Highway parking risk dynamic early warning method and system based on spatial-temporal characteristics and adaptive threshold

Abstract

The invention relates to the technical field of intelligent traffic, and discloses an expressway parking risk dynamic early warning method and system based on spatial-temporal characteristics and an adaptive threshold, and the method comprises the steps: collecting vehicle driving data through a multi-type sensor, obtaining vehicle internal data through 5G communication, and employing a D-S evidence theory for fusion; performing spatio-temporal feature processing on the fused data, constructing a spatio-temporal matrix, and extracting features by using a spatio-temporal convolutional neural network; constructing a parking risk assessment model by combining a gating circulation unit of an attention mechanism based on the extracted features, and outputting a parking risk probability in a period of time in the future; according to road sections, time periods and weather conditions, an early warning threshold value is dynamically adjusted by using a reinforcement learning algorithm, multi-channel early warning is triggered when the threshold value is exceeded, and the system comprises a data acquisition and transmission subsystem, a data processing and analysis subsystem and an early warning decision and release subsystem. The objective of the invention is to realize prediction and timely and effective early warning of the parking risk of the high-speed vehicle.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a method and system for dynamic early warning of highway parking risks based on spatiotemporal characteristics and adaptive thresholds. Background Art

[0002] As the traffic volume on highways continues to grow, the safety risks of vehicle driving continue to rise. If the parking behavior of highway vehicles cannot be detected and warned in time, it is very easy to cause serious traffic accidents such as rear-end collisions, which not only seriously threatens the lives of drivers and passengers, but also greatly reduces road traffic efficiency. At present, with the vigorous development of intelligent transportation systems, the use of big data technology to accurately predict and timely warn the risks of high-speed vehicle parking has become a core requirement for improving the level of highway traffic safety. With its ability to integrate multi-source heterogeneous data and mine potential information, big data technology can provide a more comprehensive and accurate basis for vehicle parking warnings, which is of great significance to ensuring the safe and efficient operation of highways.

[0003] At present, there are various means of data collection, but the data collected by different sensors have problems such as accuracy differences and inconsistent data formats. For example, the camera is greatly affected by the weather, and the capture effect of the vehicle shape and driving trajectory is not good in rainy and foggy weather; the geomagnetic sensor is easily disturbed by the surrounding environment, resulting in inaccurate speed change monitoring. In addition, the multi-source data fusion technology is not yet mature, and it is difficult to fully explore the internal connection between the data. The quality and accuracy of the fused data are difficult to meet the high requirements of parking risk warning; most of the existing parking risk assessment models consider a single factor and cannot comprehensively integrate multiple factors such as vehicle driving status, road environment, and vehicle status. Moreover, the traditional model has poor adaptability to complex traffic scenes and emergencies. Under different road sections, time periods and weather conditions, the prediction accuracy of the model fluctuates greatly, making it difficult to achieve stable and reliable parking risk prediction. At the same time, the warning threshold usually adopts a fixed value and cannot be dynamically adjusted according to the actual traffic conditions. This leads to false alarms or missed alarms in cases of large traffic flow and complex road conditions. At the same time, the existing warning system has delays in the information transmission process, and cannot convey the warning information to relevant personnel in a timely manner, which affects the driver's effective emergency measures and reduces the actual application value of the warning system.

[0004] For example, patent number CN202311656749.7, the name of the patent is a highway event warning system and comprehensive perception method, which uses millimeter wave radar to collect vehicle driving feature information, and performs noise removal and behavior recognition through data processing and event detection units to judge abnormal events and determine the degree of driving danger. For ramp merging areas and multi-lane driving areas, relevant parameters are calculated through specific algorithms to evaluate potential risks, and then warning classification is carried out according to the abnormal event situation, and the information is transmitted to the comprehensive perception warning prompt platform and the user's mobile terminal with the help of the warning prompt unit; the advantage is that the millimeter wave radar has high resolution and strong anti-interference ability, and is less affected by bad weather. It can collect various characteristics of vehicles in real time in complex environments, providing a reliable data basis for subsequent analysis. Special warning algorithms are designed for common dangerous scenes on highways, such as ramp merging areas and multi-lane driving areas. By calculating key parameters such as the time and space occupancy rate of vehicles arriving at potential conflict points, risks can be accurately assessed, and monitoring and graded warnings of various abnormal events can be achieved. However, this solution mainly relies on millimeter-wave radar to collect data, and the data source is relatively single, and it may not be able to fully obtain all the information that affects highway driving safety, such as the vehicle's own mechanical fault information, the driver's physical condition information, etc. Algorithm models are mostly built based on specific scenarios and assumptions. For example, when calculating the time for ramp vehicles and main road vehicles to reach potential conflict points, simplified assumptions are made about the vehicle's driving status. In actual complex and changeable traffic conditions, the accuracy and universality of the model may be affected. Although the overall system is committed to achieving real-time warnings, there may be delays in data processing, calculation of complex parameters, and information transmission, resulting in the failure of warnings to be delivered to drivers in a timely manner, affecting the driver's timely handling of dangerous situations. Moreover, most of these methods do not fully consider key issues such as data fusion and dynamic adjustment of warning thresholds, and cannot effectively solve the problems in actual production. Summary of the invention

[0005] The present invention provides a method and system for dynamic early warning of highway parking risks based on spatiotemporal characteristics and adaptive thresholds, which can predict and promptly and effectively warn of the parking risks of high-speed vehicles.

[0006] The basic solution provided by the present invention is: The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds includes the following steps: S01 Data collection and fusion: Use cameras, geomagnetic sensors, and lidar to collect vehicle driving data. The camera captures the vehicle's appearance and driving trajectory characteristics, the geomagnetic sensor monitors the vehicle's passing and speed change characteristics, and the lidar measures the vehicle's distance, speed, and angle in real time; 5G communication is used to obtain vehicle working conditions, brake system status, tire pressure, and temperature data in real time, and DS evidence theory is used for data fusion. The DS evidence theory is used for data fusion, and the formula is: in, is the basic probability distribution function of proposition A after fusion, is the i-th data source for the proposition The basic probability distribution function of , k is a normalization constant; S02 Feature extraction based on spatiotemporal features: Build a spatiotemporal matrix with continuously collected and fused data. The spatial dimensions in the spatiotemporal matrix include the position of the vehicle on the highway, the lane, and the relative position information of surrounding vehicles. Use spatiotemporal convolutional neural network to extract features, extract the spatiotemporal characteristics of vehicle speed change trend and driving trajectory change; S03 Parking risk assessment and prediction model construction: Based on the extracted features, a parking risk assessment model is constructed by combining the gated recurrent unit with the attention mechanism to output the risk probability of the vehicle parking in the future; S04 Dynamic threshold adaptive adjustment and warning triggering: According to different sections of the highway, different time periods and weather conditions, the parking risk warning threshold is dynamically adjusted using the reinforcement learning algorithm. The action of adjusting the threshold can be selected according to the current state, and rewards are fed back according to the action. The reward function formula is: in, , , is the weight coefficient, Y is the accuracy, M is the recall rate, and U is the false alarm rate. When the predicted parking risk probability is greater than the dynamically adjusted threshold, the multi-channel early warning mechanism is triggered to issue early warning information for the target vehicle and target road section.

