Mobile maintenance vehicle rear-end collision risk real-time early warning method and system based on front time window

Abstract

The invention provides a mobile maintenance vehicle rear-end collision risk real-time early warning method and system based on a front time window, and relates to the technical field of intelligent traffic. The method comprises the steps of integrating sensors such as a GNSS / IMU, a millimeter wave radar and a binocular camera on a mobile maintenance vehicle, obtaining information such as the relative distance, the speed, the acceleration and the lane position of a backward coming vehicle and the maintenance vehicle, constructing a time sequence prediction model which takes historical fragments as input and applies front time window constraint to data nearby the current moment, and predicting the time sequence of the current time. Relative movement in a future short time period is predicted, safety indexes such as collision time TTC and required deceleration DRAC are calculated, threshold values are corrected in combination with road alignment, longitudinal slopes and visibility, and a directional horn is used for achieving rear-end collision early warning. According to the method, the false alarm rate and the missing alarm rate can be reduced in complex scenes such as curves, long slopes and low visibility, the early warning accuracy and robustness are improved, and rear-end collision accidents of mobile maintenance operation are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method and system for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window. Background Technology

[0002] In mobile work zones (including mobile maintenance and short-term lane occupancy), rear-end collisions between oncoming vehicles and work / buffer vehicles (such as TMA vehicles) are among the most common and highest-risk accident types. Empirical studies show that each collision involving TMA can save on accident costs compared to a scenario without TMA, and the investment can be recovered in less than a year for high-traffic facilities; however, such measures do not reduce the probability of collisions and complement rather than replace "soft warnings".

[0003] In terms of risk assessment methods, the field of traffic safety widely adopts alternative safety measures (SSMs), such as time-to-collision (TTC), deceleration required to avoid a collision (DRAC), and conflict potential index (CPI), to measure rear-end collision risk in near real-time and establish predictability correlations with accident data. Numerous publications and conference papers have verified the correlation between TTC / DRAC and rear-end collisions and their feasibility for early warning / control. However, existing engineering implementations primarily rely on fixed thresholds or rule-based triggers, rarely incorporating road geometry and visibility (such as curvature κ, longitudinal slope g, and stopping sight distance SSD) into threshold adaptation. According to the AASHTO / FHWA design and management guidelines, stopping sight distance and the 2.5-second perception-reaction time assumption, longitudinal slope, and adhesion conditions have a decisive impact on braking distance and visibility. This means that a simple "uniform TTC threshold across all scenarios" is prone to missed detections or false alarms on curves, long longitudinal slopes, or in low visibility conditions, making it difficult to reliably serve the complex scenarios of mobile work areas.

[0004] In trajectory / motion prediction, deep learning models (LSTM / GRU, Transformer, etc.) have become the mainstream for short- to medium-term vehicle trajectory prediction in recent years, exhibiting lower mean / end-point distance errors compared to kinematic baselines such as constant speed / uniform acceleration. However, issues such as long-term prediction drift, over-extrapolation of sudden changes in operating conditions, and limitations in edge computing power / latency still restrict their large-scale deployment in real-time vehicle warnings. In particular, common models often directly use adjacent historical segments to estimate the future, easily extrapolating "recent sudden braking / abrupt changes" to the future, leading to unstable warnings. Meanwhile, although Transformer-type models perform well at urban intersections and highly interactive scenarios, their high computational cost puts pressure on edge deployment and real-time performance.

[0005] In summary, the closest existing technologies are mainly: ① radar / camera-based TTC / DRAC threshold-triggered rear-end collision warning; ② edge-cloud prediction prototypes or simulation verification based on LSTM / Transformer. Their common shortcomings are: (a) thresholds are mostly fixed values, without adaptive correction based on geometric priors such as curvature, longitudinal slope, and visibility; (b) the time-series model does not explicitly handle the "pre-emptive window" effect, resulting in large extrapolation biases for adjacent historical disturbances; (c) the lack of uncertainty and sensor confidence fusion makes it prone to false alarms / missed alarms in the threshold-sensitive zone; and (d) the lack of low-latency implementation and interpretable output for vehicle edge computing, making it difficult to meet the engineering requirements of coordinating "millisecond-hundred-millisecond" alarms with a driver's 2.5s reaction time in mobile work areas. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a real-time early warning method and system for rear-end collision risks of mobile maintenance vehicles based on a forward time window. This invention solves the technical problems of the prior art, such as fixed early warning threshold settings, inability to adapt to complex road conditions, trajectory prediction models being easily affected by adjacent historical noise leading to extrapolation instability, and insufficient early warning accuracy and real-time performance due to the lack of a multi-source uncertainty fusion mechanism.

