A smart early warning method and system for temporary construction work in highway reconstruction and expansion.

Abstract

This invention discloses an intelligent early warning method and system for temporary construction work during highway reconstruction and expansion. The method includes: acquiring point cloud information of vehicles and the environment near the construction area using radar; obtaining vehicle detection targets based on the point cloud information; maintaining continuous time-series trajectories based on the vehicle detection targets to obtain a target state sequence set; constructing vehicle social pooling interaction features based on the target state sequence set to obtain an input feature vector; obtaining trajectory prediction results by combining the input feature vector with a long short-term memory network and a hybrid density network; calculating risk indicators for all time periods based on the trajectory prediction results to obtain vehicle risk factors; classifying risk levels based on the vehicle risk factors to obtain vehicle risk results; generating early warning instructions based on the vehicle risk results; and providing on-site early warnings and vehicle-side human-machine interaction prompts based on the early warning instructions.

Description

Technical Field

[0001] This invention relates to the field of road traffic safety facilities technology, and in particular to an intelligent early warning method and system for temporary construction work during highway reconstruction and expansion. Background Technology

[0002] Highway reconstruction and expansion projects are crucial for improving road network capacity. However, due to complex factors such as lane reduction, abrupt changes in alignment, limited working space, and the interplay of traffic flow and construction activities, these construction zones have become special areas with significantly higher accident risks than ordinary road sections. Among traffic accidents occurring within highway reconstruction and expansion construction zones, those caused by vehicles losing control and entering the construction area account for a prominent proportion. These accidents also result in a higher rate of injuries and fatalities compared to those on ordinary road sections, posing a serious threat to the lives of construction workers and drivers, and impacting project progress and social order.

[0003] To ensure safety in the construction area, existing construction methods typically involve installing a series of static traffic warning signs, markings, and cone-shaped traffic signs upstream of the work area. Under ideal weather conditions, these facilities can provide some guidance and warnings. However, existing technical solutions have the following significant drawbacks:

[0004] (1) Visibility is severely limited, especially in severe weather: The optical performance of widely used reflective signs is highly dependent on the illumination and reflection of external light sources. In low visibility weather conditions such as rain, fog, and haze, water droplets or particles suspended in the air will scatter light significantly, causing the effective visibility distance of reflective signs to drop sharply. Actual measurement data shows that in moderate rain or light fog, the effective visibility distance often drops to less than 50 meters. This distance is far lower than the safe braking distance required for vehicles at common speeds on highways. When drivers suddenly discover a construction zone ahead in severe weather, there is no sufficient distance to take effective braking or evasive action, which can easily lead to rear-end collisions or serious accidents such as running into the construction zone.

[0005] (2) Passive guidance, lack of proactive early warning capability: Existing temporary traffic control facilities (such as signs, cones, water-filled barriers, etc.) are essentially static and passive means of prompting and physical isolation. They can only provide predetermined route information to drivers who follow the rules and are focused, and are completely unable to proactively sense the status of approaching vehicles (speeding, deviating from the predetermined lane) or detect in real time whether personnel, equipment or vehicles have entered the boundary of the construction area. At night, on curves or in areas with limited visibility, drivers are prone to improper operation due to fatigue, distraction or misjudgment of the alignment. If timely and enhanced early warning information is lacking at this time, vehicles are very likely to deviate from the safe path and enter high-risk construction areas.

[0006] (3) Lack of environmental adaptation and dynamic response mechanism: The display mode of existing warning facilities is fixed and cannot be dynamically adjusted and intelligently responded according to ambient lighting conditions, weather conditions or real-time traffic risks. For example, more conspicuous and penetrating light signals are needed in rainy and foggy weather, or strong and targeted avoidance instructions need to be issued immediately when there is a risk of vehicle collision, which cannot be achieved by existing technologies.

[0007] (4) Information silos and delayed emergency response: When serious intrusion incidents (such as vehicles or personnel entering the construction area) occur, existing facilities cannot automatically and quickly report information such as the location and nature of the incident to the monitoring center in real time. The discovery of the accident mainly relies on manual patrols or alarms from passing vehicles. The information transmission chain is long and the positioning is inaccurate, resulting in delayed emergency response, missing the best rescue and traffic control opportunities, and causing the consequences of the accident to expand.

[0008] In summary, existing traffic safety facilities in construction areas, based on static reflective markings and physical barriers, suffer from severely insufficient visibility in low-visibility conditions such as rain and fog. Furthermore, they completely lack the ability to actively detect the risk of approaching high-speed vehicles, to perceive sudden intrusion events in real time, and to adaptively respond dynamically based on environmental and risk levels. These technical deficiencies are one of the key reasons for the high accident rate in highway reconstruction and expansion construction areas, especially the serious injuries and fatalities caused by vehicles crashing into the work area.

[0009] Therefore, a new technical solution is urgently needed to address the technical problem of how to conduct intelligent early warning based on proactive risk detection. Summary of the Invention

[0010] This invention provides an intelligent early warning method and system for temporary construction work in highway reconstruction and expansion, in order to solve the technical problem of how to conduct intelligent early warning based on proactive risk detection.

[0011] To achieve the above objectives, the present invention provides an intelligent early warning method for temporary construction work during highway reconstruction and expansion, characterized in that it includes:

[0012] Point cloud information of vehicles and environment near the construction area is obtained by radar; vehicle detection targets are obtained based on the point cloud information; continuous time series trajectories are maintained based on the vehicle detection targets to obtain a set of target state sequences.

[0013] Based on the target state sequence set, a vehicle social pooling interaction feature is constructed to obtain the input feature vector; the trajectory prediction result is obtained by combining the input feature vector with a long short-term memory network and a hybrid density network.

[0014] Based on the trajectory prediction results, risk indicators are calculated for all time periods to obtain vehicle risk factors; risk levels are classified according to vehicle risk factors to obtain vehicle risk results; early warning instructions are generated based on vehicle risk results; on-site early warnings and vehicle-side human-machine interaction prompts are provided based on the early warning instructions.

[0015] Preferably, point cloud information of vehicles and the environment near the construction area obtained by radar includes:

[0016] The vehicle and environment near the construction area are sensed by millimeter-wave radar to obtain echo signals; the echo signals are mixed and demodulated and sampled quickly, and then processed by three-dimensional fast Fourier transform to obtain an initial point cloud dataset; the initial point cloud dataset is subjected to constant false alarm rate detection to filter out stray points and false alarm points, and retain the valid point cloud with a confidence level higher than the preset value to obtain point cloud information.