[0007] The highway parking risk dynamic warning system based on spatiotemporal characteristics and adaptive thresholds is characterized by including the following systems Data collection and transmission subsystem: deploy cameras to cover the highway section, and the cameras have auto-focus and low-light enhancement functions; The geomagnetic sensor can detect tiny vehicle vibrations with a response time of less than 0.02 seconds; LiDAR scanning generates point cloud data in real time; The Internet of Vehicles communication module supports 5G high-speed communication; Data processing and analysis subsystem: The data fusion unit performs data fusion processing based on DS evidence theory to eliminate data errors and conflicts; The spatiotemporal feature extraction unit uses the ST-CNN model to extract features from the fused data; the parking risk assessment model unit is based on the GRU-Attention model and combines the attention mechanism to assess and predict vehicle parking risks; Early warning decision and release subsystem: The dynamic threshold adjustment module uses a reinforcement learning algorithm to dynamically adjust the warning threshold according to real-time road conditions and environmental factors; The early warning release module releases early warning information through 5G short messages, highway electronic display screens, and traffic broadcasts.

[0008] The principles and advantages of the present invention are: The data acquisition and transmission subsystem uses cameras, geomagnetic sensors, lidar, and Internet of Vehicles communication modules to collect vehicle driving data, vehicle operating conditions, brake system status, tire pressure and temperature data from multiple dimensions. Using DS evidence theory, the data is integrated in the data processing and analysis subsystem and converted into a unified and usable form to reduce data errors and conflicts.

[0009] Next, a spatiotemporal convolutional neural network is used to construct a spatiotemporal matrix, and the fused data is subjected to feature extraction to mine the spatiotemporal characteristics of vehicle travel, such as speed change trends and driving trajectory changes. Then, a parking risk assessment model is constructed by combining a gated recurrent unit with an attention mechanism. This model can effectively process time series features, focus on key information through an attention mechanism, and then accurately output the risk probability of a vehicle parking in the future. Finally, the warning decision and release subsystem uses a reinforcement learning algorithm to dynamically adjust the parking risk warning threshold based on real-time road conditions and environmental factors such as different sections of the highway, different time periods, and weather conditions. When the predicted parking risk probability is greater than the dynamically adjusted threshold, warning information is issued in a timely manner to the target vehicle and target section through multiple channels such as 5G short messages, highway electronic display screens, and traffic broadcasts, so as to achieve effective warning of parking risks.

[0010] Compared with the existing technology, this solution adopts multi-sensor collaborative collection, covering rich data such as vehicle appearance, driving trajectory, speed changes, vehicle operating conditions, etc. The comprehensiveness of the data far exceeds the single sensor collection method. At the same time, the application of DS evidence theory can better integrate multi-source data, reduce errors and conflicts, and improve data quality. Traditional methods often have problems of information loss or inaccuracy when processing multi-source data.

[0011] In feature extraction and risk assessment model construction, the combination of spatiotemporal convolutional neural network and gated recurrent unit combined with attention mechanism can more accurately capture the spatiotemporal characteristics of vehicle travel and assess parking risks. Compared with simple machine learning algorithms, this solution considers more comprehensive factors, has stronger model generalization ability, can adapt to complex and changeable traffic scenarios, and improves the accuracy of parking risk prediction.

[0012] In the process of adjusting and issuing warning thresholds, the present invention utilizes reinforcement learning algorithms to dynamically adjust thresholds, optimizes warning strategies according to actual road conditions and environmental factors, effectively reduces false alarm and missed alarm rates, and uses multi-channel warning issuance methods to ensure that warning information is promptly conveyed to target vehicles and relevant personnel on the road sections. The traditional fixed threshold warning method lacks flexibility, and the warning issuance channels are single and untimely, making it difficult to meet the needs of actual traffic scenarios.

[0013] Furthermore, in S02, a spatiotemporal convolutional neural network is used for feature extraction. The network structure includes a spatiotemporal convolution layer, a pooling layer, and a fully connected layer. In the spatiotemporal convolution layer, the spatiotemporal convolution kernel slides in the time and space dimensions. The convolution formula is:

[0014] Among them, the It is the Lth layer at position The output, is the spatiotemporal convolution kernel weight of the Lth layer, It is L Layer in position Input, is a bias term. Through the spatiotemporal convolution operation, the spatiotemporal characteristics of vehicle speed change trend and driving trajectory change are extracted. The spatiotemporal convolution kernel slides in the time and space dimensions, and can simultaneously consider the information of vehicles at different times and different spatial positions. The driving status of vehicles on the highway is constantly changing. The speed change trend and driving trajectory change are key dynamic information for judging whether the vehicle has a parking risk. Through this convolution operation, the speed fluctuation during the vehicle driving process can be effectively captured, such as whether the vehicle suddenly decelerates or accelerates, and whether the driving trajectory deviates abnormally; the highway scene is complex, and factors such as vehicle density, driving direction, and weather conditions will affect vehicle driving. The speed change trend and driving trajectory change characteristics extracted by the spatiotemporal convolutional neural network can comprehensively consider these complex factors. For example, under different traffic flow conditions, by analyzing the speed change trend, it can be judged whether the vehicle is driving normally or may stop due to traffic congestion; according to the change of the driving trajectory, it can be judged whether the vehicle is unstable due to bad weather such as heavy rain and fog, thereby increasing the parking risk. This adaptability to complex scenes makes the early warning method more suitable for practical applications.

