Video Anomaly Detection Based On Deep Learning

Anomaly detection as a research challenge and key technology for intelligent video surveillance,the key issue is how to obtain a better feature representation of the object behavior.Compared with the traditional method,the biggest advantage of deep learning is that it can automatically learn useful feature from the massive data.So deep learning provides a good solution to the anomaly detection.In this dissertation,we briefly introduce the knowledge of anomaly detection,and analyze the anomaly detection models based on traditional methods and deep learning methods.Finally,we proposed an anomaly detection frame,and the main works are as follows.Firstly,we proposed an anomaly detection model based on Alex Net.First,the video frame is time-sampled to spatial-temporal block in which the features are extracted by Alex Net and classified by one-class SVMs.In classifiers training stage,in order to improve the detection rate of the model,multiple classifiers are trained in parallel。In the testing phase,based on the voting principle of "one person,one vote",we made the final judgment based on the voting.Several experiments were performed to verify the advantage of sampling one every some frames,and the optimal sampling interval was determined.Due to the different network depths different the ability to obtain feature data,the optimal network layer number of the model was determined by the traversal methodSecondly,we proposed an anomaly detection model based on Alex Net+SAE and context information.We adopted a new data processing model to obtain the spatial information and size context information of the image,and train a multi-layer sparse self-encoding network to reduce dimension of the feature data of the Alex Net network to reduce the training time and improving the detection rate of the classifier.Several experiments were done to verify the SAE's optimal hidden layer structure,which is used to detect anomalies.Experiments show that the proposed can achieve good detection results on the UCSD Ped2 dataset and the UMN dataset.