Edge-Computing-Enabled Abnormal Activity Recognition for Visual Surveillance

Abstract

Due to the ever increasing number of closed circuit television (CCTV) cameras worldwide, it is the need of the hour to automate the screening of video content. Still, the majority of video content is manually screened to detect some anomalous incidence or activity. Automatic abnormal event detection such as theft, burglary, or accidents may be helpful in many situations. However, there are significant difficulties in processing video data acquired by several cameras at a central location, such as bandwidth, latency, large computing resource needs, and so on. To address this issue, an edge-based visual surveillance technique has been implemented, in which video analytics are performed on the edge nodes to detect aberrant incidents in the video stream. Various deep learning models were trained to distinguish 13 different categories of aberrant incidences in video. A customized Bi-LSTM model outperforms existing cutting-edge approaches. This approach is used on edge nodes to process video locally. The user can receive analytics reports and notifications. The experimental findings suggest that the proposed system is appropriate for visual surveillance with increased accuracy and lower cost and processing resources.