[0015] Furthermore, in S03, a parking risk assessment model is constructed by combining a gated recurrent unit with an attention mechanism, and the attention mechanism formula is:

[0016]

[0017]

[0018] in, It is a feature With context vector The relevance score of is the attention weight, is the weighted context vector; In the GRU unit, the update gate formula is:

[0019] The reset gate formula is:

[0020] The candidate hidden state formula is:

[0021] The hidden state update formula is:

[0022] in, , are the outputs of the update gate and the reset gate, respectively. is the sigmoid activation function, , , is the weight matrix, , , is the bias vector, is the hidden state at the previous moment, is the input data at the current moment, is a candidate hidden state. The GRU unit processes the time series features and outputs the risk probability of the vehicle parking in the future. The driving state of vehicles on the highway changes over time and has obvious time series characteristics. The GRU unit can handle this kind of time series data well. Through mechanisms such as updating gates, resetting gates and candidate hidden states, it can remember the key information in the vehicle's past driving state and conduct a comprehensive analysis in combination with the current input data. For example, information such as the speed change of the vehicle in the previous period and the coherence of the driving trajectory can be effectively memorized and utilized by the GRU unit, so as to more accurately predict the future parking risk probability; traditional recurrent neural networks are prone to gradient vanishing problems when processing long time series, making it difficult for the model to learn long-distance dependencies. The GRU unit alleviates the gradient vanishing problem to a certain extent compared to traditional RNNs by simplifying the gating mechanism. This enables the model to more effectively propagate gradients and learn more comprehensive time series features when processing long time series of vehicle driving data, thereby more accurately evaluating the parking risk probability.

[0023] Furthermore, in the step of S01 data collection and fusion: The camera frame rate can be automatically adjusted according to the ambient light and vehicle speed, with an adjustment range of 15-35 frames / second. When the light is dim or the vehicle is traveling too fast, the frame rate is automatically increased to 35 frames / second; when the light is sufficient and the vehicle speed is stable, the frame rate is reduced to 15 frames / second; the geomagnetic sensor has a self-calibration function, which automatically performs a calibration every 10 minutes. By comparing with the preset standard magnetic field parameters, the sensor sensitivity and measurement deviation are adjusted in real time; the lidar uses multi-echo technology, which can perform multiple measurements on the same target and fuse the data.

[0024] The camera frame rate can be automatically adjusted according to the ambient light and vehicle speed. When the light is dim or the vehicle is traveling too fast, increasing the frame rate to 35 frames per second can capture more details and avoid information loss due to insufficient image acquisition, thereby more accurately identifying the vehicle shape and driving trajectory. When the light is sufficient and the vehicle speed is stable, reducing the frame rate to 15 frames per second can not only meet data needs, but also save resources and reduce data processing volume; the geomagnetic sensor has a self-calibration function, which automatically calibrates every 10 minutes. By comparing the preset standard magnetic field parameters, it adjusts the sensor's sensitivity and measurement deviation in real time, ensuring the accuracy and reliability of the measurement data.

[0025] LiDAR uses multi-echo technology to perform multiple measurements on the same target and fuse the data, greatly improving the accuracy and reliability of the measurement. It can better cope with complex road conditions and target environments, reduce data inaccuracies caused by single measurement errors, and thus provide more accurate and comprehensive information for the measurement of vehicle distance, speed and angle.

[0026] Further, in the feature extraction step based on spatiotemporal features in S02: When the space-time matrix is ​​constructed, the data collection window of the time dimension can be dynamically adjusted according to the traffic flow. During peak traffic hours, the collection window can be shortened to 10 seconds; during off-peak traffic hours, it can be extended to 20 seconds; During the training process of the spatiotemporal convolutional neural network, transfer learning technology is used to first pre-train on a large-scale general traffic dataset, and then fine-tune it for the highway vehicle stop warning scenario using a small amount of scene data; In the pooling layer, an adaptive pooling strategy is used to automatically adjust the pooling window size according to the degree of local change in the feature map for areas with drastic changes.

[0027] When constructing the spatiotemporal matrix, the data collection window in the time dimension can be dynamically adjusted according to the traffic flow. By shortening the collection window to 10 seconds during peak traffic hours, key vehicle driving change information can be captured more timely, because the vehicle density is high and the driving conditions are complex at this time, and a shorter collection window can respond and process real-time data more quickly; and by extending it to 20 seconds during off-peak traffic hours, more comprehensive and stable vehicle driving trend information can be obtained, reducing the frequent updating and processing pressure of data; in the training process of the spatiotemporal convolutional neural network, transfer learning technology is used. Pre-training is first performed on a large-scale general traffic data set, which can use the existing large amount of general data to learn general traffic characteristics and patterns, and then fine-tuning is performed for the highway vehicle parking warning scenario using a small amount of local scene data, which not only saves training time and computing resources, but also enables the model to better adapt to the needs of specific scenarios and improve the accuracy and generalization ability of the model.

[0028] An adaptive pooling strategy is adopted in the pooling layer. For areas with drastic changes, using a smaller pooling window can retain more detail information, thereby more accurately extracting the key features of vehicle driving; while for relatively stable areas, using a larger pooling window can reduce the amount of calculation and data redundancy, thereby improving the efficiency and performance of the model.

[0029] Furthermore, in the step S03 of building a parking risk assessment and prediction model: The weight matrix of the GRU unit is sparsely processed. Through the L1 regularization method, some weight values ​​are set to 0 and the number of model parameters is reduced. By setting some weight values ​​to 0 and reducing the number of model parameters, the complexity and computational complexity of the model can be reduced, so that during the training and prediction process, fewer computing resources are required and the calculation time is shorter, thereby improving the operating efficiency of the model and enabling it to provide parking risk assessment and prediction results more quickly.

