Efficient-MobileNet: A Feature Fusion Approach for Vehicle Driver Behavior Detection and Classification

Driver behavior detection is a crucial aspect of enhancing road safety and preventing accidents by identifying and classifying driving actions in real-time. This paper presents "Efficient-MobileNet," a deep learning-based model for automated driver behavior detection. The proposed model employs a feature fusion approach, leveraging EfficientNetB7 for feature extraction from input data, combined with Atrous Spatial Pyramid Pooling (ASPP), attention blocks, and residual blocks to enhance the model's performance. ASPP captures multi-scale contextual information, critical for detecting driver actions of varying sizes and positions within an image or video frame. Attention blocks improve the model's focus by emphasizing relevant features, such as hand movements or facial expressions, while ignoring irrelevant background details. Residual blocks address the vanishing gradient problem, enabling the model to learn complex behavior patterns by maintaining information flow through identity mappings. The fused features are passed through MobileNetV3 for final classification, resulting in a robust and accurate system capable of reliably classifying both safe and risky driving behaviors. Extensive experiments on a real-world dataset demonstrate the effectiveness of the proposed model, outperforming existing methods in terms of accuracy, precision, recall, f1 score, and AUC of 99.62%, 99.91%, 99.96%, 99.93%, and 1.00% respectively.