Anomaly Motion Detection and Tracking for Real-Time Security System

Almost every corner of the world is under surveillance. However, crimes (anomalies) can occur anywhere and anytime. Duration of anomalous activities is for a very short time as compared to normal activities. Advancements in computer vision have proved Convolutional Neural Networks (CNN) to be a compelling solution for anomaly detection in videos. This paper proposes to bring together different methods for effective anomaly detection. The dependency of Machine Learning on massive sets of the hand-labeled training dataset is time-consuming. To avoid this, a predictive model is constructed that learns with weak supervision using Multiple Instance Learning (MIL) approaches. The combination of motion information (optical flow) along with spatial information (RGB) obtained from 3-dimensional Convolutional Networks (C3D), using 2-stream architecture, offers a stronger visual representation of videos, thus enhancing anomaly detection performance. TVL1 is used to obtain the optical flow representation of videos. After detection, suspicious behaviors in videos are tracked on the UCF Crime dataset videos.