[0007] To achieve the above objectives, the present invention provides the following solution: A real-time early warning method for rear-end collision risk of mobile maintenance vehicles based on a forward time window includes: By integrating a GNSS / IMU all-in-one unit, a rearward millimeter-wave radar, and a binocular camera on the mobile maintenance vehicle, the positioning and attitude data of the mobile maintenance vehicle, the radar measurement data of the approaching vehicle, and the rearward video images are collected. Based on the positioning and attitude data, the radar measurement data, and the rearward video image, the relative motion parameters, lane relationship features, and road geometry and environmental features of the approaching vehicle are extracted in a unified coordinate system, and a multi-source temporal feature sequence with the current time as the cutoff point is constructed based on the extracted data. Apply a pre-window masking or continuous weight reduction processing to the historical segments in the multi-source time-series feature sequence that are immediately adjacent to the current time to obtain the feature sequence after pre-window processing; The feature sequence processed by the pre-window is input into a pre-constructed two-level temporal neural network, and the two-level temporal neural network is used to predict the relative motion trajectory of the following vehicle within a preset time period in the future. Based on the predicted relative motion trajectory, calculate alternative safety indicators; Based on the road geometry and environmental characteristics, a safety threshold correction model is established. The initial threshold of the alternative safety indicator is then modified according to the scenario using the safety threshold correction model to obtain a dynamic threshold. The alternative safety indicators modified according to the scenario, the dynamic threshold, the prediction uncertainty of the two-level temporal neural network, and the confidence level of the sensor are used as risk evidence. A unified risk score is calculated using a multi-source information fusion method, and early warning information is issued based on the unified risk score.

[0008] A real-time rear-end collision risk warning system for mobile maintenance vehicles based on a forward time window includes: The vehicle-mounted sensing module includes a GNSS / IMU integrated unit, a rear-facing millimeter-wave radar, and a binocular camera, used for data collection; An electronic control unit, connected to the vehicle-mounted sensing module, is used to perform the method as described in any one of claims 1 to 9; The early warning interaction module is connected to the electronic control unit and is used to provide audio-visual prompts and directional amplification based on the early warning information output by the electronic control unit.

