Computer Vision: Human Pose Tracking & Action Recognition

Track human pose across challenging video and classify actions using temporal motion features, deep learning, optimized inference, and deployment-ready pipelines.

Categories

Tech Used

Problem

Human activity recognition becomes difficult when videos contain fast movement, camera motion, occlusion, changing viewpoints, or noisy keypoint detections.

Approach

• Implemented and compared multiple pose-estimation backends for robust human keypoint tracking
• Smoothed temporal keypoint trajectories and engineered motion features from joint geometry and movement cues
• Built sequence-modeling and classification pipelines for recognizing actions from pose dynamics
• Prepared inference components for real-time or near-real-time application workflows
• Used deployment-oriented tooling including ONNX Runtime, TensorRT, FastAPI, Docker, and GPU acceleration

Results

• Improved stability of pose-based analysis under challenging video conditions
• Built a reusable foundation for action recognition, sports analytics, exercise analysis, ergonomics, and human-computer interaction
• Connected pose estimation and temporal classification into a product-oriented inference workflow

Demo Videos

Screenshots