Abnormal Behavior Detection Technology Based On Deep Learning And Its Application

With the improvement of the computing abilities and the increasing abundance of datasets,deep learning has achieved great success in image classification,semantic segmentation,target detection in Computer Version.In the task of Video Action Recognition,some mainstream convolutional neural networks have obtained good performance.However,these commonly used network structures are not good at learning the context information and visual properties in the videos.As a result,these deep learning algorithms have the disadvantage of modeling in area or information that has important influence on video images.Besides,the time of action taking place in the video is not invariant.Therefore,how to effectively put network attention on the action area is another problem need to be solved in video activity recognition.Thus,this thesis analyzes and researches several problems appearing in video-based human behavior recognition.The main contents and results of thesis are as following.Firstly,in order to identify abnormal behaviors in the video accurately,this thesis proposes a modeling based on generative adversarial networks.This modeling first uses a generator to predict future frames.Then it compares prediction frames with real frames by using a discriminator.In this way,whether there are abnormal behaviors can be determined according to the results of the comparison.Secondly,since there are much difficulties in training of generative adversarial networks and generated images sometimes are seriously distorted,the generation of the image would become meaningless.What's worse,model crashes might take place.Therefore this thesis chooses Wasserstein GAN.In addition,due to the inherent disadvantage of the original generative adversarial networks,model can't achieve very good results.In this way,this thesis especially adopts the features of conditional generative adversarial nets to constraint motion information of models by introducing optical flow.As for extraction of optical flow features,Flownet Network that has good performance at present to extract optical flow of scenes in video has been chosen.Then combine improved generative adversarial networks with variational automatic encoder and utilize variational automatic encoder to conduct joint training with VAE and GAN.Better results were obtained by giving full play to their advantages in feature extraction and generation performance.Besides human control has been a reality by introducing features of conditional generative adversarial Nets.Last but not the least,this thesis analyzes the realization of main function modules video analysis system and conducts its functional test.With analysis of the results,video analysis system can satisfy the needs of abnormal behavior detection,which has great value in both the theory and practice.