| Illegal overturning is a common phenomenon in all social situations,which not only causes serious negative impact on society,but also may cause serious property damage or personal injury.With the rapid development of artificial intelligence and other technologies,surveillance systems are gradually becoming integrated,networked and digitalised,and intelligent security surveillance systems have become a research hotspot of great interest.This paper proposes to introduce spatio-temporal behaviour detection technology in the field of behaviour recognition into security surveillance systems to detect abnormal overturning behaviour,based on the YOLOv5s target detection model and a spatio-temporal behaviour detection model combining Faster R-CNN and SlowFast networks,with the main research content as follows:(1)Spatio-temporal behaviour detection technology contains two modules,target detection and behaviour recognition,and this paper addresses the two modules separately,starting with pedestrian detection technology research.The YOLOv5s network is improved to solve the problems of low resolution of video images,large scale of human target frames in different postures and partial occlusion of overturned objects in surveillance video scenes.The SPD-Conv module is used to replace the stepwise convolution and pooling layers of the original network to avoid the loss of fine-grained image information;the adaptive fusion module(ASFF)is introduced in the Neck part to improve the scale invariance of the network;the Grid Mask data enhancement strategy is introduced to enhance the network’s attention to the salient features of video images under the occlusion of fence rules;finally,the improved network is proposed The accuracy of the proposed network model YOLOv5s＿SA is experimentally demonstrated to be 90.5%on the home-made dataset,and the visualisation analysis further demonstrates that the model is more suitable for assisting the behaviour recognition module to complete the detection task.(2)The spatio-temporal behaviour detection model is studied to identify four types of behaviours that will occur during overtaking:walk,climb,fall and stand.In order to improve the detection accuracy,firstly,the target detection network Faster R-CNN in the spatio-temporal behaviour detection model is replaced with YOLOv5s＿SA;secondly,human behaviour is rich in diversity and the contextual information of the same behaviour often differs,making the network highly susceptible to false detection.In order to improve the recognition accuracy of the SlowFast network for behavioural recognition,this paper introduces the ACTION spatio-temporal attention module in the Fast branch of the SlowFast network and the SE spatial attention module in the Slow branch,and proposes the SlowFast_AS model to improve the network’s learning ability of spatio-temporal features and the processing ability of key information overall.It is experimentally demonstrated that the recognition accuracy of the improved model SlowFast_AS is improved by 11.08%compared with the original SlowFast network model.(3)Combined with spatio-temporal convolutional networks to design an abnormal overturning behaviour detection system,the discriminant algorithm processes the recognition results obtained from the behaviour recognition network to detect continuous overturning behaviour,and finally tests are carried out on a test set of overturning videos taken,and the system is experimentally proven to have good detection effect in detecting abnormal overturning behaviour. |
