Evaluating the impact of lossy compression on ADAS deep learning models using fisheye cameras

The increasing deployment of Advanced Driver Assistance Systems (ADAS) alongside the continual rise in camera sensor resolution has led to high bandwidth, and generally high cost, computation, and intra-vehicle communication. While the sensor bandwidth impacts the vehicle architecture, it also affects the data collection, storage, deep learning model training, and validation infrastructures. However, if the bandwidth was low, while still achieving the goal of high accuracy ADAS perception, the time and cost associated with creating and deploying the system would be greatly reduced. This study investigates the influence of lossy compression on multi-task deep learning models for real-time perception in ADAS employing fisheye cameras. We leverage a large-scale dataset and train a representative multi-task ADAS perception model for pedestrian, kerb, line, and soiling classification. The testing dataset is subjected to compression using the popular H.264 video codec at varying compression ratios. Through rigorous evaluation, we analyse the effects of compression on model performance, providing insights into the feasibility of employing lossy compression techniques in ADAS applications. Our results reveal that lossy compression could be deployed in automotive perception applications and that a compression ratio of up to 98% (720Mb/s to 12Mb/s), could be utilised with negligible performance degradation.