TRACKING AND ABNORMAL BEHAVIOR DETECTION IN VIDEO SURVEILLANCE USING OPTICAL FLOW AND NEURAL NETWORKS

Abstract

A target tracking algorithm that should also detect abnormal behavior detection is aimed to correctly identify the targets as being in a normal scenario or some chaotic movement. This dissertation is aimed for applications like highway traffic monitoring, railway stations, entry into restricted area etc. A model is developed here that presents the results in the form of normal and chaotic classes. The algorithm first extracts the information from the optical flow model using Lucas-Kanade approach. Optical flow model provides with the information of horizontal and vertical displacements of the objects of interest and the directions associated with each pixel. The features extracted from the model are then fed into neural network that is used for training as well as classification of the data. The uniqueness of this algorithm is that it uses foreground detection with Gaussian mixture model before passing the video frames to optical flow model. In this way the noise is being eliminated at the very initial stage and only the objects in motion are correctly identified. The study is being conducted on the real time videos taken from camera directly and some synthesized videos as well. The accuracy of method has been calculated by confusion matrix and mean square error of the neural network. The overall accuracy of the system is calculated as 97.5% and the percentage wrong classifications are 2.5%. The mean square error for the implementation was calculated as 3.5e-02.