Abnormal Behavior Detection Based On Riding Trajectory

With the common use of shared bicycles,chauffeur-driven electric bicycles and other bicycles,urban residents are experiencing the convenience of travel while facing a large number of traffic accidents caused by riding anomalies such as illegal steering and retrograde.Road traffic safety is a huge concern,and there is an urgent need to design appropriate riding anomaly detection methods to help traffic management departments to provide real-time warning of riding anomalies.The positioning devices equipped with various types of bicycles continuously generate a huge scale of location data in the process of riding,and they have the characteristics of wide coverage,low cost,real-time,and massive,etc.By analyzing them efficiently,the movement behavior pattern and trend of cyclists can be extracted to ensure the timely detection of abnormal riding events.However,there is a lack of research in academia and industry on using riding trajectory data for anomaly detection,and the existing trajectory anomaly detection methods only focus on the identification of trajectory anomalies of vehicles such as cabs and online vehicles,and cannot be directly used to deal with the riding trajectory data that turn more randomly and frequently.Riding anomalies mainly include two types of irregular steering anomalies and reverse driving,and they differ greatly in both the location and angle where the steering occurs.In addition,the urban cycling road network often undergoes changes in road topology,and the above problems bring huge challenges for abnormality detection using cycling trajectory data:(1)due to sampling errors and missing samples of positioning equipment,and the missing position correction strategy that comes with the equipment,resulting in data quality problems such as noisy points,drifting trajectory segments and missing trajectory segments in cycling trajectory data,how to effectively identify noise,correct drifting How to effectively identify noise,correct drift segments and recover missing segments becomes the first difficulty to be solved for accurate detection of abnormal riding?(2)the complex and variable steering position of riding vehicles and the differential steering frequency on different roads bring the second challenge for the detection of cyclists’ irregular steering?(3)the topological changes of riding road network caused by road closure etc.occur from time to time,bringing the third challenge for correct identification of reverse riding behavior.In response to the above challenges,this paper designs a three-stage riding data calibration method based on trajectory clustering for riding anomaly detection based on real riding trajectory data from Shanghai and Xiamen cities,based on which a steering violation detection method combined with an attention mechanism and a real-time retrograde identification method based on gated recurrent network are designed.(1)Calibration of cycling data based on trajectory clustering: The quality problems of cycling trajectory data are analyzed in depth around two types of cycling abnormal behaviors,such as illegal steering and reverse driving? then,on the basis of systematically sorting out the shortcomings of existing trajectory data calibration methods,a three-stage data calibration mechanism is proposed for the quality problems of cycling data,such as noise points,drifting trajectory segments and missing trajectory segments,including abnormal trajectory point elimination,drifting trajectory segment correction and missing trajectory segment recovery.The effectiveness and robustness of the proposed data calibration method are verified by experiments based on real cycling trajectory data,in which the accurate calibration length is improved by 6.85%.(2)Combined attention mechanism for steering violation detection:A steering violation detection method combining graph attention network and self-attention mechanism is designed for the influence of cyclists’ complex and variable steering positions and steering angles on steering violation detection.In the construction of the steering violation detection model,the graph attention network is introduced to extract the spatial features affecting the steering violation,and the trajectory sequence features are combined with the self-attention mechanism.Finally,an extensive experimental study is carried out on real riding trajectory data,and the experimental results show that the proposed method improves the F-score by 9.98% in detecting riding steering violations compared with the comparison method.(3)Reverse ride recognition based on gated recurrent network: For the problem of main direction change of cycling lane caused by cycling road topology change,a spatio-temporal feature extraction method of main direction is proposed by fusing graph attention network and gated cyclic unit,and then a detection model of main direction change of cycling lane is constructed to realize lane direction calibration based on main direction change? subsequently,an online map matching of cycling trajectory data is used to realize the sliding window model to Real-time detection of retrograde section range.Experiments based on real cycling trajectory data verified that the proposed retrograde detection method F-score improved by 7.63% compared with existing methods.In summary,this paper aims at detecting abnormal riding behavior,and first designs a data calibration method including abnormal trajectory point elimination,drift trajectory segment correction and missing trajectory segment recovery,then proposes a riding violation steering detection method by fusing graph attention network and self-attention mechanism,and designs a reverse riding detection method including riding lane main direction calibration and sliding window model based.Finally,the efficient detection performance of the proposed method is verified through extensive comparison experiments with real cycling trajectory data in two cities,Shanghai and Xiamen.