergency response area includes: Calling the feature fusion layer of the traffic signal adaptive model, spatially correlating and encoding the dynamic offset with the location identifier of the sudden traffic event, and generating an event perception feature map; Predicting the vehicle arrival rate change curve and pedestrian gathering density change curve of the emergency response area in the future time interval based on the event perception feature graph through the time series prediction layer of the traffic signal adaptive model; Determining an initial value of the phase switching frequency parameter according to an intersection of the vehicle arrival rate variation curve and the pedestrian gathering density variation curve; Based on the initial value, the directional weight of the green light duration parameter is iteratively adjusted through the parameter optimization layer of the traffic signal adaptive model until the directional traffic priority parameter satisfies the preset conflict avoidance constraint condition.

4. The method according to claim 3, characterized in that: The traffic signal adaptive model is trained by the following steps: Collecting training data corresponding to multiple historical traffic emergencies, the training data includes real-time traffic flow data when the event occurs, traffic signal adjustment records after the event is handled, and traffic recovery efficiency indicators after the event ends; Performing spatiotemporal slicing processing on the real-time traffic flow data to generate training samples matching the dimensions of the emergency response area, and converting the traffic signal adjustment records into a label parameter set of the training samples; Constructing a neural network structure of an initial traffic signal adaptation model, wherein the neural network structure includes a graph convolution branch for extracting spatial dependencies and a recurrent neural network branch for capturing temporal dependencies; Inputting the training samples into the initial traffic signal adaptive model, calculating the mean square error loss between the label parameter set and the model output parameter, and updating the connection weights of the neural network structure by a gradient descent algorithm; When the improvement of the traffic recovery efficiency index on the validation data set reaches a preset threshold, the training is stopped and the model parameters are deployed to the edge computing node.

5. The method according to claim 1, characterized in that The dynamically adjusting the execution priority of the control instruction sequence based on the real-time change trend of the vehicle density distribution data and the pedestrian activity trajectory data includes: Monitoring the position coordinates and movement direction of new vehicles in the emergency response area, and calculating the shortest path distance between the new vehicles and the location identifier of the sudden traffic incident; Predicting the time interval for the newly added vehicles to arrive at the sudden traffic event area according to the shortest path distance and the current average speed of the vehicles; If the time interval is less than the preset emergency response time threshold, the green light duration parameter of the lane where the newly added vehicle is located is adjusted up to the first priority level; The abnormal gathering area in the pedestrian activity trajectory data is synchronously detected. If the abnormal gathering area overlaps with the location identifier of the sudden traffic event, a pedestrian-only passage time window is allocated in the directional passage priority parameter.

6. The method according to claim 5, characterized in that The allocating of a pedestrian-only passage time window in the passage priority parameter of the direction includes: Obtaining the real-time movement direction distribution of pedestrians in the abnormal gathering area, and extracting the target movement direction with the highest frequency in the real-time movement direction distribution as the dominant flow direction of pedestrians; Based on the pedestrian dominant flow direction and the vehicle travel directions of each lane in the emergency response area, determining a target conflict lane set that has an intersection conflict with the pedestrian dominant flow direction; According to the real-time vehicle flow and pedestrian density of each lane in the target conflict lane set, predict the vehicle arrival peak time and pedestrian waiting time of the target conflict lane within a preset future time interval; Based on the overlapping time period between the vehicle arrival peak time and the pedestrian waiting time, a plurality of pedestrian passage time window candidate sets are generated, each of which includes a start time and a duration; Evaluate the influence weight of each candidate set of pedestrian travel time windows on the overall travel efficiency of the emergency response area through the edge computing node, and select the candidate set with the highest influence weight as the final pedestrian-only travel time window; An independent control instruction is allocated to the final pedestrian-only passage time window in the directional passage priority parameter, and the green light signal of the target conflicting lane set within the final pedestrian-only passage time window is synchronously prohibited.

7. The method according to claim 1, characterized in that The location identifier of the traffic emergency is obtained by the following steps: Receive real-time video stream data uploaded by the monitoring device at the target intersection, and perform moving target detection on each frame of the real-time video stream data; Extracting a position change sequence of a detected moving target and calculating a position offset of the moving target between adjacent frames; If the position offset exceeds the preset abnormal displacement threshold continuously, the area where the moving target is located is determined to be a suspected sudden traffic incident area; Sending a high-resolution snapshot command to the monitoring device through the edge computing node to obtain multi-angle detailed images of the suspected sudden traffic incident area; The multi-angle detail image is classified into event types, and if the classification result is a vehicle collision or a road obstacle, the suspected sudden traffic event area is marked as a location identifier of the sudden traffic event.

8. The method according to claim 7, characterized in that The classifying the event type of the multi-angle detail image includes: Inputting the multi-angle detail image into a pre-trained event classification model to extract key object contour features and scene context features in the image; Recognize visual patterns of scattered vehicle parts, cracks on the road surface, or fallen pedestrians based on the key object contour features; Analyze the sudden change in speed of adjacent vehicles or the collective turning behavior of the movement direction of a group of pedestrians based on the scene context features; If the scattered vehicle parts are identified and there is a sudden change in speed, the event type is determined to be a vehicle collision; If the road crack is identified and there is a collective turning behavior of the pedestrian group in the direction of movement, the event type is determined to be a road obstacle.

9. The method according to claim 1, characterized in that: After dynamically adjusting the execution priority of the control instruction sequence, the method further includes: Continuously collecting traffic flow recovery indicators in the emergency response area, wherein the traffic flow recovery indicators include a vehicle queue length reduction rate, a pedestrian average waiting time reduction rate, and a lane capacity recovery coefficient; Comparing the traffic flow recovery index with a preset recovery benchmark curve, and calculating the deviation between the actual recovery progress and the expected recovery progress; If the deviation exceeds the allowable error range, a parameter recalibration mechanism of the traffic signal adaptive model is triggered, and the parameter recalibration mechanism includes the following steps: Freeze the weight parameters of the graph convolution branch of the traffic signal adaptive model, and unlock the hidden layer nodes of the recurrent neural network branch; Injecting the current traffic state feature vector into the recurrent neural network branch, and recalculating the direction weight distribution ratio of the green light duration parameter; When the deviation drops to within the allowable error range, the parameter recalibration mechanism is released and the model full parameter update mode is restored; The activation conditions of the parameter recalibration mechanism include: It is detected that the vehicle queue length reduction rate does not reach the expected threshold value for three consecutive control cycles; Or, the pedestrian average waiting time reduction rate shows a reverse growth trend within a preset time period; Or, the difference between the lane capacity recovery coefficient and the collaborative control index of the adjacent intersection exceeds the collaborative tolerance.

10. A control device, characterized in that: include: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1 to 9.

Images