This project explores computer vision techniques for identifying violent actions based on posture and motion cues. Using YOLOv3 for person detection and a light-weight posture analysis pipeline, the model achieved approximately 90% accuracy on a curated dataset. The work culminated in a published research paper.

Highlights:

• Real-time inference with GPU acceleration.
• Modular design for model swaps and experimentation.
• Clear evaluation metrics and ablation studies.

If you want this entry to link to a live demo or repository, update the link or repo fields in the frontmatter.