RA-Distill: Relational Alignment Distillation Based on Gram Matrix for Semantic Segmentation of Country Club Environments

Electric vehicles in leisure facilities pose significant safety concerns, demanding robust autonomous driving systems with precise visual perception algorithms. This paper introduces a novel knowledge distillation framework, relational alignment distillation (RA-Distill), for semantic segmentation of country club environments. The proposed method addresses critical challenges for achieving reliable accuracy in complex environments while ensuring computational efficiency for deployment on resource-limited hardware. RA-Distill extracts rich relational knowledge by computing Gram matrices from channel attention maps to analyze inter-channel correlations and global contexts. This structural information is then transferred from a complex teacher network to a lightweight student network using a similarity metric based on the centered kernel alignment for ensuring the invariance to scaling and orthogonal transformations. Experiments were conducted on a real-world country club dataset and the public CamVid dataset. The experimental results demonstrate that the proposed RA-Distill significantly outperforms previous distillation methods. Our lightweight student model surpasses the performance of the teacher network in the country club environments, enhancing the reliability of the collision avoidance system.