Facial Feature Extraction Algorithm Based on Improved YOLOv7-Tiny

Facial feature extraction is a critical step in driver fatigue detection, essential for improving driving safety. This paper proposes a novel driver facial feature extraction algorithm based on YOLOv7-Tiny to address the challenges of low precision and practical deployment in existing fatigue detection systems. The algorithm employs depthwise-separable convolution combined with spatial depth convolution to effectively extract key facial features. An attention module is integrated between the backbone and neck networks to enhance contextual understanding and filter out irrelevant information, enabling precise extraction of detailed and distinguishable features, even under varying lighting conditions and driver poses. Additionally, a feature fusion module is introduced to merge features from different receptive fields, improving multi-scale feature extraction and reducing the miss rate for small objects. The proposed algorithm achieves a detection accuracy of 59.8% mAP on the Drowsy-Driving-Det dataset, marking an 8.5% improvement over the original method, alongside a 60.9% reduction in model parameters. This improved algorithm not only meets real-time deployment requirements but also maintains high detection accuracy, making it well-suited for facial fatigue feature extraction in complex driving environments and edge deployment scenarios.