Research Of Abnormal Behavior Detection Based On Lightweight Model

Family meetings between prisoners and their families and communication with legal aid personnel play an important role in promoting active rehabilitation of prisoners.As a new way of meeting,remote video meeting has been paid more and more attention and become one of the important directions in the construction of intelligent prison.The traditional supervision method of meeting process has low intelligence level,large police input and high labor intensity.In this thesis,the abnormal behavior detection and early warning technology based on deep learning is studied for the application of remote video meeting,so as to improve the intelligence level of meeting process supervision and the work efficiency of police monitoring.According to the practical application requirements,a lightweight object detection and pose estimation model is constructed,and a fast and effective abnormal behavior detection algorithm is designed to realize automatic supervision and early warning of video meeting process.The main research work is as follows:1.Based on the object detection algorithm YOLO-V3,a human object detector with high recognition accuracy and real-time detection is constructed by improving and optimizing it.As YOLO-V3 is a relatively large model with high computing power requirements,the network should be reduced according to requirements.Firstly,the scaling factor y in the batch normalization layer is used as the index to carry out network sparsification.After the sparsification,network channel pruning is carried out.Since channel pruning would cause loss of network prediction accuracy,knowledge distillation method is adopted to compensate for the loss of accuracy caused by pruning with the help of high-performance network model.Finally,after pruning and knowledge distillation,the model size of yolo-v3 is reduced from the original 248MB to 24MB,which is one tenth of the original size.The recognition accuracy of the human category on MSCOCO 2017 data set achieved by the newly constructed is 68.6%.2.Based on the AlphaPose network model,a human pose detector with high recognition accuracy and real-time detection is constructed by improving and optimizing it.The AlphaPose network adopts a top-down detection method,which is composed of two parts.The first part is the human object detector,which is responsible for detecting human body,and the detected human target is sent to the second part by a single pose estimator to detect the coordinates of skeletal joints.In order to reduce the model size of AlphaPose network and meanwhile maintain high recognition accuracy,the pruned YOLO-V3 network is used instead of the original Faster-RCNN human target detector,and an efficient and lightweight backbone network Efficientnet-B0 is employed instead of the original Stacked Hourglass Networks single pose estimator.Finally,the newly constructed human pose detector achieves an average accuracy of 71.2%on MSCOCO 2017 pose estimation dataset.Compared with the mainstream models,the newly constructed human pose detector can reach the processing speed of 86FPS,which makes it possible to meet the project application.3.Based on the constructed real-time posture detection model,algorithms are designed to detect violent body movements and illegally leaving the supervision area,and thus realize the function of abnormal behavior detection and warning for video meeting.The proposed algorithms take the coordinate information of human body skeletal joints output by posture detection model as starting point,calculate the movement speed as the detection criterion,track the abnormal behaviors of the meeting participant in real time and give early warning.Experiment results on the constructed dataset show that the accuracy rate and recall rate achieved are 94%and 97.9%respectively.