Object Tracking And Abnormal Loitering Detection Based On Deep Learning

Computer and artificial intelligence is playing an increasingly important role with the rapid development of information society today. Video surveillance system is being increasingly improved and larger owing to intensive distribution of cameras, which plays an irreplaceable role in the maintenance of social stability. The general researching path of intelligent video surveillance technology is object tracking and anomaly detection. Great quantities of approaches proved to be effective have been proposed on this researching path after recent years of development. However, achieving a robust and real-time tracker based on video object tracking and anomaly detection is still a challenging and promising problem due to the complexity of the surveillance environment. Moreover, the algorithm needs to overcome a variety of interference so that numerous algorithms for intelligent monitoring have drawn many researchers’ attention.Based on the research of existing algorithms of object tracking in video sequence and analysis of deep learning’s successful application in the field of image processing and after thoroughly studied on the object tracking algorithms on the basis of deep learning, this dissertation intensively investigated an effective algorithm of object tracking based on convolution neural network using pre-trained features combined with mainstream object tracking framework, which also has been applied to the field of intelligent video surveillance. In addition, abnormal loitering detection framework was proposed in the dissertation, including discrimination of the abnormal behavior and detection and identification of suspicious objects with multi-camera. The major achievements of the dissertation are as follows:Firstly, proposed a new algorithm of object tracking in video sequence based on deep learning, including Particle Filter, pre-trained features, convolution neural network, discrimination classifier, online parameters updating, etc. The mainstream object tracking method and framework was combined with deep learning model closely and deeply in the proposed approach, which has achieved more prominent performance and strong robustness in actual use, available to adapt to multiple interferences in the process of tracking and changeability of tracked object itself.Secondly, proposed an abnormal loitering detection method based on object tracking framework on the basis of deep learning, making an effective attempt to apply the method to the field of intelligent video surveillance. Abnormal loitering detection method proposed in the dissertation can accurately and efficiently find the suspicious objects of video surveillance. Therefore, the algorithm achieved detection and identification of suspicious objects with multi-camera within the feature extraction capability based on deep learning model.