Publication

Abstract : Autonomous Mobility (AM) is the next milestone in the history of the automotive industry. Radar, lidar, cameras, ultrasonic, and other sensor units provide the required information for highly automated driving and pave the way for AM. This popular and fast-developing technology aims to increase vehicle safety through various sensors and advanced-level Artificial Intelligence (AI)- based algorithms. The camera is one of the vital components of AM systems, which captures real-time videos, processes, encodes, packetizes and transfers them to different processing units for further applications. Video compression plays a crucial role in video data compression for multi-camera systems and this technique needs to be perception-aware. The capabilities of video compression are improved by perception awareness, but it is complicated and requires a method to obtain optimal values for its parameters. This paper proposes a framework for semantic-aware video compression with reduced functional complexity and a proper flow for quantization parameter optimization. The optimized quantization factors used for encoding the region of interest and the region out of interest of the input video frame. The proposed framework was evaluated with H.264, HEVC, VP8, and AV1 video compression on the public cityscape dataset. Our experimental results demonstrated that our proposed framework with the AV1 codec achieved the highest efficiency (compression ratio of 12:1). HEVC produced the highest quality results, with a notable improvement in both SSIM (0.9981) and PSNR (53.5656). The fastest execution speeds were achieved by H.264 (CPU: 237 ms, GPU: 65) and VP8 (CPU: 246 ms, GPU: 64).