Split Computing in Autonomous Mobility for Efficient Semantic Segmentation using Transformers

With the advancement of deep learning in Connected Autonomous Vehicles (CAVs), real-time semantic segmentation has emerged as a crucial task. The integration of semantic segmentation vision transformers into CAV perception systems capitalizes on the efficiency of the vision transformers and their capacity to capture a global context. To further optimize transformer’s performance, we propose the integration of a Split Computing (SC) into the transformer architecture to enable the processing of the computation related to transformers to be distributed between the CAVs and an Edge Computing Server (ECS). The objective of our work is to evaluate SC efficiency in minimizing latency of intensive computational tasks of the transformers with minimal loss in semantic segmentation’s accuracy. Simulations demonstrate that vision transformer architectures are well-suited for the SC integration and outperform both ECSOnly and CAV-Only baseline approaches in various scenarios, offering an acceptable latency-accuracy trade-offs. Specifically, our findings indicate that SC outperforms baseline methods in terms of latency by up to 79.36%, incurring a minor accuracy reduction. This makes them particularly suitable for applications where real-time processing and minimal latency are critical considerations.