Posture positioning estimation for players based on attention mechanism and hierarchical context

As a basic method to study human motion, pose estimation has become a research hotspot in the field of computer vision. Its main task is to detect the coordinate positions of human joints and key parts in the image, so as to obtain partial or all limb information of the human body, so as to judge the behavior. For the issue of complicated background, non-rigid changes in posture and low efficiency of posture estimation model in athletes' posture positioning estimation methods, this work proposes a new posture positioning strategy. Based on the attention mechanism, this work designs a large receptive field hourglass attention network and a large receptive field residual module to improve the traditional residual module. The large receptive field residual module expands the effective receptive field area, which can enable the model to more effectively use image's multi-scale information, and improves the accuracy and robustness of attitude estimation. Second, an efficient human pose estimation model framework is proposed based on the context learning. The key point range area and the specific local area are, respectively, input into the model, and layer-by-layer prediction is performed, which realizes that the model can be efficiently trained and deployed while retaining strong generalization capabilities. The large receptive field hourglass attention network is used as the stage backbone network of the hierarchical context network to achieve a balance between model accuracy and efficiency. Finally, comprehensive and systematic experiments are carried out to verify the superiority and feasibility of the method designed in this work.

Automated summary

This paper proposes a new posture positioning strategy based on attention mechanism and context learning, including a large receptive field residual module and an efficient human pose estimation model framework. By designing a large receptive field network and incorporating context information, the accuracy and robustness of the model are improved while maintaining efficiency.