[0017] Preferably, the vehicle detection target is obtained based on the point cloud information, and a continuous time-series trajectory is maintained based on the vehicle detection target to obtain a target state sequence set including:

[0018] VoxelNet network is used to encode voxel features of point cloud information to obtain the encoding vector of each voxel; the encoding vector of each voxel is used to model the spatial context in a sparse convolutional network to obtain a set of candidate target boxes; multi-target tracking is performed based on the set of candidate target boxes and an interactive multi-model filter to obtain vehicle detection targets; each detection target is assigned a unique identifier and a continuous time series trajectory is maintained to obtain a set of target state sequences.

[0019] Preferably, the vehicle social pooling interaction features are constructed based on the target state sequence set, and the resulting input feature vector includes:

[0020] After organizing the target state sequence set into standardized spatiotemporal input data, the Z-Score method is used for standardization to obtain a standardized time series sequence.

[0021] Based on the standardized time series, vehicle social pooling interaction features are constructed to obtain the input feature vector, including:

[0022] Treat all vehicles as nodes and their relative positions and speeds as edges, and construct a vehicle interaction graph at each time step; establish weighted connection edges, with the weights determined by the relative position distance and speed difference between the two vehicles; perform pooling operations on the features of all neighboring vehicles of the target to obtain pooled features; and concatenate the standardized state of the target vehicle with the pooled features to obtain the input feature vector.

[0023] Preferably, the trajectory prediction results obtained by combining the input feature vector with a long short-term memory network and a hybrid density network include:

[0024] Based on the input feature vector and a long short-term memory network, a temporal encoding representation of the future trajectory is obtained. Based on the temporal encoding representation of the future trajectory and a hybrid density network, multimodal modeling and probability prediction of the future trajectory distribution are performed to obtain the future trajectory distribution. Based on the future trajectory distribution, road and construction zone constraints are fused and interactive multi-model filters are applied to obtain the trajectory prediction result.

[0025] Preferably, risk indicators are calculated for all time periods based on the trajectory prediction results, resulting in vehicle risk factors including:

[0026] Based on the trajectory prediction results, a preset number of candidate trajectories are generated, and the prediction variance of each prediction point is calculated. Based on the trajectory prediction results, a preset risk index is calculated, and each risk index is normalized and then dynamically weighted in combination with the prediction variance to obtain a single-frame risk factor. The single-frame comprehensive risk factors at all times are aggregated in time series to obtain the vehicle risk factor.

[0027] Preferably, after normalizing each risk indicator, dynamic weighting is performed based on the prediction variance to obtain single-frame risk factors, including:

[0028] Risk indicators include lateral distance, collision time, intrusion speed, trajectory curvature, and longitudinal deceleration. Lateral distance measures the degree of deviation of the vehicle from the boundary of the construction zone. Collision time describes the remaining time of longitudinal collision between the vehicle and the construction zone. Intrusion speed describes the lateral cutting trend. Trajectory curvature describes the degree of vehicle steering abruptly. Longitudinal deceleration reflects whether the driver takes emergency braking measures.

[0029] The collision time and lateral distance metrics were normalized using the inverse mapping function; the intrusion speed, trajectory curvature, and longitudinal deceleration metrics were normalized using the Sigmoid function.

[0030] After obtaining the normalized results of each indicator, the normalized results are dynamically weighted based on the prediction uncertainty and the prediction variance to obtain the single-frame risk factor.

[0031] Preferably, the risk level is classified according to the vehicle risk factors, and the resulting vehicle risk results include:

[0032] Risk levels are classified based on vehicle risk factors, and risk levels are defined using a segmented threshold method. Three risk ranges are obtained based on two preset thresholds. The risk level is determined as low risk, medium risk, or high risk based on the position of the vehicle risk factor within the three risk ranges. The risk level, vehicle risk factors, and the target vehicle identifier are integrated to obtain the vehicle risk result.

[0033] Preferably, a warning command is generated based on the vehicle risk results; on-site warnings and vehicle-side human-machine interaction prompts are provided based on the warning command, including:

[0034] The warning instructions include the risk level, the identifier of the target vehicle, and the preset intervention actions; the warning instructions are distributed to warning devices at all levels through a message distribution mechanism to realize on-site warning and vehicle-side human-machine interaction prompts.

[0035] The present invention also provides an intelligent early warning system for temporary construction of highway reconstruction and expansion, and the system includes a first module, a second module, a third module and a fourth module for the method of the present invention;

[0036] The first module is used to acquire point cloud information of vehicles and the environment near the construction area through radar; obtain vehicle detection targets based on the point cloud information; maintain continuous time series trajectories based on the vehicle detection targets; and obtain a set of target state sequences.

[0037] The second module is used to construct vehicle social pooling interaction features based on the target state sequence set to obtain the input feature vector; and to obtain the trajectory prediction result by combining the input feature vector with a long short-term memory network and a hybrid density network.

[0038] The third module is used to calculate risk indicators for all times based on the trajectory prediction results, and obtain vehicle risk factors; based on the vehicle risk factors, risk levels are classified to obtain vehicle risk results.

[0039] The fourth module is used to generate early warning instructions based on vehicle risk results; and to provide on-site early warnings and vehicle-side human-machine interaction prompts based on the early warning instructions.

[0040] The present invention has the following beneficial effects:

[0041] The method of this invention has significant advantages such as high-precision identification, highly robust prediction, intelligent assessment, and hierarchical early warning, which can effectively improve traffic safety and the level of intelligence in construction sections. Specific beneficial effects include:

[0042] (1) It has achieved accurate perception and continuous tracking of vehicles in the construction area.

[0043] This invention directly acquires high-density point cloud information around the construction area using radar sensors, and utilizes the VoxelNet network to achieve end-to-end 3D detection of vehicle targets. It can accurately identify multiple vehicle targets and their motion states, such as position, speed, and acceleration, in complex construction environments. Subsequently, an Interactive Multi-Model (IMM) filter is introduced, combining various motion models such as uniform velocity (CV), uniform acceleration (CA), and coordinated turning (CT) to perform temporal filtering and data association on the detected targets. This achieves continuous and stable maintenance of multi-target trajectories, effectively suppressing trajectory interruptions caused by noise and frame drops.

[0044] (2) Improved the accuracy of trajectory prediction and the ability to model complex interactive behaviors.

