A Traffic Anomaly Detection and Identification Approach Based on Multi-instance Learning

Abstract

Traffic anomaly detection plays an important role in effective prevention and timely handling of traffic accidents. However, currently traffic anomaly detection is still in its infancy, mainly depending on manual intervention, which not only consumes a lot of manpower, but also is unfavorable in timeliness. According to the characteristics of urban road traffic scenes, this paper proposes a MFnet network structure based on multiple fiber modules, aiming to realize fast extraction and real-time calculation of video streaming features via group convolution and sparse connectivity. In addition, the weakly-supervised multi-instance learning method is introduced for the training on application of traffic anomaly detection model, which reduces the difficulty of labeling sample videos and improves the capacity of traffic anomaly detection in complex scenarios. Experimental results based on real traffic video data show that this method proposed herein, compared to the existing traffic anomaly detection methods, is good in terms of detection accuracy and recall rate, and is efficient in traffic anomaly detection in actual traffic scenarios.