[0009] The present invention discloses the following technical effects: This invention provides a real-time early warning method and system for rear-end collision risks of mobile maintenance vehicles based on a forward time window. The core technical features of this invention are "radar-video-GNSS / IMU multi-source trajectory-level fusion + forward time window masking + two-level temporal neural network + geometric prior correction + multi-index fusion (TTC / DRAC / collision probability) + low-latency implementation at the edge," resulting in the following inherent technical effects: First, by performing spatiotemporal alignment and trajectory-level fusion of millimeter-wave radar, binocular cameras, GNSS / IMU, and high-precision maps under a unified coordinate system, a multi-dimensional target state with high-precision longitudinal distance / speed and lateral lane relationships is obtained, providing a stable and complete input foundation for subsequent prediction and early warning. Second, by introducing a forward time window masking or continuous weight reduction mechanism, the "sudden extrapolation" of abnormal data such as sudden braking and lane changes immediately adjacent to the current moment is suppressed, significantly reducing trajectory drift and jitter in 0–6 s (especially 0–3 s) predictions, and improving the stability and accuracy of short-term predictions. Third, geometric and perceptual priors such as road curvature, longitudinal slope, speed limit, and visibility distance are incorporated into the adaptive correction of alternative safety indicator thresholds such as TTC, DRAC, and collision probability. This makes the identification of rear-end collision risks in complex conditions such as curves, long slopes, and low visibility more in line with actual safety margins, reducing false alarms. Fourth, geometrically corrected TTC / DRAC, collision probability, prediction uncertainty, and sensor confidence are fused as multi-source evidence. Compared with a single TTC trigger logic, this significantly reduces the false alarm rate while maintaining a high recall rate, and outputs warning reasons such as trigger indicators and data quality, facilitating operation and maintenance review and accountability. Fifth, by using cloud-based training, edge distillation, and model compression, lightweight inference for vehicles is achieved. The closed-loop latency from data acquisition, fusion, prediction to early warning output is controlled to the order of approximately 100 ms. Early warnings can be issued within the driver's perception-reaction time of approximately 2.5 s, which is generally taken in road design, significantly expanding the driver's effective response window. At the same time, it can form a collaborative system of "collision prevention + harm reduction" with hard protection facilities such as TMA, reducing the risk of injury to workers and the economic losses caused by work stoppage due to accidents. It has good engineering feasibility and social and economic benefits for large-scale promotion. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart of a real-time early warning method for rear-end collision risk of mobile maintenance vehicles based on a forward time window is provided in an embodiment of the present invention; Figure 2Installation diagram of the integrated early warning device for rear-end collision risk of mobile maintenance vehicles provided in this embodiment of the invention; Figure 3 This is a schematic diagram illustrating the use of millimeter-wave radar to determine the relative distance between a vehicle behind and a maintenance vehicle, as provided in an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] like Figure 1 As shown, this invention provides a real-time early warning method for rear-end collision risk of mobile maintenance vehicles based on a forward time window, including: Step 100: Collect the positioning and attitude data of the mobile maintenance vehicle, the radar measurement data of the approaching vehicle, and the rearward video images by using the GNSS / IMU integrated unit, rearward millimeter-wave radar, and binocular camera integrated on the mobile maintenance vehicle. Step 200: Based on the positioning and attitude data, the radar measurement data and the rearward video image, extract the relative motion parameters, lane relationship features and road geometry and environmental features of the approaching vehicle in a unified coordinate system, and construct a multi-source temporal feature sequence with the current time as the cutoff point based on the extracted data; Step 300: Apply a pre-window masking or continuous weight reduction processing to the historical segments in the multi-source time-series feature sequence that are immediately adjacent to the current time to obtain the feature sequence after pre-window processing; Step 400: Input the feature sequence processed by the pre-window into a pre-constructed two-level temporal neural network, and use the two-level temporal neural network to predict the relative motion trajectory of the following vehicle within a preset time period in the future; Step 500: Calculate alternative safety indicators based on the predicted relative motion trajectory; Step 600: Based on the road geometry and environmental features, establish a safety threshold correction model, and use the safety threshold correction model to modify the initial threshold of the alternative safety indicator in a scenario-based manner to obtain a dynamic threshold; Step 700: Using the scenario-corrected alternative safety indicators, the dynamic threshold, the prediction uncertainty of the two-level temporal neural network, and the confidence level of the sensor as risk evidence, a unified risk score is calculated using a multi-source information fusion method, and a warning information is issued based on the unified risk score.

[0015] Specifically, the specific process of the method in this embodiment is as follows: The vehicle integrates a binocular camera, a rear-facing millimeter-wave radar, a GNSS / IMU all-in-one unit, and a vehicle CAN interface. GNSS / IMU is used to acquire the vehicle's precise position, heading, and attitude in the road coordinate system. Combined with a high-precision map, the extrinsic parameter relationships between the map coordinate system, road coordinate system, and vehicle coordinate system are established. Intrinsic and extrinsic parameter calibrations are performed on the binocular camera to obtain the projection relationship from the pixel plane to the vehicle coordinate system. The millimeter-wave radar is installed and calibrated to obtain the transformation matrix from the radar coordinate system to the vehicle coordinate system. Using the GNSS or ECU's local high-precision clock as a reference, the radar, camera, and IMU data are timestamped to form a unified time base.

[0016] Under a unified coordinate system, radar trajectories and video-detected targets are correlated using methods such as gating matching, Hungarian algorithm, or JPDA. For successfully matched targets, weighted Kalman filtering or Bayesian fusion is used to combine the high accuracy of radar in longitudinal distance / velocity with the advantages of video in lateral position / lane relationship, resulting in a multi-dimensional target state vector containing information such as relative distance, longitudinal / lateral velocity, acceleration, lane affiliation, and contour scale. For unmatched targets, sensor-independent trajectories are retained and confidence levels are marked.

[0017] Based on the multi-source fusion state of the mobile maintenance vehicle and various backward targets, a temporal feature window covering the most recent period is constructed. Features include relative distance, relative speed, relative acceleration, relative azimuth, lane affiliation, lateral offset, cut-in / cut-out trends, road curvature, longitudinal slope, speed limit, visible distance, and the maintenance vehicle's own speed, braking, steering, and guide vane operation status. At each prediction time, a pre-window masking or continuous weight reduction is applied to a time segment immediately adjacent to the current time in the historical window. High-frequency noise signals such as abnormal sudden braking and sharp steering within this time segment are reduced in weight or masked in the temporal model input, making the model rely more on earlier stable driving behavior patterns. This suppresses the excessive extrapolation of sudden conditions on future trajectory prediction, forming a multi-dimensional temporal feature sequence processed by the pre-window.

