An Efficient Deep Learning Model for Simultaneous Detection of Human Body, Hands, and Face

Abstract

Accurate and efficient detection of multi-scale human body parts in a single pass, particularly the body, hands, and face, is essential for real-world applications such as pedestrian safety systems and fine-grained human-computer interaction. Existing detectors typically rely on multi-model pipelines or heavyweight architectures, resulting in increased complexity and limited real-time performance on edge devices. This study proposes a deep learning-based simultaneous human-body, hand, and face detection approach by modifying the YOLO11-nano architecture to obtain a more lightweight and efficient model on the Human Parts dataset. The proposed modifications focus on reducing the number of channels in the backbone and neck, thereby achieving an approximately 10% reduction, decreasing 2,590,425 to 2,330,687 of the model parameters and the computational complexity from 6.4 GFLOPs to 6.2 GFLOPs. Experimental results demonstrate that the proposed model achieves 93.5% mAP@0.5 and with 17.19 frame per second, while maintaining comparable detection accuracy to one-stage detectors. These results indicate a more favorable accuracy and efficiency trade-off, making the proposed model suitable for deployment in safety-critical and interactive real-world applications with limited computational resources.