Road Accident Detection From Surveillance Videos
BKU Team 2018
An implementation and a modified version of Real-world Anomaly Detection in Surveillance Videos (Sultani, Waqas and Chen) on Road_Accident dataset.
Dataset
Road accident dataset consists of 796 videos under *.mp4 format (330 normal, 366 abnormal, 100 testing).
- Dataset link: updating
- C3D Extractor: Learning Spatiotemporal Features with 3D Convolutional Networks (Du Tran et al.).
- Extract C3D feature of video using Google Colab (this jupyter notebook)
Follow the instruction in the notebook to extract video feature.
Training
Check this notebook Train_Test_Code to see the documentation as well as training/testing process.
- Keras 1.1.0
- Theano 0.9.0
- Python 3
Visualize the results
Django web application. See WebApp directory for more details.
File structure
| File/Directory | Decscription |
|---|---|
| C3D | Extract C3D video feature |
| Scripts | Python, Matlab ultility scripts |
| Temporal Annotation | Groudtruth annotation of testing videos |
| Makefile.config | Configuration file to build C3D Caffe model |
| Train/Test Code | Jupyter notebook for Traning/Testing process |
If you find any bug, or have some questions, feel free to contact any of these: Bien Do ([email protected]), Hoai Do ([email protected]), Dat Nguyen ([email protected]).
References
[1] W. Sultani, C. Chen, and M. Shah, โReal-world anomaly detection in surveillance videos,โ in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2018.
[2] D. Tran, L. Bourdev, R. Fergus, et al., โLearning spatiotemporal features with 3d convolutional networks,โ in The IEEE International Conference on Computer Vision (ICCV), Dec. 2015 .