This repository contains Jupyter notebooks for fine-tuning and evaluating the MoViNet model on the UCF Crime dataset using frame-level labels. The project is part of the Sentinel system — an edge-powered, real-time surveillance anomaly detection framework designed for seamless integration with existing IP camera networks.
sentinel_data_loading.ipynb: Downloads, extracts, and splits the UCF Crime dataset into train/test directories.sentinel_model_training.ipynb: Fine-tunes the MoViNet A0 model using frame-level labels.sentinel_model_evaluation.ipynb: Evaluates model performance on the test set.sentinel_model_testing.ipynb: Tests the trained model on a random video and outputs anomaly scores per frame.
We fine-tune the MoViNet A0 model, a memory-efficient and mobile-friendly video classification architecture, on frame-level annotations from the UCF Crime dataset. MoViNet’s design includes:
- Depthwise separable convolutions
- Bottleneck layers
- Temporal convolutions
- Global average pooling
These characteristics make MoViNet ideal for edge deployment with minimal computational cost.
We use the UCF Crime Dataset, a large-scale real-world surveillance video dataset containing:
- 1900+ untrimmed videos
- 13 anomalous activity categories (e.g., Fighting, Arson, Burglary, Abuse, etc.)
- Roughly equal distribution of anomalous and normal videos
The original UCF Crime dataset provides weak video-level labels, meaning that anomalous events may only appear in a small portion of each labeled anomalous video.
To enable fully supervised training, we incorporated frame-level annotations from this external resource. These detailed labels allowed us to train the MoViNet model using precise frame-wise anomaly information, significantly improving the model's learning and detection performance.
- Confusion Matrix: Low false positives and negatives with an accuracy of 81%.
- ROC-AUC Score: 89%, which is competitive with the highest reported results on the UCF Crime dataset.
More information and visual results are available on our Sentinel Project Page.
- Deployed on NVIDIA Jetson Nano
- Performs local inference, logs anomalies, records flagged video clips
- Supports on-demand live streaming from edge to mobile
- Register and link edge devices
- Receive push notifications for anomaly alerts
- Access live camera feed remotely
- Receives alerts from edge devices
- Stores anomaly logs and associated footage
- Relays push notifications to mobile devices
- Streams live video on request