[0030] Furthermore, the reinforcement learning algorithm adopts a combination of a deep Q network and a double Q network. In the exploration phase, the deep Q network is used to quickly learn environmental information; in the convergence phase, it is switched to the double Q network. In the exploration phase, the deep Q network can quickly learn environmental information and can widely try different actions, thereby quickly collecting diverse information about the environment; in the convergence phase, it is switched to the double Q network. The double Q network reduces the risk of overestimating the value of actions by decoupling the selection and evaluation of actions, and can more accurately estimate the value of the optimal action, thereby making the learning process more stable and reliable, avoiding falling into a local optimal solution, and improving the quality of the final learned strategy. Furthermore, in the S04, in the multi-channel warning mechanism, personalized warning messages are sent for different types of vehicles. Personalized warning messages can better meet the needs and concerns of drivers of different types of vehicles. Different types of vehicles have different braking performance, reaction time and handling characteristics. Personalized warnings can provide drivers with guidance that is more in line with the actual situation based on these characteristics, thereby increasing the possibility of them taking correct avoidance measures and reducing the risk of accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a logic block diagram of the highway parking risk dynamic warning system based on spatiotemporal characteristics and adaptive thresholds of the present invention. DETAILED DESCRIPTION

[0032] The following is further described in detail through specific implementation methods: Embodiment 1 is basically as attached Figure 1 As shown: The highway parking risk dynamic warning system based on spatiotemporal characteristics and adaptive thresholds includes the following systems Data collection and transmission subsystem: deploy cameras to cover the highway section, and the cameras have auto-focus and low-light enhancement functions; The geomagnetic sensor can detect tiny vehicle vibrations with a response time of less than 0.02 seconds; LiDAR scanning generates point cloud data in real time; The Internet of Vehicles communication module supports 5G high-speed communication; Data processing and analysis subsystem: The data fusion unit performs data fusion processing based on DS evidence theory to eliminate data errors and conflicts; The spatiotemporal feature extraction unit uses the ST-CNN model to extract features from the fused data; the parking risk assessment model unit is based on the GRU-Attention model and combines the attention mechanism to assess and predict vehicle parking risks; Early warning decision and release subsystem: The dynamic threshold adjustment module uses a reinforcement learning algorithm to dynamically adjust the warning threshold according to real-time road conditions and environmental factors; The early warning release module releases early warning information through 5G short messages, highway electronic display screens, and traffic broadcasts.

[0033] Example 2 The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to Example 1 includes the following steps: S01 Data collection and fusion: Use cameras, geomagnetic sensors, and lidar to collect vehicle driving data. The camera captures the vehicle's appearance and driving trajectory characteristics, the geomagnetic sensor monitors the vehicle's passing and speed change characteristics, and the lidar measures the vehicle's distance, speed, and angle in real time; 5G communication is used to obtain vehicle working conditions, brake system status, tire pressure, and temperature data in real time, and DS evidence theory is used for data fusion. The DS evidence theory is used for data fusion, and the formula is: in, is the basic probability distribution function of proposition A after fusion, is the i-th data source for the proposition The basic probability distribution function of , k is a normalization constant; S02 Feature extraction based on spatiotemporal features: Build a spatiotemporal matrix with continuously collected and fused data. The spatial dimensions in the spatiotemporal matrix include the position of the vehicle on the highway, the lane, and the relative position information of surrounding vehicles. Use a spatiotemporal convolutional neural network for feature extraction. The network structure includes a spatiotemporal convolution layer, a pooling layer, and a fully connected layer. Through the spatiotemporal convolution operation, the spatiotemporal characteristics of the vehicle's speed change trend and driving trajectory change are extracted. S03 Parking risk assessment and prediction model construction: Based on the extracted features, a parking risk assessment model is constructed by combining the gated recurrent unit with the attention mechanism to output the risk probability of the vehicle parking in the future; S04 Dynamic threshold adaptive adjustment and warning triggering: According to different sections of the highway, different time periods and weather conditions, the parking risk warning threshold is dynamically adjusted using the reinforcement learning algorithm. The action of adjusting the threshold can be selected according to the current state, and rewards are fed back according to the action. The reward function formula is: in, , , is the weight coefficient, Y is the accuracy, M is the recall rate, and U is the false alarm rate. Different sections mainly refer to the parts of the highway that are different in terms of geographical location, road structure, and distribution of traffic facilities. These differences will lead to different risk conditions of vehicles during driving, which in turn affects the setting of the parking risk warning threshold. There are often many bends and steep slopes on mountain highway sections. When vehicles are driving on such sections, due to the centrifugal force of the bends and the influence of steep slopes, they need to frequently accelerate, decelerate, and shift gears, which increases the risk of vehicle failure or loss of control. For example, continuous curves will obstruct the driver's vision. If the vehicle in front suddenly stops, it will be difficult for the vehicle behind to react in time, and the parking risk will increase significantly. On the highway sections in the plains, the roads are relatively flat and straight, and vehicles travel relatively smoothly, so the parking risk is relatively low. In addition, on sections near service areas and toll stations, the driving status of vehicles changes frequently, such as deceleration, stopping, starting, acceleration, etc., and the traffic flow will also fluctuate greatly. The parking risk of these sections is also different from that of other sections; during the morning and evening rush hours on weekdays, the traffic volume on highways around the city increases dramatically, the vehicle speed is slow, and the distance between vehicles is small. In this case, once a vehicle stops, it is easy to cause a rear-end collision, so the parking risk warning threshold needs to be lowered accordingly in order to issue a warning more timely. In the late night period, the traffic volume on the highway is greatly reduced, but the driver is prone to fatigue at this time, and the probability of the vehicle stopping due to a breakdown is relatively increased. In addition, due to the small traffic volume, rescue may not be timely, so the parking risk warning threshold also needs to be adjusted; in addition, during holidays, the number of traveling vehicles increases, and the traffic volume and driving conditions on the highway are different from weekdays. The warning threshold also needs to be adjusted according to the specific characteristics of the time period; in foggy weather, visibility is extremely low, and it is difficult for the driver to see the vehicle and road conditions in front, and the braking distance of the vehicle will also increase significantly. At this time, even if the actual distance between the vehicles is far, the parking risk is still very high because the driver cannot make an accurate judgment, and the warning threshold should be lowered accordingly. For example, in heavy rain, the road surface will become slippery, the vehicle's handling performance will decrease, and it is easy to slip and lose control, and the parking risk will increase. On sunny days, the driver has a good line of sight, the vehicle is stable, and the parking risk is relatively low. The warning threshold can be appropriately increased.

[0034] When the predicted parking risk probability is greater than the dynamically adjusted threshold, the multi-channel early warning mechanism is triggered to issue early warning information for the target vehicle and target road section.

[0035] The data acquisition and transmission subsystem uses cameras, geomagnetic sensors, lidar, and Internet of Vehicles communication modules to collect vehicle driving data, vehicle operating conditions, brake system status, tire pressure and temperature data from multiple dimensions. Using DS evidence theory, the data is integrated in the data processing and analysis subsystem and converted into a unified and usable form to reduce data errors and conflicts.

