An Improved Lightweight Human Pose Estimation Method in Video

Human pose estimation in video confronts the challenges with complex models, large amount of parameters, and high computational complexity. This paper proposes a lightweight human pose estimation method in video that fuses spatial features and time constraints. The spatial feature network is based on the modified HRNet incorporating lightweight modules. The time constraint network models contextual dependencies with time constraints, and processes temporal information through the Long Short-Term Memory units to increase the accuracy of pose estimation. Furthermore, the Atrous Spatial Pyramid Pooling module is utilized to integrate time series feature maps on different scales to achieve stable sequence output. The experimental results on two classic video datasets demonstrate that the model can reduce the parameters and calculations while enhancing the accuracy effectively. The ablation experiment indicates the ability of the time constraint module to improve the performance of pose estimation.