Research Advanced in Human Pose Estimation based on Deep Learning

Human pose estimation constitutes a crucial task within the realm of computer vision. Its objective is to precisely identify and locate the main features of the human body as shown in images or videos, thereby enabling the understanding and analysis of human behavior. In the early stages, pose estimation efforts predominantly focused on single-person scenes, relying mainly on template matching, direct regression, and detection methods. Thanks to the rapid progress of deep learning technologies, research on human pose estimation has evolved from single-person scenarios to complex multi-person ones. Multi-person pose estimation can be separated into top-down and bottom-up approaches. The top-down method initially detects the human body and then conducts pose estimation. Although relatively straightforward to implement, it has limitations when dealing with multi-person scenes and occlusion situations. In contrast, the bottom-up method first detects main features and then assigns them to different individuals, proving more effective in handling occlusion and multi-person scenes. Considering the above factors, this paper undertakes a comprehensive and in-depth study on human pose estimation based on deep learning, covering multiple aspects such as single-person pose estimation, multi-person pose estimation, datasets, evaluation metrics, algorithms, and experimental methods. Additionally, it discusses the existing problems and future development directions from perspectives including improving the model architecture, optimizing data processing, and integrating multi-modal information. Moreover, recent advancements in deep learning architectures, including convolutional neural networks and recurrent neural networks, have further enhanced the performance and accuracy of human pose estimation. The combination of different modalities, such as depth information and infrared imaging, also holds promise for addressing some of the challenges in complex scenes.