A Semantic-Aware Compression Strategy for Intelligent Vehicles

Autonomous Intelligent vehicles (AIV) utilize multiple sensors like cameras, radars, LiDAR’s etc. Vision sensors, i.e., cameras, are essential to Autonomous Mobility (AM) functions. Cameras are used to capture the real-time surroundings of vehicles. Vehicle networks need to handle the enormous volume of video data from multi-cameras. Hence, video compression plays a significant role in handling larger volumes of video data. The popular H.264 video compression format is used for vehicle video data compression due to its operational flexibility. Video compression quality is important for computer-vision applications and AM functionalities to maintain its operational accuracy. The advancement of computer-vision based deep learning approaches brings Semantic-Aware Video Compression (SAVC). In this method, video compression is achieved based on the Region Of Interest (RoI) and Region-Out-of-Interest (non-RoI) by applying different quantization factors. In this paper, the video compression is based on RoI and non-RoI classification at the pixel level of semantic segmentation. A parameter-tuning approach has been proposed to obtain better results. With this strategy, the optimal value of the quantization factor is selected for SAVC. The proposed method shows a significant improvement of 2.07% in Peak Signal-to-Noise Ratio (PSNR) over the existing methods.