This repository belongs to a team of 5 students from BITS-Pilani. The project was done as a part of the PS1 program, at CEERI-Pilani
This project aims to use unsupervised deep learning methods , to detect anomalies in videos. Our model architecture is based on 3D Convolutional - Generative Adversarial Networks (#D-GANs). It consists of 2 parts, an autoencoder, and a discriminator. The model is trained adversarially, the autoencoder is trained to reconstruct videos well, and the discriminator is trained to classify real videos, and regenerated ones (from the autoencoder) , accurately.
As the autoencoder is trained on normal videos, it is unable to reconstruct videos containing anomalies accurately, and thus, the discriminator will be able to classify the video containing anomalies as a regenerated video.
We trained and tested the model using the UCF-Crime dataset.
Link for Kaggle Notebook: https://www.kaggle.com/abhinavlalwani/deep-learning-for-anomaly-detection-in-videos