Abstract: In recent years,with the rapid development of smart phones,head-bowing pedestrians keep browsing mobile phones when they cross the road.As a result,they often cause the traffic accidents.How to effectively detect the bow-headed people has become an urgent problem.The existing detection methods not only require a large number of datasets about real pedestrians using mobile phones,but also have the problems of low recognition accuracy and unsatisfactory speed.Considering these problems,this paper proposes a fast and effective method to detect head-bowing abnormal pedestrians.This method is based on joint points instead of images,therefore it is different from existing methods.Firstly,we propose a novel method to construct a synthetic dataset,which solves the problem of lacking dataset and corresponding labels.We leverage human joint points by adjusting the coordinates of the left and right hand to simulate the posture of device-holding.Secondly,we achieve the abnormal behavior recognition and classification by training network on our synthetic dataset.Additionally,we make full use of arm and head information to achieve a precise pedestrian abnormal behavior detection.Finally,experiment results show that the proposed method can achieve real-time detection,and the detection accuracy reaches 94.08%.Therefore,it can provide necessary reference information for video surveillance,drivers,assisted driving,and autonomous driving systems.

Key words: Abnormal detection, Bowed heads detection, Data fitting, Human joint points