[0045] This invention constructs a Social Pooling interaction feature network to explicitly characterize the spatial-temporal relationships between vehicles, mapping the relative position, speed, and direction information of neighboring vehicles to a high-dimensional interaction space, thus achieving collaborative modeling of vehicle group behavior. Based on this, a Long Short-Term Memory (LSTM) network is used to encode time-series features, capturing long-term dependencies and dynamic evolution patterns of motion; and a Hybrid Density Network (MDN) is used to output a multimodal probability distribution of future trajectories, capable of simultaneously expressing multiple driving intentions (such as deceleration, lane changing, and turning), significantly improving the robustness and uncertainty representation of predictions.

[0046] (3) A multi-dimensional dynamic risk assessment system was constructed to achieve adaptive risk quantification.

[0047] This invention proposes a five-category risk indicator system, including lateral distance, collision time, intrusion speed, trajectory curvature, and longitudinal deceleration. It considers both the spatial relationship between the vehicle and the construction area and integrates kinematic characteristics and driving behavior factors. By normalizing the indicators across different dimensions using reciprocal and sigmoid functions, and introducing a dynamic weighting mechanism based on prediction variance, the system can adaptively adjust the weights according to the confidence level of each indicator, achieving quantitative assessment and dynamic time-series tracking of vehicle risk. This method maintains the stability and reliability of risk assessment even under high prediction uncertainty.

[0048] (4) A hierarchical and interpretable multimodal proactive early warning mechanism has been implemented.

[0049] This invention classifies risks into low, medium, and high levels based on comprehensive risk factors and dynamically adjusts the warning mode in conjunction with optical and acoustic warning subsystems. In the medium-risk stage, the system increases the brightness of the LED array and displays text warnings; in the high-risk stage, the LED array enters a strobe mode and switches to a red background, achieving forced visual intervention. Simultaneously, the system triggers the C-V2X communication module, which generates standardized RSM / CPM warning messages from the roadside unit (RSU) and broadcasts them to surrounding vehicles in real time via the PC5 interface. After the receiving on-board unit (OBU) completes authentication, it combines visual, voice, and other human-machine interaction methods to ensure that the driver perceives the risk in a timely manner and takes evasive action.

[0050] (5) Closed-loop feedback and traceable safety management have been achieved.

[0051] The system generates a log record after each warning is triggered, including vehicle ID, risk factor, risk level, control command, and precise timestamp. Through log data analysis, the backend can dynamically optimize risk criteria, weighting parameters, and thresholds, providing a basis for subsequent safety performance assessments, system tuning, and accident liability tracing, thus forming a complete closed-loop mechanism of "detection—prediction—early warning—recording—optimization".

[0052] The intelligent early warning system for temporary construction of highway reconstruction and expansion projects of the present invention, when used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0053] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0054] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0055] Figure 1 This is a schematic diagram of the method flow of a preferred embodiment of the present invention. Detailed Implementation

[0056] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0057] See Figure 1 In a preferred embodiment of the present invention, an intelligent early warning method for temporary construction work during highway reconstruction and expansion is provided, characterized by comprising:

[0058] S1. Obtain point cloud information of vehicles and environment near the construction area through radar; obtain vehicle detection targets based on point cloud information; maintain continuous time series trajectories based on vehicle detection targets to obtain a set of target state sequences.

[0059] In a preferred embodiment of the present invention, acquiring point cloud information of vehicles and the environment near the construction area via radar includes:

[0060] Millimeter-wave radar is used to sense vehicles and the environment near the construction area, obtaining echo signals. These echo signals are then mixed, demodulated, and quickly sampled before being processed by a 3D Fast Fourier Transform to obtain an initial point cloud dataset. To reduce noise, the initial point cloud dataset undergoes constant false alarm rate (CFAR) detection to filter out stray and false alarm points, retaining valid point clouds with confidence levels higher than a preset value. This yields distance-velocity-angle point cloud information, providing reliable input for subsequent depth feature extraction. In a preferred embodiment of this invention, the point cloud information... Represented as:

[0061] ;

[0062] in, Indicates the spatial position of the target scattering point in the radar coordinate system; Indicates echo intensity; Represents a point cloud information vector; Indicates the point cloud number; This indicates the number of point clouds collected.

[0063] In a preferred embodiment of the present invention, the vehicle detection target is obtained based on point cloud information, and a continuous time-series trajectory is maintained based on the vehicle detection target to obtain a target state sequence set, including:

[0064] VoxelNet is used to encode voxelized features of point cloud information, resulting in encoded vectors for each voxel. These encoded vectors are then used in a sparse convolutional network to model the spatial context, yielding a set of candidate bounding boxes, specifically including:

[0065] After obtaining the cleaned point cloud data, a VoxelNet network is used for voxelized feature encoding. Specifically, the point cloud is divided into a regular three-dimensional voxel grid, and the set of points within each voxel is denoted as . , This represents the number of point clouds within a voxel. In the voxel feature encoder, the point cloud is first centered, then nonlinearly mapped using a multilayer perceptron, and finally the point features within the voxel are averaged and aggregated to obtain the encoding vector for that voxel. The calculation formula is as follows:

[0066] ;

[0067] in, Represents the geometric center of the point cloud within a voxel. It is a nonlinear transformation function.

[0068] Subsequently, all voxel features are used for spatial context modeling in a sparse convolutional network, outputting a 3D feature map. An initial set of candidate bounding boxes is then generated via the detection head. :

[0069] ;

[0070] Among them, location ,size and heading angle Describe the geometric properties of the target being detected; Indicates the confidence level of the detection; This represents the state vector of the detected target.

[0071] By setting a threshold Filter out the set of candidate target boxes that meet the conditions. :

[0072] ;

[0073] Multi-target tracking is performed based on a set of candidate target boxes combined with an interactive multi-model filter to obtain vehicle detection targets. Each detection target is assigned a unique identifier, and a continuous time-series trajectory is maintained to obtain a set of target state sequences, specifically including:

[0074] To ensure the continuity and robustness of the detected targets over time, an interactive multi-model filter is introduced for multi-target tracking. In this method, it is assumed that the target motion may follow three basic models: uniform velocity (CV), uniform acceleration (CA), and coordinated turning (CT). For each model, the state transition relation can be expressed as:

[0075] For the A model, whose state vector is defined as:

[0076] ;

[0077] in For the vehicle's position coordinates, For velocity components, For acceleration components, For heading angle, This represents the matrix transpose. The state transition equation for each model is:

[0078] ;

[0079] in Here is the state transition matrix. Here is the process noise covariance matrix, which reflects the vehicle's noise in the first and second stages. Dynamics and system uncertainties under the model; This represents the vehicle state vector in the (k-1)th frame; This represents the motion deviation of the target vehicle in the (k-1)th frame; Let represent the process noise covariance matrix. The observation equation is defined as:

[0080] ;

[0081] in, These are the observations (position and velocity) obtained from radar detection. For the observation matrix, To observe the noise covariance matrix, This represents observation noise. In the IMM framework, at each time step, the system runs multiple Kalman filters in parallel, calculating the prediction and update results of the target state based on different motion assumptions. These results are then fused using model probability weights to obtain the final state estimate. The model weight update is calculated using Bayesian principles.

