• anomalies in videos
• anomaly detection
• computer vision
• deep learning
• frame prediction
• generative models
• one-class classification

Anomaly detection consists in finding events or items which vary from the normality.
It can be a useful tool to reduce or simplify the work that humans have to do, increasing productivity and reducing errors and costs. A suitable candidate to build an automatic anomaly detection system is Deep Convolutional Neural Network (CNN) that has proven effective in Computer Vision tasks. The major feature of NNs is that they can learn from examples, with no need for any previous expertise or knowledge. This can be a useful feature in anomaly detection in which events could be unknown because of the sporadic nature of anomalies or challenging to represent.
A viable approach is to train a model to learn how normality appears in an unsupervised way and considers all the events different from it as abnormal. The main advantage of this approach is that the system can be trained using an exhaustive dataset of normality examples. Most of the recent researches on anomaly detection, and also this work, go in this direction.
This work aims to implement a frame prediction based anomaly detection system that performs as well as other state-of-the-art approaches and tests its discrimination capabilities on several types of anomalies. A custom dataset has been created to remedy the lack of some types of anomalies in publicly available datasets and tests the proposed solution with a wider range of anomalies.