[0036] Next, a spatiotemporal convolutional neural network is used to construct a spatiotemporal matrix, and the fused data is subjected to feature extraction to mine the spatiotemporal characteristics of vehicle travel, such as speed change trends and driving trajectory changes. Then, a parking risk assessment model is constructed by combining a gated recurrent unit with an attention mechanism. This model can effectively process time series features, focus on key information through an attention mechanism, and then accurately output the risk probability of vehicle parking in the future. Finally, the warning decision and release subsystem uses a reinforcement learning algorithm to dynamically adjust the parking risk warning threshold based on real-time road conditions and environmental factors such as different sections of the highway, different time periods, and weather conditions. When the predicted parking risk probability is greater than the dynamically adjusted threshold, warning information is issued in a timely manner for target vehicles and target sections through multiple channels such as 5G short messages, highway electronic display screens, and traffic broadcasts, so as to achieve effective warning of parking risks.

[0037] Compared with the existing technology, this solution adopts multi-sensor collaborative collection, covering rich data such as vehicle appearance, driving trajectory, speed changes, vehicle operating conditions, etc. The comprehensiveness of the data far exceeds the single sensor collection method. At the same time, the application of DS evidence theory can better integrate multi-source data, reduce errors and conflicts, and improve data quality. Traditional methods often have problems of information loss or inaccuracy when processing multi-source data.

[0038] In feature extraction and risk assessment model construction, the combination of spatiotemporal convolutional neural network and gated recurrent unit combined with attention mechanism can more accurately capture the spatiotemporal characteristics of vehicle travel and assess parking risks. Compared with simple machine learning algorithms, this solution considers more comprehensive factors, has stronger model generalization ability, can adapt to complex and changeable traffic scenarios, and improves the accuracy of parking risk prediction.

[0039] In the process of adjusting and issuing warning thresholds, the present invention utilizes reinforcement learning algorithms to dynamically adjust thresholds, optimizes warning strategies according to actual road conditions and environmental factors, effectively reduces false alarm and missed alarm rates, and uses multi-channel warning issuance methods to ensure that warning information is promptly conveyed to target vehicles and relevant personnel on the road sections. The traditional fixed threshold warning method lacks flexibility, and the warning issuance channels are single and untimely, making it difficult to meet the needs of actual traffic scenarios.

[0040] In S02, a spatiotemporal convolutional neural network is used for feature extraction. The network structure includes a spatiotemporal convolution layer, a pooling layer, and a fully connected layer. In the spatiotemporal convolution layer, the spatiotemporal convolution kernel slides in the time and space dimensions. The convolution formula is:

[0041] Among them, the It is the Lth layer at position The output, is the spatiotemporal convolution kernel weight of the Lth layer, It is L Layer in position Input, is a bias term. Through the spatiotemporal convolution operation, the spatiotemporal characteristics of vehicle speed change trend and driving trajectory change are extracted. The spatiotemporal convolution kernel slides in the time and space dimensions, and can simultaneously consider the information of vehicles at different times and different spatial positions. The driving status of vehicles on the highway is constantly changing. The speed change trend and driving trajectory change are key dynamic information for judging whether the vehicle has a parking risk. Through this convolution operation, the speed fluctuation during the vehicle driving process can be effectively captured, such as whether the vehicle suddenly decelerates or accelerates, and whether the driving trajectory deviates abnormally; the highway scene is complex, and factors such as vehicle density, driving direction, and weather conditions will affect vehicle driving. The speed change trend and driving trajectory change characteristics extracted by the spatiotemporal convolutional neural network can comprehensively consider these complex factors. For example, under different traffic flow conditions, by analyzing the speed change trend, it can be judged whether the vehicle is driving normally or may stop due to traffic congestion; according to the change of the driving trajectory, it can be judged whether the vehicle is unstable due to bad weather such as heavy rain and fog, thereby increasing the parking risk. This adaptability to complex scenes makes the early warning method more suitable for practical applications.

[0042] In S03, a parking risk assessment model is constructed by combining a gated recurrent unit with an attention mechanism, and the attention mechanism formula is:

[0043]

[0044]

[0045] in, It is a feature With context vector The relevance score of is the attention weight, is the weighted context vector; In the GRU unit, the update gate formula is:

[0046] The reset gate formula is:

[0047] The candidate hidden state formula is:

[0048] The hidden state update formula is:

[0049] in, , are the outputs of the update gate and the reset gate, respectively. is the sigmoid activation function, , , is the weight matrix, , , is the bias vector, is the hidden state at the previous moment, is the input data at the current moment, is a candidate hidden state. The GRU unit processes the time series features and outputs the risk probability of the vehicle parking in the future. The driving state of vehicles on the highway changes over time and has obvious time series characteristics. The GRU unit can handle this kind of time series data well. Through mechanisms such as updating gates, resetting gates and candidate hidden states, it can remember the key information in the vehicle's past driving state and conduct a comprehensive analysis in combination with the current input data. For example, information such as the speed change of the vehicle in the previous period and the coherence of the driving trajectory can be effectively memorized and utilized by the GRU unit, so as to more accurately predict the future parking risk probability; traditional recurrent neural networks are prone to gradient vanishing problems when processing long time series, making it difficult for the model to learn long-distance dependencies. The GRU unit alleviates the gradient vanishing problem to a certain extent compared to traditional RNNs by simplifying the gating mechanism. This enables the model to more effectively propagate gradients and learn more comprehensive time series features when processing long time series of vehicle driving data, thereby more accurately evaluating the parking risk probability.

[0050] In the step of S01 data collection and fusion: The camera frame rate can be automatically adjusted according to the ambient light and vehicle speed, with an adjustment range of 15-35 frames / second. When the light is dim or the vehicle is traveling too fast, the frame rate is automatically increased to 35 frames / second; when the light is sufficient and the vehicle speed is stable, the frame rate is reduced to 15 frames / second; the geomagnetic sensor has a self-calibration function, which automatically performs a calibration every 10 minutes. By comparing with the preset standard magnetic field parameters, the sensor sensitivity and measurement deviation are adjusted in real time; the lidar uses multi-echo technology, which can perform multiple measurements on the same target and fuse the data.