[0082] ;

[0083] in, Indicates the first The model at time The probability, This represents the model transition probability, i.e., the probability of the system transitioning from the model. Convert to model The possibility of satisfying , This represents the total number of motion models. Representing the model at time step The posterior probability, The system represents the model The probability of transitioning to model n, Let represent the state estimation vector of model n. Through this update formula, the system can dynamically adjust the weights of each model based on real-time observation data, achieving adaptive modeling for different driving behavior patterns. The final comprehensive state estimate, i.e., the vehicle detection target, can be expressed as:

[0084] ;

[0085] in For the model The filtering results are obtained. This weighted fusion allows the system to automatically favor the uniform velocity model during steady motion, while automatically increasing the weight of the CA or CT model during maneuvers such as acceleration, turning, or lane changing, thereby achieving a smooth and continuous target trajectory output.

[0086] After estimation using the IMM filter, each detected vehicle is assigned a unique identifier (ID) and a continuous time-series trajectory is maintained. The set of target state sequences can be represented as:

[0087] ;

[0088] in For the goal The set of target state sequences For a moment The state vector; The horizontal coordinate of the vehicle's position, The vertical coordinate of the vehicle's position. For the vehicle's lateral speed, For the longitudinal speed of the vehicle, For the lateral acceleration of the vehicle, For longitudinal acceleration, This refers to the vehicle's heading angle. In this way, the system achieves end-to-end processing from point cloud signals to a stable target trajectory, providing structured, high-precision input for subsequent trajectory prediction and risk quantification.

[0089] S2. Construct vehicle social pooling interaction features based on the target state sequence set to obtain the input feature vector; combine the input feature vector with a long short-term memory network and a hybrid density network to obtain the trajectory prediction result.

[0090] In a preferred embodiment of the present invention, the vehicle social pooling interaction features are constructed based on the target state sequence set, and the resulting input feature vector includes:

[0091] After organizing the target state sequence set into standardized spatiotemporal input data, the Z-Score method is used for standardization to obtain a standardized time series sequence.

[0092] In the initial stage of trajectory prediction, the target state sequence output from step S1 needs to be organized into standardized spatiotemporal input data. The Z-Score method is used to standardize the input data. For each target vehicle, the maintenance length is... Historical state cache standardization time sequence The single-frame state vector is defined as follows: . for The horizontal coordinate of the vehicle's position at any given time. for The vertical coordinate of the vehicle's position at any given time. for The vehicle's lateral speed at all times for The longitudinal speed of the vehicle at any given time, for The vehicle's lateral acceleration at all times for longitudinal acceleration at all times for Vehicle heading angle at any given time.

[0093] Based on the standardized time series, vehicle social pooling interaction features are constructed to obtain the input feature vector, including:

[0094] Treating all vehicles as nodes and their relative positions and speeds as edges, a vehicle interaction graph is constructed at each time step. Weighted connections are established, with weights determined by the relative distance and speed difference between two vehicles. Pooling is performed on the features of all neighboring vehicles of the target vehicle to obtain pooled features. The standardized state of the target vehicle itself is then concatenated with the pooled features to obtain the input feature vector. Specifically, this includes:

[0095] All vehicles detected by the system are considered as nodes in a graph structure, and the node features are the motion state vectors of each vehicle at the current moment. ,in Indicates the vehicle's position coordinates. For any two nodes, the velocity component is... and Establish weighted connection edges, with their weights The relative positional distance and speed difference between the two vehicles are jointly determined, and are defined as follows:

[0096] ;

[0097] in As the normalization factor, then, for the target vehicle Define its neighbor set as ,in This is the interaction threshold, used to filter the main interaction objects. For each neighboring car... Calculate its relative eigenvectors , For vehicles With vehicles Horizontal distance; For vehicles With vehicles Longitudinal distance; For vehicles With vehicles Speed ​​difference. These relative quantities characterize the dynamic relationship between neighboring vehicles and the target vehicle, and are important inputs for modeling interactions. The neighboring vehicle's state... Relative Relationship spliced ​​as Then Input mapping function Extract high-dimensional interactive feature vectors. The mapping function can take the form of a multilayer perceptron (MLP):

[0098] ;

[0099] The mapping results capture the potential motion intentions of neighboring vehicles and their impact on the target vehicle's behavior through nonlinear transformation. To aggregate information from multiple neighboring vehicles, the target vehicle... Pooling operation is performed on all neighboring vehicle features:

[0100] ;

[0101] Ultimately, the target vehicle's own standardized state is determined. With pooling characteristics The features are concatenated to form a new input feature vector. :

[0102] ;

[0103] This input feature vector preserves the target vehicle's self-kinematic information while explicitly incorporating the influence of surrounding vehicles on its motion. Through this social pooling feature construction method, the system can model the implicit interaction relationships between multiple vehicles, enabling the prediction model to better understand the driving situation and thus improve the prediction accuracy for complex traffic behaviors.

[0104] In a preferred embodiment of the present invention, obtaining the trajectory prediction result based on the input feature vector combined with a long short-term memory network and a hybrid density network includes:

[0105] Based on the input feature vector and a long short-term memory network, a temporal encoding representation of the future trajectory is obtained, specifically including:

[0106] The input feature vector sequence obtained at each time step above The input is fed into a Long Short-Term Memory (LSTM) network, which utilizes its gating mechanism to capture long-term dependencies and temporal evolution patterns. At each time step, the LSTM network updates the hidden state based on the input and the previous hidden state, using the following update formula:

[0107] ;

[0108] in Let be the hidden state vector at time t, which includes forget gate, input gate, and output gate operations. By stacking multiple layers or introducing Dropout, the model's expressiveness and generalization ability can be enhanced. Finally, the hidden state sequence... As a temporal encoding representation of future trajectories, it provides high-dimensional semantic features for the probability prediction module.

