CN-121982930-A - Mobile maintenance vehicle rear-end collision risk real-time early warning method and system based on front time window

CN121982930ACN 121982930 ACN121982930 ACN 121982930ACN-121982930-A

Abstract

The invention provides a real-time early warning method and system for rear-end collision risk of a mobile maintenance vehicle based on a front time window, and relates to the technical field of intelligent transportation. The method comprises the steps of integrating sensors such as a GNSS/IMU, a millimeter wave radar, a binocular camera and the like on a mobile maintenance vehicle, acquiring information such as relative distance, speed, acceleration, lane position and the like of a backward vehicle and the maintenance vehicle, constructing a time sequence prediction model which takes a history segment as input and applies a front time window constraint to data near the current moment, predicting relative motion in a short time period in the future, calculating safety indexes such as collision time TTC, required deceleration DRAC and the like, correcting a threshold value by combining road linearity, longitudinal slope and visibility, and realizing rear-end collision early warning by utilizing a directional loudspeaker. The invention can reduce false alarm and missing report rate under complex scenes such as curves, long slopes, low visibility and the like, improve early warning accuracy and robustness, and reduce rear-end collision accidents of mobile maintenance operation.

Inventors

ShangGuan Qiangqiang
WANG JUNHUA
FU TING
LI YUHANG

Assignees

同济大学

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (10)

1. A real-time early warning method for rear-end collision risk of a mobile maintenance vehicle based on a front time window is characterized by comprising the following steps: acquiring positioning and attitude data of the mobile maintenance vehicle, radar measurement data of a backward coming vehicle and a backward video image through a GNSS/IMU integrated machine, a backward millimeter wave radar and a binocular camera which are integrated on the mobile maintenance vehicle; Based on the positioning and attitude data, the radar measurement data and the backward video image, extracting relative motion parameters, lane relation characteristics and road geometry and environment characteristics of a backward vehicle under a unified coordinate system, and constructing a multi-source time sequence characteristic sequence taking the current moment as a cut-off point based on the extracted data; applying a pre-time window mask or continuous weight reduction treatment to a history segment adjacent to the current moment in the multi-source time sequence characteristic sequence to obtain a characteristic sequence subjected to pre-time window treatment; inputting the feature sequence processed by the pre-time window into a pre-constructed two-stage time sequence neural network, and predicting the relative motion track of the backward vehicle in a future preset time period by using the two-stage time sequence neural network; Calculating a substitute safety index based on the predicted relative motion trail; based on the road geometry and the environmental characteristics, a safety threshold correction model is established, and the safety threshold correction model is utilized to conduct scene correction on the initial threshold value of the alternative safety index, so as to obtain a dynamic threshold value; And taking the scene-corrected alternative safety index, the dynamic threshold value, the prediction uncertainty of the two-stage time sequence neural network and the confidence level of the sensor as risk evidences, calculating a unified risk score by adopting a multi-source information fusion method, and issuing early warning information according to the unified risk score.
2. The real-time early warning method for rear-end collision risk of the mobile maintenance vehicle based on the front time window according to claim 1, wherein the relative motion parameters comprise relative distance under a vehicle coordinate system, relative convergence speed along a connecting line direction and relative acceleration.
3. The real-time early warning method for rear-end collision risk of mobile maintenance vehicle based on a front time window according to claim 2, wherein the constructing a multi-source time sequence feature sequence with the current moment as a cut-off point based on the extracted data comprises the following steps: Based on one of unscented Kalman filtering or extended Kalman filtering methods, carrying out state estimation on a backward vehicle according to the radar measurement data and the positioning and gesture data to obtain the relative motion parameters; based on the matching result of the positioning and posture data and the high-precision map, obtaining road curvature, longitudinal slope, speed limit and visible distance estimation, and obtaining the road geometry and environment characteristics; and stacking the relative motion parameters, the lane relation features, the road geometry and the environment features in time sequence to obtain the multi-source time sequence feature sequence.
4. The method for real-time early warning of rear-end collision risk of mobile maintenance vehicle based on a pre-time window according to claim 1, wherein the method for real-time early warning of rear-end collision risk of mobile maintenance vehicle based on the pre-time window is characterized by applying a pre-time window mask or continuous weight reduction process to a historical segment in the multi-source time sequence feature sequence, so as to obtain a feature sequence processed by the pre-time window, and comprises the following steps: setting a suppression zone taking the current moment as the tail end; If the front time window mask is executed, setting the feature weight falling into the suppression zone to zero; And if the continuous weight reduction is executed, processing the feature weights falling into the suppression zone according to a preset attenuation function, so that features in the suppression zone which are closer to the current moment have smaller weights.
5. The real-time early warning method for rear-end collision risk of the mobile maintenance vehicle based on the front time window of claim 1 is characterized in that the two-stage time sequence neural network comprises a first-stage subnetwork and a second-stage subnetwork which are cascaded.
6. The method for real-time early warning of rear-end collision risk of a mobile maintenance vehicle based on a pre-time window according to claim 5, wherein the predicting the relative motion track of the backward vehicle in the future preset time period by using the two-stage time sequence neural network comprises the following steps: Outputting a short-term relative track prediction result within 0 to 3 seconds in the future by using the lightweight cyclic neural network of the first-stage subnetwork; And outputting a mid-term relative track prediction result within 3 to 6 seconds in the future by using a causal mask converter encoder of the second level sub-network, wherein the two-level time sequence neural network is trained by a multi-task loss function, and the multi-task loss function comprises constraints on relative displacement, relative speed, lateral acceleration and physical constraint penalty items.
7. The real-time early warning method for rear-end collision risk of mobile maintenance vehicle based on a front time window according to claim 1, wherein the replacement safety index comprises: Time of collision and deceleration required to avoid collision.
8. The real-time early warning method for rear-end collision risk of mobile maintenance vehicle based on a pre-positioned time window according to claim 7, wherein the method for performing scene correction on the initial threshold value of the substitute safety index by using the safety threshold value correction model to obtain a dynamic threshold value comprises the following steps: A parking sight distance model is introduced, and the driver perception-reaction time, road adhesion coefficient, road longitudinal slope and speed limit in the road geometry and environmental characteristics are input into the parking sight distance model so as to calculate the minimum safe distance or minimum safe deceleration of scene; Constructing an adaptive threshold function based on the minimum safe distance or minimum safe deceleration; And tightening or relaxing the initial threshold value of the collision time and the deceleration required by collision avoidance according to the road curvature and the visible distance estimation in the road geometry and the environment characteristics by using the self-adaptive threshold function to obtain the dynamic threshold value.
9. The real-time early warning method for rear-end collision risk of mobile maintenance vehicle based on the front time window of claim 8, wherein the calculating the unified risk score by adopting the multi-source information fusion method comprises the following steps: converting the collision time, the deceleration required to avoid the collision, the prediction uncertainty and the confidence into corresponding evidence; fusing the evidence bodies by utilizing a combination rule of a D-S evidence theory to obtain the normalized unified risk score; And triggering the directional loudspeaker to perform early warning when the unified risk score exceeds a preset level threshold.
10. A real-time early warning system for rear-end collision risk of a mobile maintenance vehicle based on a front time window is characterized by comprising: the vehicle-mounted sensing module comprises a GNSS/IMU integrated machine, a backward millimeter wave radar and a binocular camera and is used for acquiring data; An electronic control unit, connected to the on-board perception module, for performing the method according to any one of claims 1 to 9; and the early warning interaction module is connected with the electronic control unit and used for carrying out acousto-optic prompt and directional sound amplification according to the early warning information output by the electronic control unit.