[0018] The temporal features processed by the pre-window are input into a two-level temporal neural network model: the first level is a lightweight recurrent neural network (such as GRU or LSTM) for short-term relative trajectory prediction from 0 to 3 seconds, ensuring a rapid response to approaching risks; the second level is a Transformer encoder with causal masking or a one-dimensional convolutional-attention hybrid network for medium-term relative trajectory prediction from 3 to 6 seconds, characterizing the complex behavioral trends of rearward vehicles such as lane changing, lane switching, and overtaking. A multi-task loss function simultaneously constrains multiple physical quantities such as relative position, relative velocity, and lateral acceleration, and incorporates physical constraint regularization terms that limit the range of maximum longitudinal deceleration and lateral acceleration to ensure the physical feasibility of the predicted trajectory. The model is trained offline in the cloud using a large amount of measured and simulated data. After training, it is deployed to the vehicle ECU through model compression and knowledge distillation to achieve low-latency inference at the edge. Simultaneously, the variance or confidence interval for each prediction time point is output to quantify the prediction uncertainty.

[0019] Based on predicted relative trajectories, alternative safety indicators such as Time to Collision (TTC) and Deceleration Required to Avoid Collision (DRAC) are calculated in real time. A Stop Sight Distance (SSD) model is introduced, incorporating parameters such as driver perception-reaction time, road adhesion coefficient, road longitudinal slope, speed limit, and visibility distance into the threshold adaptive correction. A dynamic threshold function suitable for different scenarios is constructed: under straight, flat, and good visibility conditions, the threshold is basically consistent with the conventional setting; under conditions such as curves, long slopes, and low visibility, the warning threshold is automatically tightened or relaxed based on high-precision maps and visibility distance estimation, so that the risk assessment matches the driver's actual safety margin.

[0020] Using geometrically corrected TTC, DRAC and their temporal evolution characteristics, prediction uncertainty, multi-source sensor confidence and target behavior characteristics as multiple risk evidences, Kalman fusion, Bayesian inference or DS evidence theory are used to fuse multi-source information to obtain a normalized unified risk score R. According to preset thresholds, R is divided into multiple warning levels, such as red emergency warning, orange severe warning, yellow alert warning, etc., and the current value of TTC / DRAC and its changing trend are combined to determine whether to trigger different levels of warning.

[0021] For high-level warnings, the ECU controls the audible and visual alarms in the driver's cab, the vibration of the seat or steering wheel, and the external guide panels and warning lights to provide alerts. It can optionally broadcast rear-end collision risk information to the rear vehicle via the V2X communication module. In the event of partial sensor failure such as GNSS lockout, camera obstruction, or radar malfunction, the system automatically switches to a simplified fusion strategy and tightens the threshold based on data quality, while still maintaining a conservative rear-end collision warning capability.

[0022] Furthermore, the relative motion parameters include the relative distance in the vehicle coordinate system, the relative convergence velocity along the connecting line direction, and the relative acceleration.

[0023] Furthermore, based on the extracted data, a multi-source time-series feature sequence is constructed with the current time as the cutoff point, including: Based on one of the unscented Kalman filtering or extended Kalman filtering methods, the state of the approaching vehicle is estimated according to the radar measurement data and the positioning and attitude data to obtain the relative motion parameters; based on the rearward video image, the lateral position, lane number and cutting-in or cutting-out trend of the approaching vehicle are identified to obtain the lane relationship features. Based on the matching results of the positioning and attitude data and the high-precision map, the road curvature, longitudinal slope, speed limit and visibility distance are estimated to obtain the road geometry and environmental features; The relative motion parameters, lane relationship features, and road geometry and environment features are stacked in chronological order to obtain the multi-source temporal feature sequence.

