Task-Oriented Deep Joint Source-Channel Coding with Semantic-Aware Adaptive Quantization for Autonomous Driving

In autonomous driving and intelligent transportation, vehicles need to share visual perception information to extend sensing range, but conventional separate source-channel coding can suffer from the cliff effect under harsh vehicular channels, causing downstream perception failures. This paper proposes an end-to-end semantic communication system for autonomous-driving semantic segmentation. The system adopts SegFormer-B2 as the semantic encoder, performs feature modulation conditioned on the signal-to-noise ratio SNR, applies semantic-aware adaptive bit-width quantization based on Gumbel-Softmax, and uses a task-oriented decoder to directly produce segmentation maps. Experiments on Cityscapes evaluate the system under additive white Gaussian noise (AWGN) and per-channel independent block Rayleigh fading channels, with comparisons against a JPEG-based separate baseline and fixed-bit-width variants. Under AWGN, the proposed system maintains a mean intersection over union above 0.70 at SNR=−6dB, while the conventional baseline fails at SNR=0dB with an mIoU of 0.02, indicating improved robustness against the cliff effect. The results also show a non-monotonic relationship between quantization precision and segmentation performance: fixed 4-bit quantization can outperform fixed 6-bit quantization because finer quantization is more sensitive to channel noise. Under Rayleigh fading, adaptive quantization may suffer from semantic-channel mismatch, providing useful observations for future channel-aware semantic resource allocation.