Semantic-Aware Video Compression for Automotive Cameras

Assisted and automated driving functions in vehicles exploit sensor data to build situational awareness, however, the data amount required by these functions might exceed the bandwidth of current wired vehicle communication networks. Consequently, sensor data reduction, and automotive camera video compression need investigation. However, conventional video compression schemes, such as H.264 and H.265, have been mainly optimised for human vision. In this paper, we propose a semantic-aware (SA) video compression (SAC) framework that compresses separately and simultaneously region-of-interest and region-out-of-interest of automotive camera video frames, before transmitting them to processing unit(s), where the data are used for perception tasks, such as object detection, semantic segmentation, etc. Using our newly proposed technique, the region-of-interest (ROI), encapsulating most of the road stakeholders, retains higher quality using lower compression ratio. The experimental results show that under the same overall compression ratio, our proposed SAC scheme maintains a similar or better image quality, measured accordingly to traditional metrics and to our newly proposed semantic-aware metrics. The newly proposed metrics, namely SA-PSNR, SA-SSIM, and iIoU, give more emphasis to ROI quality, which has an immediate impact on the planning and decisions of assisted and automated driving functions. Using our SA-X264 compression, SA-PSNR and SA-SSIM have an increase of 2.864 and 0.008 respectively compared to traditional H.264, with higher ROI quality and the same compression ratio. Finally, a segmentation-based perception algorithm has been used to compare reconstructed frames, demonstrating a 2.7% mIOU improvement, when using the proposed SAC method versus traditional compression techniques.