SafeCommunity AI
A multi-channel intelligent video analytics system spanning tracking, rule-based alerts, anomaly recognition, and vision-language review.
- My role
- Core algorithms and system engineering
- Methods & tools
- PyTorch · VideoMAE · CLIP · Qwen-VL
01
Problem & challenge
Rule-only security systems can be brittle, generate false alerts, and provide little explanation. The project needed to connect long-video understanding, open-set events, and readable language feedback.
02
Key contributions
Combined VideoMAE and CLIP representations with Attention Pooling and Top-K MIL for key-segment aggregation.
Adapted a vision-language model with QLoRA and structured prompts for event review and explanation.
Built a web workflow for multi-stream video, ROI rules, tracking, and alert review.
03
Outcome & disclosure boundary
The system connects multi-stream inputs with alert review and result display. The public repository includes the architecture and implementation.
View on GitHub