Anomaly Activity Detection on Live Camera Feed via Machine Learning and Statistical Analysis.

Abstract

Anomaly detection is important in security camera systems in order to differentiate normal and suspicious activities. In this research, a method is proposeci to detect anomaly activities in a live video feed using machine learning algorithms based on a statistical analysis of the video. The developed system is based on the detection of moving objects along the vista. In addition, it's very important to track down the behaviour of the object in time. In order to achieve these aspects, a statistical analysis of the video which gives a lot of information about the pixel variations with the time is used. The videos can be modelled into a variation of pixel intensities and time, using a mathematical model. Once these variations are tracked, KNN algorithm is used to predict whether there is a malignant activity which should be notified. For the basic prototype, the machine learning problem is solved as a supervised problem having a large data set of possible activities. This approach can be used in unsupervised models too. In this research, the proposed method is applied and tested to grayscale videos and this can be extended to colour videos too