Combining dynamic split convolutions and lightweight inverse residual module for human pose estimation
Li Wang, Rihan Gu · 2024
Human pose estimation has been increasing in importance in fields such as animation design, security monitoring, and motion analysis. However, current human pose estimation algorithms focus on accuracy, resulting in complex networks and high computational costs, making it difficult to be applied on mobile devices and embedded platforms. To alleviate this problem, this paper proposes L-HRNet, a lightweight human posture estimation network that combines dynamic split convolution and lightweight inverse residual module. firstly, the bottleneck layer DKASCneck of the high-resolution network is re-designed by using dynamic split convolution and dynamic kernel aggregation operation, which avoids the computational cost of using a large convolution kernel and at the same time enhances the network’s ability to extract useful features; the high-resolution network is better utilized to extract useful features; and the high-resolution network is more efficient in terms of computation cost. The high-resolution network can retain the spatial location information well, and in order to further reduce the number of model parameters, a lightweight inverted residual module is proposed. Finally using this network, the human body pose can be estimated efficiently. Compared with the benchmark network, the model achieves 73.9% and 89.7% accuracy on the COCO and MPII datasets.