[0051] The camera frame rate can be automatically adjusted according to the ambient light and vehicle speed. When the light is dim or the vehicle is traveling too fast, increasing the frame rate to 35 frames per second can capture more details and avoid information loss due to insufficient image acquisition, thereby more accurately identifying the vehicle shape and driving trajectory. When the light is sufficient and the vehicle speed is stable, reducing the frame rate to 15 frames per second can not only meet data needs, but also save resources and reduce data processing volume; the geomagnetic sensor has a self-calibration function, which automatically calibrates every 10 minutes. By comparing the preset standard magnetic field parameters, it adjusts the sensor's sensitivity and measurement deviation in real time, ensuring the accuracy and reliability of the measurement data.

[0052] LiDAR uses multi-echo technology to perform multiple measurements on the same target and fuse the data, greatly improving the accuracy and reliability of the measurement. It can better cope with complex road conditions and target environments, reduce data inaccuracies caused by single measurement errors, and thus provide more accurate and comprehensive information for the measurement of vehicle distance, speed and angle.

[0053] In the feature extraction step based on spatiotemporal features in S02: When the space-time matrix is ​​constructed, the data collection window of the time dimension can be dynamically adjusted according to the traffic flow. During peak traffic hours, the collection window can be shortened to 10 seconds; during off-peak traffic hours, it can be extended to 20 seconds; During the training process of the spatiotemporal convolutional neural network, transfer learning technology is used to first pre-train on a large-scale general traffic dataset, and then fine-tune it for the highway vehicle stop warning scenario using a small amount of scene data; In the pooling layer, an adaptive pooling strategy is used to automatically adjust the pooling window size according to the degree of local change in the feature map for areas with drastic changes.

[0054] When constructing the spatiotemporal matrix, the data collection window in the time dimension can be dynamically adjusted according to the traffic flow. By shortening the collection window to 10 seconds during peak traffic hours, key vehicle driving change information can be captured more timely, because the vehicle density is high and the driving conditions are complex at this time, and a shorter collection window can respond and process real-time data more quickly; and by extending it to 20 seconds during off-peak traffic hours, more comprehensive and stable vehicle driving trend information can be obtained, reducing the frequent updating and processing pressure of data; in the training process of the spatiotemporal convolutional neural network, transfer learning technology is used. Pre-training is first performed on a large-scale general traffic data set, which can use the existing large amount of general data to learn general traffic characteristics and patterns, and then fine-tuning is performed for the highway vehicle parking warning scenario using a small amount of local scene data, which not only saves training time and computing resources, but also enables the model to better adapt to the needs of specific scenarios and improve the accuracy and generalization ability of the model.

[0055] An adaptive pooling strategy is adopted in the pooling layer. For areas with drastic changes, using a smaller pooling window can retain more detail information, thereby more accurately extracting the key features of vehicle driving; while for relatively stable areas, using a larger pooling window can reduce the amount of calculation and data redundancy, thereby improving the efficiency and performance of the model.

[0056] In the step S03 of building a parking risk assessment and prediction model: The weight matrix of the GRU unit is sparsely processed. Through the L1 regularization method, some weight values ​​are set to 0 and the number of model parameters is reduced. By setting some weight values ​​to 0 and reducing the number of model parameters, the complexity and computational complexity of the model can be reduced, so that during the training and prediction process, fewer computing resources are required and the calculation time is shorter, thereby improving the operating efficiency of the model and enabling it to provide parking risk assessment and prediction results more quickly.

[0057] The reinforcement learning algorithm adopts a combination of a deep Q network and a double Q network. In the exploration phase, the deep Q network is used to quickly learn environmental information; in the convergence phase, it is switched to the double Q network. In the exploration phase, the deep Q network can quickly learn environmental information and can widely try different actions, thereby quickly collecting diverse information about the environment; in the convergence phase, it is switched to the double Q network. The double Q network reduces the risk of overestimating the value of actions by decoupling the selection and evaluation of actions, and can more accurately estimate the value of the optimal action, thereby making the learning process more stable and reliable, avoiding falling into the local optimal solution, and improving the quality of the final learned strategy. In the S04, in the multi-channel warning mechanism, personalized warning messages are sent to different types of vehicles. Personalized warning messages can better meet the needs and concerns of drivers of different types of vehicles. Different types of vehicles have different braking performance, reaction time and handling characteristics. Personalized warnings can provide drivers with guidance that is more in line with the actual situation based on these characteristics, thereby increasing the possibility of them taking correct avoidance measures and reducing the risk of accidents.

[0058] In actual application, 1. System Construction Data acquisition and transmission subsystem Cameras with autofocus and low-light enhancement capabilities are deployed at key sections of highways to ensure that the vehicle shape and driving trajectory can be clearly captured under various lighting conditions.

[0059] Install a geomagnetic sensor that can detect tiny vehicle vibrations and has a response time of less than 0.02 seconds.

[0060] Set up the LiDAR so it can scan and generate point cloud data in real time.

[0061] Equipped with an Internet of Vehicles communication module that supports 5G high-speed communication to ensure real-time data transmission.

[0062] Data processing and analysis subsystem Construct a data fusion unit to fuse data from different sensors and inside the vehicle based on DS evidence theory to eliminate data errors and conflicts.

[0063] A spatiotemporal feature extraction unit is set up, and the ST-CNN model is used to extract features from the fused data.

[0064] A parking risk assessment model unit is established to evaluate and predict vehicle parking risks based on the gated recurrent unit (GRU-Attention model) combined with the attention mechanism.

[0065] Early warning decision and release subsystem Develop a dynamic threshold adjustment module, and use the reinforcement learning algorithm that combines the deep Q network and the dual Q network to dynamically adjust the warning threshold according to real-time road conditions and environmental factors.

[0066] Build an early warning release module and release early warning information through multiple channels such as 5G short messages, highway electronic display screens, and traffic broadcasts.

[0067] Data collection and fusion (S01) Cameras, geomagnetic sensors and lidar collect real-time information about the vehicle's appearance, driving trajectory, speed changes, location, etc. At the same time, 5G communications are used to obtain data such as the vehicle's engine operating conditions, brake system status, tire pressure and temperature, etc.

[0068] When the light is dim or the vehicle is traveling too fast, the camera frame rate automatically increases to 35 frames per second; when the light is sufficient and the vehicle is traveling at a stable speed, the frame rate decreases to 15 frames per second.

[0069] The geomagnetic sensor is automatically calibrated every 10 minutes, and the sensor's sensitivity and measurement deviation are adjusted in real time by comparing with the preset standard magnetic field parameters.