Description

Mobile maintenance vehicle rear-end collision risk real-time early warning method and system based on front time window Technical Field The invention relates to the technical field of intelligent transportation, in particular to a real-time early warning method and system for rear-end collision risk of a mobile maintenance vehicle based on a front time window. Background Rear-end collisions with work/buffer vehicles (e.g., TMA vehicles) by the oncoming vehicles in mobile work areas (including mobile maintenance, short occupancy) are one of the most common and highest risk accident configurations. Proved by research, each collision involving TMA can save accident cost compared with the 'no TMA' situation, and investment can be recovered in less than one year on high traffic facilities, but the probability of collision is not reduced by such measures, and the collision and the 'soft early warning' form complementary but not alternative relation. In the aspect of risk discrimination methods, alternative safety indexes (Surrogate Safety Measures, SSM) such as collision time TTC, deceleration DRAC required for avoiding collision, and collision potential index CPI are widely adopted in the traffic safety field to measure rear-end collision risk in near real time and establish predictability association with accident data. A large number of literature-on-meeting papers verify the relevance of TTC/DRAC to rear-end collision and feasibility of use for early warning/management. However, the existing engineering implementation is mainly triggered by fixed threshold values or rules, and the road geometry and visibility (such as curvature kappa, longitudinal slope g and parking sight distance SSD) are rarely included in the threshold value self-adaption, while the perception-reaction time assumption, longitudinal slope and attachment conditions of the parking sight distance and 2.5 s have decisive influence on the braking distance and the visible distance according to the design and management guidelines of AASHTO/FHWA. This means that a simple "full scene unified TTC threshold" is prone to false alarms or missed checks in curves, long hills or low visibility, and it is difficult to stabilize complex scenes that serve a mobile work area. In the aspect of track/motion prediction, a deep learning model (LSTM/GRU, transformer and the like) has become the main stream of short-medium-term vehicle track prediction, has lower average/tail end distance errors compared with a constant speed/uniform acceleration and other kinematic baselines, but the problems of long-term prediction drift, over extrapolation of sudden variable working conditions, limited end-side computing force/time delay and the like still limit the large-scale landing of the vehicle-mounted real-time early warning system. In particular, the common model directly utilizes the immediate history segments to estimate the future, so that the sudden brake/mutation which happens immediately is easy to extrapolate to the future, the early warning is not steady, and meanwhile, the Transformer model has prominent performance in urban intersections and intensive interaction scenes, but has high calculation cost, and the stress is formed by opposite end side deployment and real-time performance. In summary, the closest prior art mainly comprises the following steps of trigger type rear-end collision early warning of TTC/DRAC threshold based on radar/camera ① and end cloud prediction prototype or simulation verification based on LSTM/transducer ②. The method has the common defects that (a) the threshold value is a fixed value, the geometric priors such as curvature, longitudinal slope, visibility and the like are not combined for self-adaptive correction, (b) a time sequence model does not explicitly process a front time window effect, extrapolation deviation on the adjacent history disturbance is large, (c) uncertainty and sensing confidence coefficient fusion is lacked, false alarm/false alarm is easy to generate in a threshold sensitive zone, and (d) low-delay realization and interpretable output facing to the vehicle-mounted edge are lacked, so that engineering requirements of cooperation of a mobile operation area' millisecond-hundred millisecond alarm and 2.5s reaction time of a driver are difficult to meet. Disclosure of Invention In order to overcome the defects of the prior art, the invention aims to provide a real-time early warning method and system for the rear-end collision risk of a mobile maintenance vehicle based on a front time window, and solves the technical problems that the early warning threshold setting is solidified, the complex road working condition cannot be adapted, the track prediction model is easy to be interfered by the adjacent history noise to cause the extrapolation instability, and the early warning accuracy and the real-time performance are insufficient due to the lack of a multi-source uncertainty fusion mechanism in the prior art. In