Aegis-Sentinel-Net
Violence detection model for video, built on a ResNet50 + Temporal Attention MLP architecture, with activation gated by a pose-based contact trigger to reduce false positives.
Description
Aegis-Sentinel-Net analyzes video sequences to identify violent behavior. Instead of running inference continuously on every frame, the pipeline uses a lightweight pose model (YOLOv11-nano-pose) to detect physical contact between people (IoU >= 0.40 between person bounding boxes). Only when contact is detected does the temporal classification model (ResNet50 + Attention MLP) get triggered. This design reduces computational cost and false positive rate compared to a continuous inference approach.
Architecture
- Backbone: ResNet50 (ImageNet-pretrained initialization)
- Temporal head: Attention MLP over per-frame extracted features
- Trigger model: YOLOv11-nano-pose, runs on every frame for person detection and contact estimation
- Activation threshold: IoU >= 0.40 between detected person bounding boxes
Training data
- Approximately 4,632 videos
- Input tensors: (16, 3, 224, 224), float16
- Normalization: standard ImageNet
Metrics
| Metric | Value |
|---|
| F1 (validation) | 0.9391 |
| Accuracy (validation) | 93.96% |
| Best checkpoint epoch | 10 |
Model format
- Trained checkpoint in PyTorch
- Also exported to ONNX (approximately 102.9 MB)
Intended use
This model was developed for research and technical portfolio purposes within the Aegis project ecosystem, focused on computer vision applied to industrial and public safety. It has not been validated for deployment in critical production environments without additional evaluation, bias auditing, and robustness testing under real-world camera, lighting, and crowd density conditions.
Known limitations
- The model was trained and validated on a moderately sized dataset (~4,632 videos); generalization to scenarios, camera angles, or populations not represented in the training data is not guaranteed.
- The contact trigger (IoU >= 0.40) may not capture forms of violence without direct physical contact (for example, use of weapons at a distance, verbal aggression, or violence involving thrown objects).
- The model's behavior across different demographic conditions has not been formally evaluated, which is relevant for any real-world deployment oriented toward safety.
License
This model is distributed under the CC BY-NC-ND 4.0 license (Attribution - NonCommercial - NoDerivatives).
This means:
- Use and reference of the model is permitted with credit to the author.
- No commercial use is permitted under any circumstances.
- No derivative works are permitted: this includes fine-tuning, retraining, extracting weights to incorporate into another model, or any modification of the distributed checkpoint.
- The model must be used as published, strictly for academic and non-commercial research purposes.
Any use outside these terms, including commercial use or adaptation of the model, requires explicit written authorization from the author.
Author
Developed by Beau as part of the Aegis portfolio of computer vision projects applied to safety.