[0070] LiDAR uses multi-echo technology to perform multiple measurements on the same target and fuse the data.

[0071] The DS evidence theory is used to fuse the collected data to obtain comprehensive and accurate data.

[0072] Feature extraction based on spatiotemporal features (S02) The fused data is processed for spatiotemporal features to construct a spatiotemporal matrix. During peak traffic hours, the data collection window in the time dimension is shortened to 10 seconds; during off-peak traffic hours, it is extended to 20 seconds.

[0073] The spatiotemporal convolutional neural network is used for feature extraction. The network structure includes spatiotemporal convolution layer, pooling layer and fully connected layer. The spatiotemporal convolution kernel slides in the time and space dimensions to extract key features such as the speed change trend and driving trajectory change of the vehicle.

[0074] During the training process, the spatiotemporal convolutional neural network is first pre-trained on a large-scale general traffic dataset, and then fine-tuned for the highway vehicle stop warning scenario using a small amount of scene data.

[0075] The pooling layer uses an adaptive pooling strategy to automatically adjust the pooling window size according to the local change degree of the feature map. For areas with drastic changes, a smaller pooling window is used; for relatively stable areas, a larger pooling window is used.

[0076] Parking risk assessment and prediction model construction (S03) Based on the extracted features, a parking risk assessment model is constructed by combining a gated recurrent unit with an attention mechanism.

[0077] The weight matrix of the GRU unit is sparsely processed, and some weight values ​​are changed to 0 through the L1 regularization method to reduce the number of model parameters.

[0078] The model processes time series features and outputs the risk probability of the vehicle stopping within the next 10 seconds.

[0079] Dynamic threshold adaptive adjustment and warning triggering (S04) According to different sections of the highway, different time periods and weather conditions, the parking risk warning threshold is dynamically adjusted using a reinforcement learning algorithm that combines a deep Q network with a dual Q network. In the exploration phase, the deep Q network is used to quickly learn environmental information; in the convergence phase, it is switched to the dual Q network.

[0080] When the predicted parking risk probability is greater than the dynamically adjusted threshold, the multi-channel warning mechanism is triggered. Personalized warning messages are sent to different types of vehicles, including warning locations, expected parking vehicle information, and recommended avoidance routes. At the same time, warning information is displayed in eye-catching signs on the electronic display screen of the highway, and the warning content is broadcast in real time through traffic broadcasting to guide vehicles to take measures in advance. Specific embodiment 3 System construction and deployment On the selected 50-kilometer-long highway section, high-definition cameras with autofocus and low-light enhancement functions are evenly deployed every 1 kilometer to ensure full coverage of the road.

[0082] Geomagnetic sensors are buried every 50 meters along the road section, and their response time is strictly controlled within 0.02 seconds to accurately detect tiny vehicle vibrations.

[0083] High-precision lidars are installed at the starting point, midpoint and end point of the road section, which can perform 360-degree real-time scanning and generate accurate point cloud data.

[0084] Equipped with advanced Internet of Vehicles communication module, it supports 5G high-speed communication and ensures real-time and fast data transmission.

[0085] Data Center Setup Establish a data processing center near the road section and configure multiple high-performance servers: It is used to run the data fusion unit, fuse multi-source data based on DS evidence theory, and eliminate errors and conflicts.

[0086] The spatiotemporal feature extraction unit uses the ST-CNN model to extract features from the fused data.

[0087] The parking risk assessment model unit is based on the GRU-Attention model and combines the attention mechanism to assess and predict vehicle parking risks.

[0088] Installation of early warning decision-making and release facilities Large electronic highway display screens are set up every 5 kilometers on the road section to ensure that drivers can clearly see the warning information from a long distance.

[0089] Establish a stable data interface with the local traffic radio station to realize the broadcast of real-time warning information.

[0090] Method Implementation Data collection and fusion During the day when there is sufficient light and the traffic volume is moderate, the camera frame rate remains at 15 frames per second; in the evening when the light becomes dim and the traffic volume increases, the frame rate automatically increases to 35 frames per second.

[0091] The geomagnetic sensor is automatically calibrated every 10 minutes, adjusting the sensitivity and measurement deviation according to the preset standard magnetic field parameters.

[0092] LiDAR performs multiple echo measurements on the same vehicle and fuses the data to obtain accurate vehicle distance, speed, and angle information.

[0093] Through 5G communication, key data such as vehicle engine operating conditions, braking system status, tire pressure and temperature can be obtained in real time.

[0094] DS evidence theory is used for data fusion to integrate multi-source data.

[0095] Feature extraction based on spatiotemporal features During the morning peak period (7-9 o'clock), the data collection window in the time dimension is shortened to 10 seconds; during the lunchtime off-peak period (12-14 o'clock), it is extended to 20 seconds.

[0096] A spatiotemporal convolutional neural network is used for feature extraction. The network model is first pre-trained on a large-scale general traffic dataset, and then fine-tuned using a small amount of data collected on this road section.

[0097] In the pooling layer, a smaller pooling window is used for areas where the vehicle's trajectory changes dramatically, and a larger pooling window is used for relatively stable areas.

[0098] Parking risk assessment and prediction model construction Based on the extracted features, a gated recurrent unit model combined with an attention mechanism is constructed.

[0099] The L1 regularization method is used to perform sparse processing on the weight matrix of the GRU unit to reduce the number of model parameters.

[0100] The model outputs the risk probability that the vehicle will stop within the next 10 seconds.

[0101] Dynamic threshold adaptive adjustment and warning triggering The warning threshold is dynamically adjusted based on the distribution of curves and straight roads on the road, rush hours on weekdays, holidays, and rain, snow, fog, and clear weather using a reinforcement learning algorithm that combines a deep Q network with a double Q network.

[0102] In the exploration phase, the deep Q network is used to extensively try different threshold adjustment actions to quickly collect environmental information; in the convergence phase, it switches to the dual Q network to more accurately estimate the optimal threshold and avoid overestimating the action value.

[0103] When the predicted parking risk probability exceeds the dynamically adjusted threshold, a multi-channel early warning mechanism is triggered: For large passenger buses, warning messages are sent to highlight their long bodies and large inertia, and it is recommended to plan avoidance routes in advance.

[0104] For small private cars, sending early warning messages emphasizes quick response and flexible avoidance.