[0109] Based on the temporal encoding representation of future trajectories and combined with a hybrid density network, multimodal modeling and probabilistic prediction of future trajectory distribution are performed to obtain the future trajectory distribution, specifically including:

[0110] After encoding the temporal features, the system needs to probabilistically model the future spatiotemporal trajectory of the target vehicle. Because vehicle driving behavior near construction zones exhibits significant uncertainty and diversity (such as decelerating to go straight, changing lanes to avoid obstacles, and accelerating to overtake), a single deterministic prediction model cannot simultaneously cover multiple driving modes. Therefore, this invention introduces a Mixture Density Network (MDN) structure based on the temporal feature vector output by the LSTM to perform multimodal modeling and probabilistic prediction of the future trajectory distribution.

[0111] Specifically, LSTM at time Output hidden state vector This vector contains time-dependent information and contextual dynamic features of the vehicle's historical trajectory. MDN will... As input, it is mapped to the future through several fully connected layers. The set of Gaussian mixture distribution parameters for the first step, including the first step The Gaussian components at time... Mixed weights Mean vector and covariance matrix .in Indicates the component number of the mixture. The number of Gaussian components is preset. Therefore, the future... Frame prediction points The conditional probability distribution is defined as:

[0112] ;

[0113] During the prediction phase, the model outputs a set of mixed distribution parameters for each step. This allows us to obtain a multimodal probability description of the future trajectory. At this point, we can obtain the future trajectory by taking the expected value of this distribution. The average trajectory points of the frame, i.e., the future trajectory distribution:

[0114] ;

[0115] This expected value represents the vehicle's performance in the future. The most likely location of a frame is the central trend of the multimodal distribution.

[0116] Based on the future trajectory distribution, road and construction zone constraints are fused and interactive multi-model filters are applied for correction to obtain trajectory prediction results, specifically including:

[0117] To ensure the physical plausibility of the prediction results, this invention applies road and construction zone constraints to the trajectory distribution output by the MDN. If the predicted mean point is not in a passable area... Then, by constraining the loss Penalize network parameters:

[0118] ;

[0119] in, This represents the road constraint penalty coefficient.

[0120] Simultaneously, the constrained deep learning (MDN) trajectory distribution is fused with the physical model (IMM) prediction results (i.e., the vehicle detection target). The IMM filter in S1 operates in parallel based on three motion models: uniform velocity (CV), uniform acceleration (CA), and coordinated turning (CT), outputting a weighted estimate of the future state. Meanwhile, the deep learning model calculates the desired location based on the output distribution of the hybrid density network. .in This is the fusion coefficient, used to adjust the relative contributions of the deep model and the physical model. A larger value can be taken when the deep model has high confidence in structured scenes (e.g., ...). When environmental noise is strong or data is missing, reduce the speed appropriately. To maintain trajectory continuity and stability, the physical model constraint ratio is increased. The two are then fused using a weighted formula to obtain the trajectory prediction result.

[0121] ;

[0122] in, The fusion coefficient; This is the prediction result of the physical model given by IMM. This design balances the expressive power of deep learning models with the stability of traditional models.

[0123] S3. Calculate the risk indicators for all times based on the trajectory prediction results to obtain the vehicle risk factor; classify the risk level based on the vehicle risk factor to obtain the vehicle risk result.

[0124] In a preferred embodiment of the present invention, risk indicators are calculated for all time periods based on the trajectory prediction results to obtain vehicle risk factors, including:

[0125] Based on the trajectory prediction results, a preset number of candidate trajectories are generated, and the prediction variance of each prediction point is calculated, specifically including:

[0126] During the inference phase, multiple samples can be taken from the MDN distribution (trajectory prediction results) to generate several candidate trajectories. And according to the mixed weights Calculate the confidence scores for different trajectories. .

[0127] ;

[0128] Simultaneously, the expected value and variance are calculated for each prediction point, with the expected value being:

[0129] ;

[0130] in, For the first Gaussian component mixing weights; For the first The mean vector of Gaussian components.

[0131] The variance is:

[0132] ;

[0133] in, For the first The covariance matrix of Gaussian components.

[0134] This uncertainty information (variance) will be used in subsequent steps for confidence-weighted risk metrics.

[0135] Based on the trajectory prediction results, preset risk indicators are calculated. These indicators are then normalized and dynamically weighted using the prediction variance to obtain a single-frame risk factor. The combined single-frame risk factors from all time points are then time-series aggregated to obtain the vehicle risk factor.

[0136] In a preferred embodiment of the present invention, after normalizing each risk indicator, dynamic weighting is performed based on the prediction variance to obtain a single-frame risk factor, including:

[0137] Risk indicators include lateral distance, collision time, intrusion speed, trajectory curvature, and longitudinal deceleration. Lateral distance measures the degree of deviation of the vehicle from the boundary of the construction zone. Collision time describes the remaining time of longitudinal collision between the vehicle and the construction zone. Intrusion speed describes the lateral cutting trend. Trajectory curvature describes the degree of vehicle steering abruptly. Longitudinal deceleration reflects whether the driver takes emergency braking measures.

[0138] A lateral distance index is defined to measure the degree of deviation of a vehicle from the boundary of the construction area. Its calculation formula is as follows:

[0139] ;

[0140] in, Vehicles in the future The horizontal position of the frame, This indicates the boundary location of the construction zone. When this value approaches zero, it indicates that the vehicle is closer to the construction zone, posing a higher risk of intrusion. Next, the Time-of-Collision (TTC) index is defined, which describes the remaining time between a longitudinal collision between the vehicle and the construction zone. The formula is:

[0141] ;

[0142] in Indicates the longitudinal distance from the leading edge of the construction area. The longitudinal velocity component is represented by TTC, with a smaller TTC indicating a more pressing risk. To further characterize the lateral intrusion trend, an intrusion velocity index is introduced. This represents the lateral velocity component; the larger the value, the more likely the vehicle is to enter the construction area laterally.

[0143] ;

[0144] At the same time, the trajectory curvature index, which reflects the degree of vehicle steering aggression, has also been included in the risk indicators. Its calculation formula is as follows:

[0145] ;

[0146] in The first derivative, This is the second derivative. Excessive curvature indicates that the vehicle may be making a sharp turn, increasing the probability of a collision. Finally, the longitudinal deceleration index is defined as:

[0147] ;

[0148] in, This is the longitudinal deceleration, used to reflect whether the driver is likely to take emergency braking measures; a larger value often indicates that a dangerous situation is occurring.

[0149] The collision time and lateral distance metrics were normalized using the inverse mapping function; the intrusion speed, trajectory curvature, and longitudinal deceleration metrics were normalized using the Sigmoid function.

