ARGO

Inspiration:

criminal activities in metro and train stations

What it does:

ARGO detects violence,armed violence and snatching

How we built it:

I fine-tuned YOLO26s-pose for pose estimation, using TrackTrack as the tracker. A separate YOLO26s model handles weapon detection, and an ST-GCN classifies actions from the tracked skeletons. Almost all training ran on Kaggle’s free GPUs, using public, homemade, and self annotated datasets.

Challenges we ran into:

The biggest challenge was the lack of training data. The model is already strong at spotting violence, but snatching needs more training. The second was the low resolution of CCTV cameras, which initially gave near zero accuracy on weapon detection.Weapon detection accuracy is now around 20/25%, limited by the small dataset.

Accomplishments that we're proud of:

I'm proud of the violence detection that hits around 90% accuracy

What we learned:

This was my first project training a model on video. I learned to work around hardware limitations using only free alternatives

What's next for ARGO:

I plan to replace ST-GCN with a graph transformer and train the model on more snatching and CCTV weapon detection data.

Built With

• claudecode
• kaggle
• numpy
• python
• pytorch