Research And Application Of Abnormal Event Detection Method Based On Video Sequence

The vehicle collision accident in the surveillance video is an important type of abnormal event.In particular,a certain proportion of casualties are caused by delayed treatment and secondary accidents.This is because the rescue organization and the vehicles around the accident cannot get a quick response to the accident.Therefore,it is crucial to develop an effective accident detection method,which can greatly reduce the number of deaths and injuries,as well as the impact and severity of accidents.At present,the detection speed of methods for detecting accidents in videos is relatively slow,which leads to low practical application value.Based on this,this article focuses on vehicle collision accidents in surveillance videos.This article first introduces the relevant theoretical knowledge about neurons,convolutional neural networks and YOLO series algorithms,and then improves the existing RFBNet network to better detect the targets in the accident,compared with the original YOLOv3,the improved YOLOv3 has better detection performance,and finally the accident in the video is detected by the combination of the improved YOLOv3 and the mean shift algorithm.The main contributions of this paper are as follows:1.If you want to detect abnormal accidents in the video,you must detect the target in the video.Only by accurately detecting the target can we better locate and detect abnormal accidents.If the structure of the target detection model is complex,it will affect the detection effect of the accident.Therefore,in view of the large amount of redundancy in the structure of the deep learning target detection algorithm,this paper proposes a pruning method based on the scaling factor to simplify the model.First,the neural network is trained on channel sparsity.The training process also trains the network weights,scaling factors,and biases.Sort the values of the scaling factors in the network structure,determine the pruning threshold according to the pruning ratio,and then pruning the network structure according to the threshold.2.RFBNet has achieved excellent performance in target detection,but the detection speed is not ideal,which is caused by the RFB module in the network and the excessive parameters of its backbone network.Aiming at the problem of excessive redundancy of RFBNet algorithm parameters,this paper simplifies the model by reducing the number of channels of the RFBNet algorithm backbone network and simplifying the RFB structure.At the same time,the OSA module is introduced and densely connected to enhance the feature extraction capability of the backbone network,and CIOU loss is used to replace the original smooth L1 loss of the RFBNet network to avoid errors caused by the inconsistency of loss functions and evaluation indicators.In the simplified model,the parameter amount is reduced by 41.6%,the speed is increased by 28.9%,and the accuracy remains basically unchanged.3.The YOLOv3 detection model has good detection performance,but the model structure is more complex,which affects the performance of the model to a certain extent.Aiming at the complex network structure of the YOLOV3 network,the pruning strategy in this article considers the pruning of all convolutional layers..Use the pruning mask to merge the shortcut layer,and use the pruning mask to connect the routing layer.Through the experiment of pruning method in this article,a higher pruning rate can be obtained.After pruning,the parameter amount of the YOLOv3 model becomes 11% of the original,which greatly reduces the parameters and calculations,but the accuracy remains almost unchanged from the original.4.This paper uses the pruned YOLOv3 algorithm to detect vehicles.By using the improved YOLOv3 and the traditional target tracking algorithm mean shift algorithm,the vehicle position and acceleration related parameters are extracted,and then the vehicle collision accident is identified according to the extracted parameters.