Human Action Recognition System

Security and factory cameras record far more footage than anyone will ever watch. This system watches it as it comes in, recognises what people are doing and flags the moments that matter.

It works in three steps. YOLOv8 finds every person in each frame. ByteTrack keeps the same ID on each person from frame to frame, even when they are briefly hidden or move fast. MediaPipe then extracts each person's skeleton, and an ST-GCN model classifies the action from how the joints move over time.

Looking at joint movement over a short window, rather than one still image at a time, is what makes it reliable: it sees movement rather than a single pose.

The system recognises more than 20 actions that matter on site, such as falling, fighting, running, crouching, operating machinery, carrying objects and standing idle. Each one is tuned differently. Fall detection is set to never miss a fall, while fight detection is set to avoid false alarms.