Human posture estimation in intelligent video surveillance systems: modern approaches and challenges

Abstract

The article presents an overview of modern methods for estimating human pose in intelligent video surveillance systems. 2D and 3D approaches are considered, including both classical methods and models based on deep learning. The features of top-down and bottom-up strategies, their advantages and limitations are analyzed. The main challenges associated with real-world surveillance conditions, such as occlusions, lighting changes, and limited image quality, are identified. Prospects for the development of the industry, including the use of 6D pose estimation, multisensory data, and self-learning models, are outlined.