Research On Safety Warning Technology Of Expressway Tunnel

With the rapid development of highway tunnels in China,their accident multiplicity and severity have a profound impact on society,and the research on tunnel safety early warning technology is imminent.With the rapid development of deep learning technology,many studies have proposed the use of target detection algorithms and target tracking algorithms to achieve tunnel safety warning,but such algorithms are always unable to achieve a comprehensive and continuous dynamic tunnel target due to the limited camera view.sensing.In this paper,we propose a high-speed tunnel safety warning scheme using cross-camera multi-target tracking algorithm,and design an AI chip-based hardware backplane to realize a complete safety warning system,which can not only locate risky targets in real time,but also complete cross-camera tracking and display of tunnel vehicles.This paper first introduces the knowledge related to cross-camera multi-target tracking algorithm.Then the system implementation framework is proposed,including both hardware and algorithm parts.The hardware part selects Horizon X3 M as the core board and focuses on the bottom board circuit design,mainly including data transmission and communication modules;the algorithm framework includes three modules of risky target detection,single-shot vehicle tracking and vehicle trajectory fusion,and introduces the algorithm implementation and board-end deployment process in detail.Next,we focus on the design and implementation of the algorithm.In the detection algorithm part,a fast method of producing data sets is proposed to solve the problem of time-consuming and laborious traditional manual labeling methods;in addition,to improve the detection performance at the edge end,the edge loss function of the lightweight target detection algorithm is improved and the angle loss is introduced to improve the robustness of the model.In the tracking algorithm,the focus is on modeling the vehicle appearance model and adding center loss to compensate for the effect of intra-class differences;and the generalized intersection ratio method is used in the matching algorithm to obtain a better intersection metric of matching frames.In the part of trajectory fusion algorithm,a lane based temporal masking algorithm is proposed to filter the vehicle trajectories,and a reordering algorithm is designed to optimize the vehicle appearance similarity matrix,and finally a hierarchical clustering algorithm is used to achieve cross-shot trajectory unification.The paper concludes with a verification of the system functionality and a comparison of the algorithm performance by module.The experimental results show that the recognition accuracy of the improved target detection algorithm reaches 93.86%,which is 6.71% higher compared with the original algorithm;the re-recognition accuracy of the constructed vehicle appearance feature rerecognition model reaches 99.49% after optimization from three aspects of learning rate strategy,network structure and loss function,which is 2.88% higher compared with the original algorithm;the tracking accuracy of the vehicle tracking algorithm is 3.3% higher after optimization;The track fusion algorithm is improved to improve the matching accuracy by more than 22%,and the clustering evaluation index is improved to more than 0.8;and the accuracy and feasibility of this paper’s algorithm deployment at the edge-end system level is verified by the actual highway tunnel test video.