Anomaly Detection in Video Sequences using Multiple Object Tracking and Autoencoder
Preeti Sharma, Mandlem Gangadharappa · International Conference on Computing for Sustainable Global Development · 2021
Identifying anomaly in video sequences is a major topic in computer vision community. Therefore, it is important to automate the camera system for understanding video activities. We used a semi-supervised learning method by extracting features such as the pixel value of the video frame and the coordinates of the bounding boxes of objects which further uses multiple object tracking systems along with an autoencoder model. While tracking, the stationary object pixels which are treated as fixed pixels start crossing the boundary of video frame then those points are considered as an anomaly. We use the autoencoder model which gives a score of reconstruction error automatically through the difference between the input value and previously constructed value. The experiments are performed on real-time video covering the objects having normal and abnormal activities. The observed results have shown that the reconstruction error of static object moving abruptly and touching boundary is higher than that of expected motion of instance and are considered to be anomalous. This shows the effectiveness of the proposed method.