A Lightweight and Robust Spatio-Temporal Deep Learning Framework for Multi-Class Driver Emotion Recognition Under Challenging Illumination and Image Degradation

Introduction Road accidents often happen due to driver tiredness, stress, or lack of attention. Facial expression recognition helps to understand a driver’s emotional state by analyzing face movements such as eyes and mouth. However, real driving conditions include poor lighting, blur, and noise, which make recognition difficult. This research focuses on developing a reliable system to detect driver facial expressions accurately in real-world situations. Materials and Methods This study uses the Driver Attention Dataset, which contains real driving images under different lighting, blur, and noise conditions. Images are first preprocessed using skin segmentation to remove background noise. A deep learning model combining CNN and BiLSTM is then used. The CNN extracts facial features, while BiLSTM analyzes changes across video frames to understand facial expressions over time. Results The proposed CNN-BiLSTM model achieved very high performance in recognizing ten driver facial expressions. The system reached about 99% accuracy with high precision and recall for all emotion classes. It performed well even under blur, noise, and low-light conditions. Compared to existing models like DSFD, RetinaFace, and YOLO, the proposed approach showed better accuracy and more stable results. Discussion and Conclusion The results show that combining spatial features and temporal information improves facial expression recognition in challenging driving environments. Skin segmentation helped the model focus only on facial regions, reducing errors caused by lighting and noise. The proposed method is suitable for real-time driver monitoring systems and can improve road safety. Future work can include testing larger datasets and integrating additional vehicle safety features.