[0105] The electronic display screens on highways display eye-catching warning text and icons, including the warning location, the type of vehicle expected to stop and the recommended deceleration range.

[0106] Traffic broadcasts the warning content in real time and in detail, guiding vehicles to take measures in an orderly manner.

[0107] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A dynamic early warning method for highway parking risk based on spatiotemporal characteristics and adaptive thresholds, comprising the following steps: S01 Data collection and fusion: Use cameras, geomagnetic sensors, and lidar to collect vehicle driving data. The camera captures the vehicle's appearance and driving trajectory characteristics, the geomagnetic sensor monitors the vehicle's passing and speed change characteristics, and the lidar measures the vehicle's distance, speed, and angle in real time; 5G communication is used to obtain vehicle working conditions, brake system status, tire pressure, and temperature data in real time, and DS evidence theory is used for data fusion. The DS evidence theory is used for data fusion, and the formula is: in, is the basic probability distribution function of proposition A after fusion, is the i-th data source for the proposition The basic probability distribution function of , k is a normalization constant; S02 Feature extraction based on spatiotemporal features: Build a spatiotemporal matrix with continuously collected and fused data. The spatial dimensions in the spatiotemporal matrix include the position of the vehicle on the highway, the lane, and the relative position information of surrounding vehicles. Use spatiotemporal convolutional neural network to extract features, extract the spatiotemporal characteristics of vehicle speed change trend and driving trajectory change; S03 Parking risk assessment and prediction model construction: Based on the extracted features, a parking risk assessment model is constructed by combining the gated recurrent unit with the attention mechanism to output the risk probability of the vehicle parking in the future; S04 Dynamic threshold adaptive adjustment and warning triggering: According to different sections of the highway, different time periods and weather conditions, the parking risk warning threshold is dynamically adjusted using the reinforcement learning algorithm. The action of adjusting the threshold can be selected according to the current state, and rewards are fed back according to the action. The reward function formula is: in, , , is the weight coefficient, Y is the accuracy, M is the recall rate, and U is the false alarm rate. When the predicted parking risk probability is greater than the dynamically adjusted threshold, the multi-channel early warning mechanism is triggered to issue early warning information for the target vehicle and target road section.

2. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In S02, a spatiotemporal convolutional neural network is used for feature extraction. The network structure includes a spatiotemporal convolution layer, a pooling layer, and a fully connected layer. In the spatiotemporal convolution layer, the spatiotemporal convolution kernel slides in the time and space dimensions. The convolution formula is: Among them, the It is the Lth layer at position The output, is the spatiotemporal convolution kernel weight of the Lth layer, It is L Layer in position Input, It is a bias term, which extracts the spatiotemporal characteristic speed change trend and driving trajectory change of the vehicle through spatiotemporal convolution operation.

3. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In S03, a parking risk assessment model is constructed by combining a gated recurrent unit with an attention mechanism, and the attention mechanism formula is: in, It is a feature With context vector The relevance score of is the attention weight, is the weighted context vector; In the GRU unit, the update gate formula is: The reset gate formula is: The candidate hidden state formula is: The hidden state update formula is: in, , are the outputs of the update gate and the reset gate, respectively. is the sigmoid activation function, , , is the weight matrix, , , is the bias vector, is the hidden state at the previous moment, is the input data at the current moment, It is a candidate hidden state. The time series features are processed by the GRU unit to output the risk probability of the vehicle parking in the future.

4. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In the S01 data collection and fusion step: The camera frame rate can be automatically adjusted according to the ambient light and vehicle speed, with an adjustment range of 15-35 frames / second. When the light is dim or the vehicle is traveling too fast, the frame rate is automatically increased to 35 frames / second; when the light is sufficient and the vehicle speed is stable, the frame rate is reduced to 15 frames / second; the geomagnetic sensor has a self-calibration function, which automatically performs a calibration every 10 minutes. By comparing with the preset standard magnetic field parameters, the sensor sensitivity and measurement deviation are adjusted in real time; the lidar uses multi-echo technology, which can perform multiple measurements on the same target and fuse the data.

5. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In the feature extraction step based on spatiotemporal features in S02: When the space-time matrix is ​​constructed, the data collection window of the time dimension can be dynamically adjusted according to the traffic flow. During peak traffic hours, the collection window can be shortened to 10 seconds; during off-peak traffic hours, it can be extended to 20 seconds; During the training process of the spatiotemporal convolutional neural network, transfer learning technology is used to first pre-train on a large-scale general traffic dataset, and then fine-tune it for the highway vehicle stop warning scenario using a small amount of scene data; In the pooling layer, an adaptive pooling strategy is used to automatically adjust the pooling window size according to the degree of local change in the feature map for areas with drastic changes.

6. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In the step S03 of building a parking risk assessment and prediction model: The weight matrix of the GRU unit is sparsely processed, and some weight values ​​are changed to 0 through the L1 regularization method.

7. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In S04, the reinforcement learning algorithm combines a deep Q network with a dual Q network. In the exploration phase, the deep Q network is used to quickly learn environmental information; in the convergence phase, the algorithm switches to the dual Q network.

8. The method for dynamic early warning of highway parking risk based on spatiotemporal characteristics and adaptive thresholds according to claim 1 is characterized by: In S04, in the multi-channel warning mechanism, personalized warning messages are sent for different types of vehicles.

9. A dynamic early warning system for highway parking risks based on spatiotemporal characteristics and adaptive thresholds, characterized in that: Including the following systems Data collection and transmission subsystem: deploy cameras to cover the highway section, and the cameras have auto-focus and low-light enhancement functions; The geomagnetic sensor can detect tiny vehicle vibrations with a response time of less than 0.02 seconds; LiDAR scanning generates point cloud data in real time; The Internet of Vehicles communication module supports 5G high-speed communication; Data processing and analysis subsystem: The data fusion unit performs data fusion processing based on DS evidence theory to eliminate data errors and conflicts; The spatiotemporal feature extraction unit uses the ST-CNN model to extract features from the fused data; the parking risk assessment model unit is based on the GRU-Attention model and combines the attention mechanism to assess and predict vehicle parking risks; Early warning decision and release subsystem: The dynamic threshold adjustment module uses a reinforcement learning algorithm to dynamically adjust the warning threshold according to real-time road conditions and environmental factors; The early warning release module releases early warning information through 5G short messages, highway electronic display screens, and traffic broadcasts.

Images