CompCAST: A System to Evaluate Offloading Strategies for Cloud-Assisted Autonomous Vehicles

Abstract

Cloud-assisted autonomous driving (AV) lets vehicles offload perception to edge or cloud servers, but transmitting high-resolution camera streams over wireless links introduces communication overhead that image and video compression must manage. Existing AV simulators offer limited support for jointly studying compression, offloading, and network conditions within one reproducible pipeline. We present CompCAST, a COMPression-aware Cloud-ASsisTed system built on CARLA and Pylot that integrates offloading-aware encoding, GPU-accelerated compression, cloud-based object detection, and realistic wireless network emulation. CompCAST's encoding strategy, combining sparse I-frames with adaptive P-frames, reduces encoded frame size by 40-50%, while fine-tuning the cloud detector on compressed inputs recovers up to 0.20 F2-score under heavy compression. Using KITTI with real 4G, 5G, and satellite traces, we show that H.265 and AV1 provide the best latency-accuracy trade-off, outperforming JPEG and H.264 with up to 45% and 34%, resp., while maintaining over 0.77 F2-score.