Lightweight Human Key Point Detection Model Based on MobileNetV3
Chao Ma, Mingkun Zhang, Zhaoyang Zhao, Chunlei Zhang, Jian-wei Ma · IEICE Transactions on Information and Systems · 2025
Aiming at the problems of a large number of network parameters and high computational complexity of human key point detection models in the fields of human-machine collaboration and action recognition, a lightweight human key point detection model was proposed. First, we fine-tune MobileNetV3 as the backbone network to reduce the overall parameter amount of the model; Second, we propose a Feature Pyramid Network (FPN) and an adaptive upsampling module to improve the model detection accuracy with the introduction of a small number of parameters; Third, we combine the joint length information loss function with the MSE loss function to accelerate the convergence of the model during training to improve the prediction accuracy. Ablation experiments verified the effectiveness of each module. The experimental results on the COCO2017 dataset showed that the average accuracy of the algorithm proposed in this paper is 68.6%, the number of model parameters is 2.2M, the computational complexity is 1.8G, and the inference speed on a single CPU platform reaches 36 frames/s, meeting the real-time requirements.