[0150] Because the dimensions and value ranges of various risk indicators differ, they need to be normalized for easier integration. For indicators such as collision time and lateral distance, where smaller values ​​indicate higher risk, a reciprocal mapping function is used for normalization, defined as:

[0151] ;

[0152] in, To adjust parameters and balance sensitivity across different numerical ranges, the Sigmoid function is used for mapping indices such as invasion velocity, curvature, and deceleration. The expression is:

[0153] ;

[0154] in, This is the scaling parameter. In this way, all risk indicators are mapped to a range. Furthermore, the closer the value is to 1, the higher the level of risk.

[0155] After obtaining the normalized results of each indicator, the normalized results are dynamically weighted based on the prediction uncertainty and the prediction variance to obtain the single-frame risk factor.

[0156] This invention proposes a dynamic weighting mechanism based on prediction uncertainty, that is, for each index at a future time... The comprehensive calculation formula is as follows:

[0157] ;

[0158] Among them, weight It is inversely proportional to the uncertainty of the indicator's prediction, specifically defined as:

[0159] ;

[0160] in, For the first The risk indicator in the first Frame prediction variance; For the first The risk indicator in the first The prediction variance of a frame. When the prediction of a certain indicator is more reliable, its contribution to the overall risk also increases accordingly, thereby achieving adaptive dynamic adjustment.

[0161] By time-series aggregation of the single-frame comprehensive risk factors at all times, the vehicle risk factor is obtained:

[0162] A single-frame risk factor reflects the degree of danger at a specific moment, but to obtain a comprehensive assessment of the entire future window, the results from all moments need to be aggregated. This method adopts the worst-case principle, taking the maximum risk value among the next N steps as the final result. To ensure the system can respond more sensitively to early risks, a time decay factor is also introduced. This gives higher weight to risk values ​​at earlier stages. The vehicle risk factor is calculated as follows:

[0163] ;

[0164] This method ensures that the system prioritizes responses to the most dangerous moments in multi-step prediction and provides reasonable time sensitivity.

[0165] In a preferred embodiment of the present invention, the risk level is classified according to vehicle risk factors to obtain vehicle risk results, including:

[0166] Risk levels are classified based on vehicle risk factors, and risk levels are defined using a segmented threshold method. Three risk ranges are obtained based on two preset thresholds. The risk level is determined as low risk, medium risk, or high risk based on the position of the vehicle risk factor within the three risk ranges. The risk level, vehicle risk factors, and the target vehicle identifier are integrated to obtain the vehicle risk result.

[0167] Once the vehicle risk factors are obtained, they need to be mapped to specific risk levels to drive subsequent early warning execution. Risk levels are defined using a segmented threshold method. If the following conditions are met, the vehicle is deemed low-risk; if the following conditions are met... If it is, it is judged as medium risk; if If so, it is judged as high risk. and The calibration can be performed based on experimental statistics and actual needs in the construction area, for example, taking values ​​of 0.4 and 0.85 respectively.

[0168] The final output is Where ID is the identifier of the target vehicle. For vehicle risk factors, The risk level is indicated. This result not only serves as input for subsequent warning execution steps but is also archived for future system optimization and accountability.

[0169] S4. Generate a warning command based on the vehicle risk result; provide on-site warnings and vehicle-side human-machine interaction prompts based on the warning command. In a preferred embodiment of the present invention, S4 specifically includes:

[0170] The warning instructions include the risk level, the identifier of the target vehicle, and the preset intervention actions; the warning instructions are distributed to warning devices at all levels through a message distribution mechanism to realize on-site warning and vehicle-side human-machine interaction prompts.

[0171] In a preferred embodiment of the present invention, a corresponding warning instruction is first generated based on the vehicle risk result. For the target vehicle... The output is a triplet. ,in It serves as a unique identifier for the vehicle. As a risk factor, The risk level is defined as follows. Based on this, the generated instruction is represented as... ,in The system records the intervention actions to be taken, such as enhancing optical cues or broadcasting emergency V2X messages. In this way, the numerical risk outcome is transformed into discretized execution commands, ensuring the operability and versatility of the system in subsequent modules.

[0172] The generated instructions are distributed via a message middleware, employing a publish-subscribe pattern to ensure decoupling and real-time performance between different modules. To this end, the system defines different topic channels based on risk level; if the risk level is low, the corresponding topic is... If it is medium risk, then the corresponding If it is high risk, then the corresponding The dispatch function is defined as follows:

[0173] ;

[0174] Different execution subsystems subscribe to different topics according to their needs. The optical alert unit subscribes to medium- and high-risk topics, the V2X communication module subscribes only to high-risk topics, and the background log system subscribes to all topics. This mechanism ensures message delivery reliability while prioritizing messages, allowing high-risk commands to receive the shortest latency transmission.

[0175] When the message is sent to the optical warning subsystem, the system dynamically adjusts the operating mode of the LED array according to the risk level. In medium-risk situations, the brightness of the LED display is increased from the baseline value, and warning text is displayed to guide vehicles to slow down and avoid the area. In high-risk situations, the LED array enters strobe mode, the background flashes red, and the brightness is at its maximum to achieve a forced warning effect. The brightness adjustment formula is:

[0176] ;

[0177] in, As the reference brightness, and As an indicator variable, it takes the value 1 when the risk level is medium or high, and 0 otherwise. For medium-risk brightness enhancement coefficient, This is a high-risk brightness enhancement factor. Through this graded control, the system can ensure driving comfort while ensuring that dangerous situations quickly attract the driver's attention.

[0178] In high-risk scenarios, the system further triggers the V2X communication module, which generates a standardized warning message from the roadside unit. This message includes the target vehicle's ID, current location, speed, distance to the construction zone, and suggested driving actions, and is encapsulated as a data packet in RSM or CPM format. The message can be formalized as follows:

[0179] ;

[0180] The Type is fixed as Roadwork Hazard, used to identify the message type. As a unique identifier for the target vehicle, The coordinates of the current position. The current velocity vector, This represents the shortest distance between the vehicle and the construction area. To suggest driving behavior, the generated message is broadcast to all vehicles within the coverage area via the C-V2X PC5 interface in the direct connection band. Upon receiving the message, the onboard unit of nearby vehicles verifies the identity and integrity in the safety module and then transmits the message to the vehicle's human-machine interface system, ensuring that information is delivered to the driver quickly, reliably, and securely.

