skeleton-based action recognition

**Skeleton-based action recognition** is the **approach that models human actions from body joint coordinates instead of raw RGB pixels** - by focusing on articulated pose dynamics, it becomes robust to background clutter, lighting changes, and appearance variation. **What Is Skeleton-Based Recognition?** - **Definition**: Action classification from 2D or 3D keypoint sequences representing body joints over time. - **Input Structure**: Graph-like skeleton with joints as nodes and bones as edges. - **Temporal Signal**: Motion trajectory of joints carries action semantics. - **Typical Models**: Spatial-temporal graph convolution networks and transformer variants. **Why Skeleton-Based Methods Matter** - **Appearance Invariance**: Less sensitive to color, texture, and scene distractions. - **Data Efficiency**: Compact pose representation lowers input dimensionality. - **Interpretability**: Joint trajectories are easier to inspect than latent pixel features. - **Cross-Domain Robustness**: Better transfer across camera and illumination conditions. - **Realtime Potential**: Lightweight models can run efficiently on edge hardware. **Core Modeling Components** **Pose Extraction**: - Detect keypoints with human pose estimator. - Track joints across time with identity consistency. **Graph Temporal Encoding**: - Apply graph convolution across body topology. - Apply temporal convolution or attention across frame sequence. **Action Classification Head**: - Aggregate graph features and output action probabilities. - Optional multi-person interaction modeling. **How It Works** **Step 1**: - Convert video to sequence of skeleton graphs and normalize joint coordinates. - Build adjacency matrix for body structure and temporal links. **Step 2**: - Encode spatial-temporal graph features and classify action with supervised loss. - Evaluate with top-k accuracy and robustness across viewpoints. **Tools & Platforms** - **OpenPose and pose estimators**: Keypoint extraction front-end. - **ST-GCN frameworks**: Graph-based action recognition baselines. - **Edge deployment runtimes**: Efficient inference for low-power systems. Skeleton-based action recognition is **a pose-centric pathway that captures motion intent while ignoring irrelevant visual noise** - it is a practical solution when robustness and interpretability are priorities.