Design Of Detection System For Abnormal Video Incident Based On Deep Learning

With the development of information technology,mankind has been carrying on the artificial intelligence from the information era.Various kinds of artificial intelligence technologies are or will change our lives,study and working.In the field of computer vision,driven by deep learning technology,the trend of rapid development in recent years has been shown.For example,the accuracy of face recognition of computer has exceeded the accuracy of human eye recognition on LFW datasets,and the technology of autonomous driving of unmanned vehicles jumps from Level 2,where drivers and machine share control of vehicle,to Level 5,full-automatic driving level.Humans simply input the place of departure and destination,and the rest of the work is handed over to the machine.Security surveillance field has also been promoted by the artificial intelligence technology.With the help of the application of dynamic facial recognition system,some stations realized face brushing and greatly improved the efficiency of security screening.Police captured a batch of' suspects through real-time face-control,which provided a new Ideas and tactics.This paper summarizes and analyzes the researches on the surveillance of anomalous events at home and abroad,and puts forward the feasibility of using deep learning technology in the surveillance of anomalous events.Due to the complexity of the definition of anomalous events,this dissertation starts with the basic visual tasks,and studies the algorithm of pedestrian tracking,face detection and pedestrian re-identification.Combining with the existing research results,a series of methods based on depth learning technology are proposed,which effectively overcome the shortcomings of the traditional program.The algorithm performance and efficiency have been effectively improved.The main research results of this paper are as follows:(1)A multi-domain target tracking algorithm based on deep learning is proposed.This algorithm first uses several video sequences labeled with a tracking target to learn a shared convolution network.On the stage of online tracking,a fully-connected layer for specified target is randomly initialized to learn the target-related features.Because of the strong ability of image feature extraction of deep learning network,the target and background have great degree of discrimination.At the same time,this paper uses the recurrent neural network to learn the motion model of the target,which makes a reliable judgment of the position information of the moving target.This algorithm is applied to the authoritative public dataset VTB for verification.Experiments show that this method can still achieve a good result even under the influence of external disturbances,such as illumination changes,scale changes,local occlusions and the deformation of the target itself.(2)A face detection algorithm based on convolutional recurrence network is proposed.Firstly,the history of face detection and the disadvantages of traditional face detection algorithms in feature representation and classifier are briefly described.Then,the face detection algorithm is designed using the deep learning network.Firstly,Gaussian mixture model is used to extract the background image and filters out irrelevant background interferences.Secondly,the feature of the image is extracted by residual network,called ResNet.Compared with the traditional texture features,the robustness of the feature is greatly improved.Using the Long Short-Term Memory network,we can learn the position and size of the face from the sequence.This method has a strong capability for detection of occlusion face,while taking advantage of the global face density information.In order to verify the validity of the algorithm,a qualitative and quantitative test of the system is conducted by using the Mall Street surveillance dataset of Chinese University of Hong Kong.The test performance is measured in terms of recall rate and accuracy rate.(3)A pedestrian re-identification algorithm based on triplet loss is proposed.The difficulty of pedestrian re-identification technology is discussed in depth.The use of distance measurement learning method makes the feature distance between the same individuals smaller and the distance between different individuals larger.Using recurrent neural network to study the spatial distribution characteristics of pedestrians,Making the expression of features more robust.In order to further verify the feasibility of the algorithm,we use the public CUKH03 dataset to train and evaluate the algorithm.The experimental results show that the proposed algorithm has excellent performance,in spite of that there are some difficulties such as light changes,posture changes,clothes fluttering and so on.