A Novel Approach of Driver Facial Expression Recognition Based on Improved Swin Transformer

Driver facial expression recognition (FER) is a fundamental yet challenging task for the perception system of intelligent vehicles. It is difficult for the current recognition model to achieve high accuracy and fast inference. To overcome the above issue, we propose a novel driver facial expression recognition method based on an improved Swin Transformer. Firstly, shallow features of facial images are extracted through CNNs to reduce the patch of image sequences and lower the computational complexity of self-attention calculations. Then, the residual coordinate attention mechanism is designed to focus on crucial facial features and suppress environmental disturbance. Finally, the structural re-parameterization method is employed to accelerate the model while maintaining recognition accuracy. Experiments are conducted on three widely used expression datasets, and the results demonstrate that our approach is competitive to several state-of-the-art methods in FER tasks. Industrial experimental results evidence the practicality of the RCST, and the overall accuracy reaches $95.04 \%$ on the HDER dataset.