skeletonbased

Skeletonbased refers to approaches that rely on a skeletal representation of the human body, typically a set of joints connected by bones, to model pose, motion, or action. In computer vision and biomechanics, skeleton-based methods use joint coordinates rather than raw pixel data, enabling robustness to clothing, lighting, and background variations.

Inputs are often 2D coordinates from monocular video or 3D coordinates from depth sensors or motion capture. The skeleton is commonly represented as a graph with joints as nodes and bones as edges, allowing models to capture spatial relationships and temporal dynamics across frames.

Modeling techniques include temporal methods such as recurrent neural networks and transformers, as well as graph neural networks that operate on the skeletal graph to learn joint interactions. Preprocessing steps frequently involve normalization, alignment to a canonical pose, and smoothing to reduce noise.

Applications of skeleton-based methods span action recognition, gesture understanding, sports analysis, rehabilitation monitoring, and human–computer interaction. Datasets used for evaluation include large 3D skeletal collections collected with depth sensors or motion capture, as well as public benchmarks designed for cross-subject and cross-view generalization. Evaluation typically reports accuracy or related metrics for recognition tasks.

Advantages of skeleton-based approaches include efficiency from compact representations and relative invariance to appearance, lighting, and background. Limitations involve dependence on accurate joint estimation, susceptibility to occlusion and missing data, and view-dependent distortions. Ongoing research aims to improve robustness through better joint estimation, multi-view fusion, data augmentation, and advanced graph-based modeling to capture complex spatio-temporal patterns.