[0024] Specifically, such as Figure 2-3 As shown, a GNSS / IMU integrated unit, a rear-facing millimeter-wave radar, and a binocular camera are integrated into a work vehicle equipped with mobile maintenance equipment and guide plates. All sensors are connected to the vehicle's electronic control unit (ECU) via Ethernet or CAN (FD). The ECU is powered by the vehicle's 12 / 24V power supply, with fuses and TVS overvoltage protection configured in the power supply branch. The binocular camera's field of view covers a distance of at least 100-150m behind the maintenance vehicle and extends to 2-4 lanes. The millimeter-wave radar preferably has a detection range greater than 200m, and its field of view covers the lane where the vehicle is located and adjacent lanes. The update frequency of the GNSS integrated unit is preferably no less than 10Hz, the update frequency of the IMU is preferably no less than 10Hz, and a map coordinate system is defined. Road coordinate system Vehicle coordinate system In addition to the coordinate systems of each sensor (camera image plane pixel coordinate system, camera coordinate system, radar coordinate system, etc.), the map coordinate system is bound to parameters such as the road centerline and station numbers; the road coordinate system is based on the tangent direction of the current road centerline. axis, normal direction is Axis; The vehicle coordinate system has its origin at the vehicle's geometric center or the center of the rear axle, extending backwards. The axis, to the left is The axis, vertically upward is The system performs intrinsic parameter calibration (focal length, principal point, distortion coefficients, etc.) and extrinsic parameter calibration on the binocular camera to obtain the rotation matrix and translation vector from the camera coordinate system to the vehicle coordinate system. It also performs installation calibration on the millimeter-wave radar to obtain the extrinsic parameter transformation from the radar coordinate system to the vehicle coordinate system. Furthermore, it obtains the transformation relationship from the vehicle coordinate system to the road coordinate system and the map coordinate system through GNSS and high-precision map matching. The system time base uses GNSSSPPS as the master clock, and the timestamps of all sensors are unified on the ECU side, employing an extrinsic parameter matrix. Complete the geographic coordinate system With vehicle coordinate system (x-front, y-left, z-up) conversion. High-precision map data (road centerline, curvature) Longitudinal slope Speed ​​limit and visibility distance estimation are pre-loaded into the ECU or sent from the cloud and are aligned with real-time positioning during operation.

[0025] Furthermore, a pre-window masking or continuous weight reduction processing is applied to the historical segments immediately adjacent to the current time in the multi-source time-series feature sequence to obtain the feature sequence after pre-window processing, including: Set a suppression zone with the current time as the end; If the aforementioned pre-window masking is executed, the feature weights falling within the suppression region will be reset to zero; If the continuous weight reduction is performed, the feature weights falling into the suppression region are processed according to a preset decay function, so that features closer to the current time in the suppression region have smaller weights.

[0026] Furthermore, the two-level temporal neural network includes: a cascaded first-level subnetwork and a second-level subnetwork.

[0027] Furthermore, the step of using the two-level temporal neural network to predict the relative motion trajectory of an approaching vehicle within a preset future time period includes: Using a lightweight recurrent neural network of the first-level subnetwork, the short-term relative trajectory prediction results within the next 0 to 3 seconds are output; The Transformer encoder with causal masking in the second-level sub-network outputs the mid-term relative trajectory prediction result within the next 3 to 6 seconds; wherein, the two-level temporal neural network is trained by a multi-task loss function, which includes constraints on relative displacement, relative velocity, lateral acceleration and physical constraint penalty terms.

[0028] Specifically, during operation, the binocular cameras acquire video images of the area behind the maintenance vehicle. Target detection and depth estimation algorithms identify vehicles and lane markings behind the vehicle, obtaining: the type of the target vehicle (e.g., passenger car, truck); the lateral position and offset of the target vehicle in image coordinates or 3D coordinates; the lane number of the target vehicle and whether it is in the same lane as the maintenance vehicle; whether the target vehicle is attempting to cut into or out of the maintenance vehicle's lane; and lane geometry information such as lane markings and directional signs. The rear-facing radar outputs target measurements in each cycle. ,in For measuring distance, Radial velocity, This is the azimuth angle. The relative position in the vehicle coordinate system is expressed as... relative velocity Obtained from multi-frame azimuth difference The system performs decomposition. It uses an unscented / extended Kalman filter to estimate the state of the target state. Satisfies the discrete constant acceleration model: ; The sampling period is : ; Measurement equation: ; Established based on geometric relationships. Filter output relative distance: ; Relative convergence velocity along the connecting line: ; and relative acceleration This constitutes the fundamental quantity of longitudinal relative motion.

