Enhanced Traffic Sign Recognition via RWKV with Deformable Attention

Autonomous vehicle systems incorporate traffic sign recognition as a foundational building block, with high accuracy and computational requirements to provide real-time performance in diverse and adverse conditions. We present here RWKV with Deformable Attention, a new vision encoder that marries the performance of RWKV with deformable attention's quality of association with diverse input features. RWKV with Deformable Attention easily achieves the linear computational cost of global attention calculation and local and global features needed for traffic sign recognition. Experimentation with the German Traffic Sign Recognition Benchmark (GTSRB) dataset shows state-of-the-art accuracy, precision, recall, and F1 score concerning baseline Vision Transformer-Tiny (ViT-T) and Convolutional Neural Networks (CNN) architectures. An overall observation of efficiency also shows that WKV with Deformable Attention outperforms transformers by as much as 12 times on high-resolution processing, proving its scalability and feasibility in a real-world deployment. Our experiments establish RWKV with Deformable Attention as an efficient, robust, and highly accurate system in traffic sign recognition, setting a new baseline in its space.