Artificial Intelligence Based Region of Interest Enhanced Video Compression
Palanivel Guruvareddiar, Praveen Prasad · 2020
Artificial Intelligence, especially deep learning based workloads for the analysis of video data are on the rise. Examples include residential and commercial security systems where the camera data are analyzed for potential intruders, in-door retail cameras that track people movement for behavior analysis etc. In an End-to-End intelligent video solution, video analytics will be carried out at multiple places from edge IP camera to high performance cloud servers. Typically, the results of the video analytics at one stage will be sent to the next stage along with the video data for efficient processing. However, a traditional video encoder does not have an understanding of the scene and/or the priority of the objects in the frame. It tries to compress the frames with a goal to produce visually pleasing video for the human viewers by maximizing the rate-distortion performance. As a result, in the compressed video frame, the details of the regions of interest to the inferencing engine may be lost due to the artifacts introduced by the lossy compression process. We propose a novel AI based low-power architecture that utilizes the results of the machine learning at the given stage to improve the performance of the overall pipeline. The proposed method will identify the regions of interest in the frame and preserve the quality of those regions without significantly reducing the quality of the remaining regions. Simulation results show that the machine vision accuracy when inferred on the compressed video streams using the proposed method are significantly better compared to the inferencing results on the compressed video streams produced by the traditional video compression methods.