[0029] Within the same sampling period, the binocular cameras detect the image frame of the backward vehicle and estimate its lateral position. Lane number and cut-in / cut-out trend labels This forms the lane and lateral relationship characteristics; GNSS and high-precision map matching yields the curvature κ and longitudinal slope of the current road segment. Speed ​​limit, visibility distance estimation Road geometry and line-of-sight features; vehicle dynamic signals are read via IMU to form vehicle operating status features. Each sensor also provides its own confidence level. .

[0030] In summary, at any time Constructing multi-source temporal feature vectors: ; in: The state estimation mainly comes from radar and IMU; From GNSS and high-precision maps; From video and lane recognition; Signal from IMU; Characterize multi-source perceived quality.

[0031] To suppress the "over-extrapolation" of future states by adjacent historical segments, the system compares the historical window at each prediction time t. Adjacent segments within Apply a front-window mask to the input sequence. ; Define the masking function: ; Obtain the masked sequence: ; Alternatively, continuous weight reduction can be used: ; By assigning smaller weights to features closer to the current moment, the overall approximation trend is maintained while mitigating the impact of recent anomalies such as sudden braking, lane changes, and momentary occlusions on the prediction. The feature sequence after masking... The data is fed into a two-stage temporal neural network: the first stage is a lightweight GRU / LSTM for short-term relative trajectory prediction (0–3 s), and the output is: ; The second stage is a Transformer Encoder with causal masking, used for mid-term prediction output from 3 to 6 seconds: ; The two-level network is trained in the cloud using a combination of real-world and simulated datasets, and then distilled into a lightweight version for edge devices, combining multi-task losses: ; Where ADE / FDE represent the average / endpoint position error and the lateral acceleration, respectively. ; The adhesion coefficient, Used to suppress unreasonable lateral / longitudinal acceleration. The adhesion coefficient, (For gravitational acceleration) The network outputs the future 0–6 s values. and uncertainty (such as variance) ).

[0032] During the risk quantification phase, the system calculates the Alternative Safety Indicator (SSM) in real time based on predicted relative motion. The Time to Collision (TTC) is considered under approach conditions. The definition is as follows: ; Otherwise take The deceleration required to avoid a collision (DRAC) is given in the longitudinal component: ; longitudinal slope angle Obtained from GNSS / IMU+ maps, reflecting the impact of slope on available braking capacity, and to reflect road geometry and sight distance constraints, an approximation of stopping sight distance (SSD) is introduced: ; Where v represents the speed characteristic of the vehicle or relative speed. The driver's perception-reaction time (typically taken as 2.5 s) is given by μ, which is estimated from the road adhesion conditions. This is combined with the curvature κ and the visible distance. Establish an adaptive threshold function: ; ; in , Obtained through offline calibration or Bayesian optimization. Automatically tightens warning thresholds for conditions such as curves, long slopes, and low visibility.

[0033] Furthermore, the alternative security indicators include: Collision time and the deceleration required to avoid a collision.

[0034] Furthermore, the initial threshold of the alternative security indicator is modified according to the scenario using the security threshold correction model to obtain a dynamic threshold, including: A parking sight distance model is introduced, in which the driver's perception-reaction time, road surface adhesion coefficient, road longitudinal slope and speed limit in the road geometry and environmental features are input into the parking sight distance model to calculate the scenario-based minimum safe distance or minimum safe deceleration; Based on the minimum safe distance or minimum safe deceleration, an adaptive threshold function is constructed; Using the adaptive threshold function, based on the road curvature and visible distance estimation in the road geometry and environmental features, the initial thresholds for the collision time and the deceleration required to avoid the collision are tightened or relaxed to obtain the dynamic threshold.

[0035] Furthermore, the calculation of the unified risk score using a multi-source information fusion method includes: The collision time, the deceleration required to avoid the collision, the prediction uncertainty, and the confidence level are converted into corresponding evidence. The evidence body is fused using the combination rules of the D-S evidence theory to obtain the normalized unified risk score; When the unified risk score exceeds a preset level threshold, a directional loudspeaker is triggered to issue an early warning.

