An R-A dual network detection model for abnormal behavior of running vehicles

Abstract

Abstract Detection of abnormal behavior of running vehicles plays an essential role in road traffic safety. In this paper, the R-A (ResNet-Adaboost) dual network detection algorithm to cope with the problem that the existing detection algorithm of abnormal running behavior cannot detect different kinds of abnormal running behavior and cannot adapt to different detection scenarios. Firstly, this paper utilizes the YOLOV5-DEEPSORT algorithm to collect the spatial and temporal information of the target vehicle. In this way, various kinds of abnormal running behavior can be detected efficiently. Secondly, based on the information matrix, the entropy method is used to determine the dynamic weight of various abnormal running information, and then to determine the type of information input to the judgment model so as to obtain the information more accurately. Finally, the accurate vehicle running information is input into the R-A detection model. In this paper, an algorithm to estimate the abnormal running information is established as the information classification basis of the R-A detection model. The algorithm calculates the difference between the input vehicle running information matrix and the normal one to judge whether the information matrix is abnormal. Moreover, it is proved by the field experiments and NGSSIM datasets that the R-A detection model is able to detect different abnormal running behavior vehicle in different scenarios. The experimental results show that the R-A model with an accuracy of 90%-95% is better than the existing detection model. Additionally, it can more accurately detect various abnormal behavior of running vehicles.