[0181] When the vehicle system receives a high-risk warning message, it will alert the driver in a multimodal manner. First, a high-risk warning icon and a directional arrow indicating a construction zone will be displayed visually on the instrument panel or central control screen. Simultaneously, the system will trigger a voice prompt, advising the driver to immediately slow down or change lanes. If the vehicle has a driver assistance system, tactile feedback such as steering wheel vibration can further enhance the warning. The combination of these three alert modes is also discussed. It can be described as:

[0182] ;

[0183] in The weighting coefficient is dynamically adjusted and its value ranges from 0 to 1. For visual cue signal components, For auditory cue signal components, This refers to tactile cue signal components. The weights differ depending on the risk level; for example, in high-risk situations, three cue methods can be activated simultaneously. Through multimodal human-computer interaction, the probability of drivers ignoring risks can be reduced, improving the effectiveness and enforceability of warnings.

[0184] All hierarchical instructions and their execution processes will be recorded in the backend system for subsequent analysis and accountability. The log format can be represented as follows:

[0185] ;

[0186] in, To consider comprehensive risk factors, Risk level, To execute the instruction, The timestamps are accurate to milliseconds to ensure the uniqueness of events. The backend not only stores risk factors and instruction information but also records the execution status of the optical unit and V2X broadcasts, facilitating the evaluation of the overall system performance. This closed-loop feedback mechanism allows for the optimization of risk criteria and threshold selection in later stages, while also providing data for accident investigation and liability determination.

[0187] The method of this invention has significant advantages such as high-precision identification, highly robust prediction, intelligent assessment, and hierarchical early warning, which can effectively improve traffic safety and the level of intelligence in construction sections. Specific beneficial effects include:

[0188] (1) It has achieved accurate perception and continuous tracking of vehicles in the construction area.

[0189] This invention directly acquires high-density point cloud information around the construction area using radar sensors, and utilizes the VoxelNet network to achieve end-to-end 3D detection of vehicle targets. It can accurately identify multiple vehicle targets and their motion states, such as position, speed, and acceleration, in complex construction environments. Subsequently, an Interactive Multi-Model (IMM) filter is introduced, combining various motion models such as uniform velocity (CV), uniform acceleration (CA), and coordinated turning (CT) to perform temporal filtering and data association on the detected targets. This achieves continuous and stable maintenance of multi-target trajectories, effectively suppressing trajectory interruptions caused by noise and frame drops.

[0190] (2) Improved the accuracy of trajectory prediction and the ability to model complex interactive behaviors.

[0191] This invention constructs a Social Pooling interaction feature network to explicitly characterize the spatial-temporal relationships between vehicles, mapping the relative position, speed, and direction information of neighboring vehicles to a high-dimensional interaction space, thus achieving collaborative modeling of vehicle group behavior. Based on this, a Long Short-Term Memory (LSTM) network is used to encode time-series features, capturing long-term dependencies and dynamic evolution patterns of motion; and a Hybrid Density Network (MDN) is used to output a multimodal probability distribution of future trajectories, capable of simultaneously expressing multiple driving intentions (such as deceleration, lane changing, and turning), significantly improving the robustness and uncertainty representation of predictions.

[0192] (3) A multi-dimensional dynamic risk assessment system was constructed to achieve adaptive risk quantification.

[0193] This invention proposes a five-category risk indicator system, including lateral distance, collision time, intrusion speed, trajectory curvature, and longitudinal deceleration. It considers both the spatial relationship between the vehicle and the construction area and integrates kinematic characteristics and driving behavior factors. By normalizing the indicators across different dimensions using reciprocal and sigmoid functions, and introducing a dynamic weighting mechanism based on prediction variance, the system can adaptively adjust the weights according to the confidence level of each indicator, achieving quantitative assessment and dynamic time-series tracking of vehicle risk. This method maintains the stability and reliability of risk assessment even under high prediction uncertainty.

[0194] (4) A hierarchical and interpretable multimodal proactive early warning mechanism has been implemented.

[0195] This invention classifies risks into low, medium, and high levels based on comprehensive risk factors and dynamically adjusts the warning mode in conjunction with optical and acoustic warning subsystems. In the medium-risk stage, the system increases the brightness of the LED array and displays text warnings; in the high-risk stage, the LED array enters a strobe mode and switches to a red background, achieving forced visual intervention. Simultaneously, the system triggers the C-V2X communication module, whereby the roadside unit generates standardized RSM / CPM warning messages, which are broadcast to surrounding vehicles in real time via the PC5 interface. After the receiving vehicle unit completes authentication, it combines visual, voice, and other human-machine interaction prompts to ensure that the driver promptly perceives the risk and takes evasive action.

[0196] (5) Closed-loop feedback and traceable safety management have been achieved.

[0197] The system generates a log record after each warning is triggered, including vehicle ID, risk factor, risk level, control command, and precise timestamp. Through log data analysis, the backend can dynamically optimize risk criteria, weighting parameters, and thresholds, providing a basis for subsequent safety performance assessments, system tuning, and accident liability tracing, thus forming a complete closed-loop mechanism of "detection—prediction—early warning—recording—optimization".

[0198] In a preferred embodiment of the present invention, an intelligent early warning system for temporary construction of highway reconstruction and expansion is also provided, which is used in the method of the present invention. The system includes a first module, a second module, a third module and a fourth module.

[0199] The first module is used to acquire point cloud information of vehicles and the environment near the construction area through radar; based on the point cloud information, vehicle detection targets are obtained; based on the vehicle detection targets, continuous time series trajectories are maintained to obtain a set of target state sequences.

[0200] The second module is used to construct vehicle social pooling interaction features based on the target state sequence set to obtain the input feature vector; and to obtain the trajectory prediction result by combining the input feature vector with the long short-term memory network and the hybrid density network.

[0201] The third module is used to calculate risk indicators for all time periods based on trajectory prediction results, thereby obtaining vehicle risk factors; and to classify risk levels based on vehicle risk factors, thereby obtaining vehicle risk results.

[0202] The fourth module is used to generate early warning instructions based on vehicle risk results; and to provide on-site early warnings and vehicle-side human-machine interaction prompts based on the early warning instructions.

