On-Camera Content Filtering for Real-Time Video Analytics

In the past several years, pervasive surveillance cameras have generated massive video records continually, and the video records can be used for several applications (e.g., tracking and object detection). Machine Learning (ML), especially Deep Learning, is a powerful method extensively used for video analytics. Typically, cameras do not have enough computational power to analyze videos locally. A commonly used strategy is gathering videos from cameras and processing them in the cloud or edge server. However, limited bandwidth between cameras and the server results in the inefficiency of existing central-based video analytics systems. In this paper, we effectively reduce the bandwidth demands by proposing a novel object filtering method while keeping the resolution of the video. Specifically, we adopt the moving object detection algorithm on the camera side to find regions in the video and encode these regions into high-resolution video. Through extensive experiments, our system can reduce the bandwidth demands of the video to be analyzed by at least 36.2% without losing analysis for main objects.