[0036] Specifically, based on predicted longitudinal / lateral trajectories and lane relationships The system further constructs conflict probability or risk intensity indicators, for example, mapping "whether to remain in the same lane and longitudinally close in the future" to... Its input depends on the lane / lateral features given in the video. Finally, the geometrically / line-of-sight corrected TTC, DRAC, and conflict probability are used as inputs. Model uncertainty and confidence level of multi-source sensors The data is mapped to several pieces of risk evidence, which are then combined using the D-S evidence theory. The rules for their combination are as follows: ; This is further extended to multiple pieces of evidence such as TTC, DRAC, and conflict probability, resulting in a unified risk score R∈[0,1]. ECU classifies risk based on the logical relationship between R and TTC, DRAC, and their thresholds: when R reaches a high threshold or An emergency warning is triggered and a directional horn is activated to alert vehicles approaching from behind. The above processing operates in a closed loop at the vehicle's electronic control unit with a fixed cycle: each cycle completes radar / video / GNSS / IMU data acquisition, multi-source feature construction, forward time window masking, two-level time series prediction, SSM calculation, D-S fusion to HMI / V2X deployment. After model compression and optimization, the total latency is controlled within approximately 100 ms, ensuring that the warning takes effect within the approximately 2.5 s driver perception-reaction time window commonly used in road design, forming a real-time rear-end collision warning link of "multi-source perception-prediction-quantification-decision".

[0037] In scenarios such as GNSS lock-off or temporary camera failure, the system automatically degrades to the "radar + IMU relative quantity" mode based on the confidence level γ, and tightens accordingly. , Thresholds are set to increase safety redundancy and ensure that conservative rear-end collision risk warnings are still provided even when some sensors are unavailable.

[0038] This embodiment also provides a real-time rear-end collision risk warning system for mobile maintenance vehicles based on a forward time window, including: The vehicle-mounted sensing module includes a GNSS / IMU integrated unit, a rear-facing millimeter-wave radar, and a binocular camera, used for data collection; The electronic control unit is connected to the vehicle-mounted sensing module; The early warning interaction module is connected to the electronic control unit and is used to provide audio-visual prompts and directional amplification based on the early warning information output by the electronic control unit.

[0039] Specifically, the GNSS / IMU integrated unit is used to acquire the positioning and attitude information of the mobile maintenance vehicle; A binocular camera is used to identify and classify vehicles behind and detect lane lines; a rearward millimeter-wave radar is used to detect the relative position and movement of vehicles approaching from behind. The high-precision map storage module is used to store information such as road centerline, curvature, longitudinal slope, speed limit, and visibility distance estimation. The vehicle electronic control unit (ECU) is used for: performing coordinate alignment and state estimation on multi-source sensor data to construct a temporal feature sequence; applying a front-end time window mask to historical window features and calling a two-level temporal neural network to predict future relative motion; calculating alternative safety indicators such as TTC and DRAC and performing threshold adaptive correction; and fusing multi-source risk evidence to generate a unified risk score and complete risk classification. The human-computer interaction and directional speaker module is used to broadcast early warning information in a visual, audio-visual manner; The vehicle electronic control unit (ECU) has the following independent modules: data preprocessing and coordinate transformation module, target state estimation module, front-end time window processing and feature management module, two-level time series prediction module, alternative safety indicator calculation and threshold correction module, and risk fusion and early warning decision module. The modules interact with each other via a bus.

[0040] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for real-time early warning of rear-end collision risk of mobile maintenance vehicles based on a preceding time window.

[0041] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0042] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, characterized in that, include: By integrating a GNSS / IMU all-in-one unit, a rearward millimeter-wave radar, and a binocular camera on the mobile maintenance vehicle, the positioning and attitude data of the mobile maintenance vehicle, the radar measurement data of the approaching vehicle, and the rearward video images are collected. Based on the positioning and attitude data, the radar measurement data, and the rearward video image, the relative motion parameters, lane relationship features, and road geometry and environmental features of the approaching vehicle are extracted in a unified coordinate system, and a multi-source temporal feature sequence with the current time as the cutoff point is constructed based on the extracted data. Apply a pre-window masking or continuous weight reduction processing to the historical segments in the multi-source time-series feature sequence that are immediately adjacent to the current time to obtain the feature sequence after pre-window processing; The feature sequence processed by the pre-window is input into a pre-constructed two-level temporal neural network, and the two-level temporal neural network is used to predict the relative motion trajectory of the following vehicle within a preset time period in the future. Based on the predicted relative motion trajectory, calculate alternative safety indicators; Based on the road geometry and environmental characteristics, a safety threshold correction model is established. The initial threshold of the alternative safety indicator is then modified according to the scenario using the safety threshold correction model to obtain a dynamic threshold. The alternative safety indicators modified according to the scenario, the dynamic threshold, the prediction uncertainty of the two-level temporal neural network, and the confidence level of the sensor are used as risk evidence. A unified risk score is calculated using a multi-source information fusion method, and early warning information is issued based on the unified risk score.

2. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 1, is characterized in that... The relative motion parameters include the relative distance in the vehicle coordinate system, the relative convergence velocity along the connecting line direction, and the relative acceleration.

3. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 2, is characterized in that... Based on the extracted data, a multi-source time-series feature sequence is constructed with the current time as the cutoff point, including: Based on one of the unscented Kalman filtering or extended Kalman filtering methods, the state of the approaching vehicle is estimated according to the radar measurement data and the positioning and attitude data to obtain the relative motion parameters; based on the rearward video image, the lateral position, lane number and cutting-in or cutting-out trend of the approaching vehicle are identified to obtain the lane relationship features. Based on the matching results of the positioning and attitude data and the high-precision map, the road curvature, longitudinal slope, speed limit and visibility distance are estimated to obtain the road geometry and environmental features; The relative motion parameters, lane relationship features, and road geometry and environment features are stacked in chronological order to obtain the multi-source temporal feature sequence.

4. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 1, is characterized in that... Applying a pre-window masking or continuous weight reduction processing to the historical segments immediately adjacent to the current time in the multi-source time-series feature sequence yields a feature sequence processed by the pre-window, including: Set a suppression zone with the current time as the end; If the aforementioned pre-window masking is executed, the feature weights falling within the suppression region will be reset to zero; If the continuous weight reduction is performed, the feature weights falling into the suppression region are processed according to a preset decay function, so that features closer to the current time in the suppression region have smaller weights.

5. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 1, is characterized in that... The two-level temporal neural network includes: a cascaded first-level subnetwork and a second-level subnetwork.

6. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 5, is characterized in that... The method of predicting the relative motion trajectory of an approaching vehicle within a preset time period using the two-level temporal neural network includes: Using a lightweight recurrent neural network of the first-level subnetwork, the short-term relative trajectory prediction results within the next 0 to 3 seconds are output; The Transformer encoder with causal masking in the second-level sub-network outputs the mid-term relative trajectory prediction result within the next 3 to 6 seconds; wherein, the two-level temporal neural network is trained by a multi-task loss function, which includes constraints on relative displacement, relative velocity, lateral acceleration and physical constraint penalty terms.

7. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 1, is characterized in that... The alternative security indicators include: Collision time and the deceleration required to avoid a collision.

8. The method for real-time early warning of rear-end collision risk for mobile maintenance vehicles based on a forward time window, as described in claim 7, is characterized in that... The initial threshold of the alternative security indicator is modified according to the scenario using the security threshold correction model to obtain a dynamic threshold, including: A parking sight distance model is introduced, in which the driver's perception-reaction time, road surface adhesion coefficient, road longitudinal slope and speed limit in the road geometry and environmental features are input into the parking sight distance model to calculate the scenario-based minimum safe distance or minimum safe deceleration; Based on the minimum safe distance or minimum safe deceleration, an adaptive threshold function is constructed; Using the adaptive threshold function, based on the road curvature and visible distance estimation in the road geometry and environmental features, the initial thresholds for the collision time and the deceleration required to avoid the collision are tightened or relaxed to obtain the dynamic threshold.

9. A real-time early warning method for rear-end collision risk of a mobile maintenance vehicle based on a forward time window, as described in claim 8, is characterized in that... The method of calculating a unified risk score using multi-source information fusion includes: The collision time, the deceleration required to avoid the collision, the prediction uncertainty, and the confidence level are converted into corresponding evidence. The evidence body is fused using the combination rules of the D-S evidence theory to obtain the normalized unified risk score; When the unified risk score exceeds a preset level threshold, a directional loudspeaker is triggered to issue an early warning.

10. A real-time early warning system for rear-end collision risk of mobile maintenance vehicles based on a forward time window, characterized in that, include: The vehicle-mounted sensing module includes a GNSS / IMU integrated unit, a rear-facing millimeter-wave radar, and a binocular camera, used for data collection; An electronic control unit, connected to the vehicle-mounted sensing module, is used to perform the method as described in any one of claims 1 to 9; The early warning interaction module is connected to the electronic control unit and is used to provide audio-visual prompts and directional amplification based on the early warning information output by the electronic control unit.

Images