[0203] The intelligent early warning system for temporary construction of highway reconstruction and expansion projects of the present invention, when used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0204] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent early warning method for temporary construction work during highway reconstruction and expansion, characterized in that, include: Point cloud information of vehicles and environment near the construction area is acquired by radar; vehicle detection targets are obtained based on the point cloud information; continuous time series trajectories are maintained based on the vehicle detection targets to obtain a set of target state sequences. Based on the target state sequence set, a vehicle social pooling interaction feature is constructed to obtain an input feature vector; based on the input feature vector, a long short-term memory network and a hybrid density network are combined to obtain the trajectory prediction result; Based on the trajectory prediction results, risk indicators are calculated for all times to obtain the vehicle risk factor; The risk level is classified according to the vehicle risk factors to obtain the vehicle risk result; A warning instruction is generated based on the vehicle risk results; Based on the aforementioned warning instructions, on-site warnings and vehicle-side human-machine interaction prompts will be provided. Based on the trajectory prediction results, risk indicators are calculated for all time periods to obtain vehicle risk factors, including: Based on the trajectory prediction results, a preset number of candidate trajectories are generated, and the prediction variance of each prediction point is calculated. Based on the trajectory prediction results, a preset risk index is calculated, and each risk index is normalized and then dynamically weighted in combination with the prediction variance to obtain a single-frame risk factor. The single-frame comprehensive risk factors at all times are time-series aggregated to obtain the vehicle risk factor. The process of normalizing each risk indicator and then dynamically weighting it based on the prediction variance to obtain a single-frame risk factor includes: The risk indicators include lateral distance, collision time, intrusion speed, trajectory curvature, and longitudinal deceleration. The lateral distance indicator measures the degree of deviation of the vehicle from the boundary of the construction area. The collision time indicator describes the remaining time of longitudinal collision between the vehicle and the construction area. The intrusion speed indicator describes the lateral cutting trend. The trajectory curvature indicator describes the degree of vehicle steering abruptly. The longitudinal deceleration indicator reflects whether the driver takes emergency braking measures. The collision time index and the lateral distance index are normalized using a reciprocal mapping function; the intrusion speed index, trajectory curvature index, and longitudinal deceleration index are normalized using a Sigmoid function. After obtaining the normalized results of each indicator, the normalized results are dynamically weighted based on the prediction uncertainty and the prediction variance to obtain the single-frame risk factor.

2. The intelligent early warning method for temporary construction work in highway reconstruction and expansion according to claim 1, characterized in that, The point cloud information of vehicles and the environment near the construction area obtained by radar includes: The vehicle and environment near the construction area are sensed by millimeter-wave radar to obtain echo signals; the echo signals are mixed and demodulated and then sampled quickly, and then processed by three-dimensional fast Fourier transform to obtain an initial point cloud dataset; the initial point cloud dataset is subjected to constant false alarm rate detection to filter out stray points and false alarm points, and retain the valid point cloud with a confidence level higher than a preset value to obtain the point cloud information.

3. The intelligent early warning method for temporary construction work in highway reconstruction and expansion according to claim 2, characterized in that, Based on the point cloud information, a vehicle detection target is obtained. Based on the vehicle detection target, a continuous time-series trajectory is maintained, resulting in a target state sequence set including: The point cloud information is voxelized and encoded using a VoxelNet network to obtain the encoding vector of each voxel. The encoding vector of each voxel is then used to model the spatial context in a sparse convolutional network to obtain a set of candidate target boxes. Multi-target tracking is performed based on the set of candidate target boxes and an interactive multi-model filter to obtain the vehicle detection target. Each detection target is assigned a unique identifier and a continuous time series trajectory is maintained to obtain a set of target state sequences.

4. The intelligent early warning method for temporary construction work in highway reconstruction and expansion according to claim 3, characterized in that, Based on the target state sequence set, a vehicle social pooling interaction feature is constructed, resulting in an input feature vector including: After organizing the target state sequence set into standardized spatiotemporal input data, the Z-Score method is used for standardization to obtain a standardized time series sequence. Based on the standardized time-series sequence, a vehicle social pooling interaction feature is constructed to obtain an input feature vector, including: Treating all vehicles as nodes and their relative positions and speeds as edges, a vehicle interaction graph is constructed at each time step. Weighted connection edges are established, with the weights determined by the relative position distance and speed difference between the two vehicles. Pooling is performed on the features of all neighboring vehicles of the target to obtain pooled features. The standardized state of the target vehicle itself is then concatenated with the pooled features to obtain the input feature vector.

5. The intelligent early warning method for temporary construction work in highway reconstruction and expansion according to claim 4, characterized in that, The trajectory prediction results obtained by combining the input feature vector with the long short-term memory network and the hybrid density network include: Based on the input feature vector and a long short-term memory network, a temporal encoding representation of the future trajectory is obtained; based on the temporal encoding representation of the future trajectory and a hybrid density network, multimodal modeling and probability prediction of the future trajectory distribution are performed to obtain the future trajectory distribution; based on the future trajectory distribution, road and construction zone constraints are fused and interactive multi-model filter correction is performed to obtain the trajectory prediction result.

6. The intelligent early warning method for temporary construction work in highway reconstruction and expansion according to claim 5, characterized in that, Based on the aforementioned vehicle risk factors, risk levels are classified to obtain vehicle risk results, including: The risk level is divided according to the vehicle risk factor. The risk level is defined by a segmented threshold method. Three risk ranges are obtained based on two preset thresholds. The vehicle risk factor is determined to be low risk, medium risk, or high risk based on its position within the three risk ranges. The risk level, the vehicle risk factor, and the identifier of the target vehicle are integrated to obtain the vehicle risk result.

7. The intelligent early warning method for temporary construction work in highway reconstruction and expansion according to claim 6, characterized in that, Generate a warning instruction based on the vehicle risk results; provide on-site warnings and vehicle-side human-machine interaction prompts based on the warning instruction, including: The warning instruction includes the risk level, the identifier of the target vehicle, and a preset intervention action; the warning instruction is distributed to warning devices at all levels through a message distribution mechanism to achieve on-site warning and vehicle-side human-machine interaction prompts.

8. An intelligent early warning system for temporary construction work on highway reconstruction and expansion, used in the method described in any one of claims 1 to 7, characterized in that, The system includes a first module, a second module, a third module, and a fourth module; The first module is used to acquire point cloud information of vehicles and environment near the construction area through radar; obtain vehicle detection targets based on the point cloud information; maintain continuous time series trajectories based on the vehicle detection targets; and obtain a set of target state sequences. The second module is used to construct vehicle social pooling interaction features based on the target state sequence set to obtain an input feature vector; and to obtain trajectory prediction results by combining the input feature vector with a long short-term memory network and a hybrid density network. The third module is used to calculate risk indicators for all times based on the trajectory prediction results, and obtain vehicle risk factors. The risk level is classified according to the vehicle risk factors to obtain the vehicle risk result; The fourth module is used to generate early warning instructions based on the vehicle risk results; The warning instructions will be used to provide on-site warnings and vehicle-side human-machine interaction prompts.

Images