Research And Application Of Pedestrian Abnormal Behavior Detection At Traffic Intersections Based On Human Pose Estimation

As our economy continues to modernise and our road system continues to improve,the level of urban modernisation is increasing,and with it,pedestrian traffic safety accidents.Unauthorised red light jumping,traffic rule violations and sudden abnormal pedestrian behaviour at urban traffic junctions can all lead to traffic accidents and personal and property safety issues.As Internet,Internet of Things and Artificial Intelligence technologies continue to innovate,it has become an inevitable trend to combine these technologies with traffic monitoring systems.Currently,there are two problems in applying deep learning models to traffic intersection scenarios:(1)the computational volume of deep learning models is too large and training data sets are scarce;(2)the lack of hardware acceleration devices in the surveillance system at traffic intersections limits the computational power and makes it difficult for deep learning models to run in real time.A new lightweight human pose estimation network SA-Tiny HRNet is proposed to address the problem that the human pose estimation network is difficult to be applied to devices with limited computing power in practice.In this paper,the lightweight module Shuffle Block,the conditional channel weighting method and the Coord Attention attention mechanism are combined to build a new lightweight module SA-Block(Shuffle AttentionBlock),and the module is applied to the lightweight network Small-HRNet to achieve a lightweight human pose estimation algorithm.The estimation algorithm is lightweight.Through experimental evaluation and visual analysis,SA-Tiny HRNet achieves m AP detection accuracy comparable to that of HRNet,DARK and Simple Baseline at a cost of 0.27 G parametric count and 1.13 M computation,while the computation and parametric count of the network are reduced by about 90% year-on-year.In response to the large number of parameters and computational effort of current action recognition networks,and the insufficient processing of long-time action timing,the AttentionBi LSTM action recognition network is proposed to achieve accurate recognition of five common actions: fall,walk,run,stand and fist in traffic intersection scenarios.In this paper,the sparse temporal sampling method is used to construct action timing features for human actions in videos,and the attention mechanism is incorporated into the Bi LSTM network,allowing the network to assign more weights to focus on the key timing features of human actions.Through evaluation experiments and visual analysis of the dataset,it is demonstrated that Attention-Bi LSTM achieves an accuracy of 85% for the recognition of five human actions and an overall recognition accuracy of about 89%,enabling the accurate detection of five common human actions at traffic intersections.The joint abnormal behaviour detection model is constructed to address the complexity of abnormal behaviour determination of pedestrians at traffic intersections.Firstly,the lightweight network YOLOv5-Lite is used to achieve dynamic recognition of pedestrians,red and green traffic lights and zebra crossings in real traffic scenarios by constructing a traffic intersection dataset;secondly,the Kalman filter algorithm is applied to achieve abnormal behaviour detection and trajectory tracking of pedestrians entering zebra crossings under red traffic lights;finally,SA-Tiny HRNet,Finally,SA-Tiny HRNet,Attention-Bi LSTM and YOLOv5-Lite are combined to achieve real-time pedestrian abnormal behaviour detection at traffic intersections.Through model complexity test experiments,video inference speed test experiments and visualization analysis,the joint model with 4.06 M computation and 1.59 G number of parameters has a real-time inference speed of 17～28FPS on the constrained device,realizing the effective detection and real-time operation of pedestrian abnormal behavior at traffic intersections.