Deep Learning-Based Video Compression for Surveillance Footage

Abstract

The utilization of closed-circuit television (CCTV) monitoring plays a pivotal role in the realm of video processing, providing an effective means for vigilant surveillance. However, a significant challenge associated with this practice is the substantial demand it places on storage space. Traditionally, surveillance footage is stored on hard disk drives, and due to limited storage capacities, it often necessitates periodic deletion. To tackle this issue, we have introduced an innovative method for compressing CCTV video, named “Detection-Based video Compression” (DBVC). Our DBVC model is a two-step process. In the first step, we employ advanced neural network approaches such as Mask-RCNN and YOLOv4 to determine active and idle frames within the surveillance video. These cutting-edge techniques enable precise identification of objects and events of interest in the video feed. In the second step, we construct a new video composed solely of the active frames, eliminating redundant or uneventful segments. After conducting a comprehensive analysis of the experimental results, it is evident that Mask-RCNN stands out with an impressive detection accuracy of 98% on the COCO dataset, making it a robust choice for identifying objects and events in the surveillance footage. Consequently, we chose to leverage the output generated by Mask-RCNN and YOLOv4 for subsequent processing stages. Our DBVC approach is a breakthrough in video compression technology, significantly reducing the storage space required for CCTV footage. In fact, it achieves an average compression ratio of up to 85% when using YOLOv4, surpassing the capabilities of existing state-of-the-art compression methods. This innovation not only optimizes storage efficiency but also maintains a high level of surveillance data integrity, making it a valuable advancement in the field of CCTV video processing and storage management.