Developing new algorithm for anomaly detection in surveillance videos

Abstract

Intelligent video surveillance has received more attention in recent years due to global security concerns. The main objective is to identify or recognize the anomalies present in the videos. In this thesis, an efficient Video Anomaly Detection (VAD) system is presented. It uses Multiple Feature Modalities (MFM) to represent a piece of a rectangular region of predefined size in a video frame called as a patch with Hybrid Classification (HC) using Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) classifiers for anomaly detection. The evaluation is done on a set of video sequences from University of California San Diego (UCSD) database which contains anomalies in a pedestrian pathway. As the name MFM implies, three different types of features; raw pixel values, gradient map, and texture energy map are extracted for anomaly detection effectively. At first, the motion estimated frame by background subtraction is partitioned into patches of predefined size. Three different sizes of patches; 15x15, 30x30 and 45x45 are used in this study. From the patches, features are extracted and combined linearly to form the feature vector of the corresponding patch. As SVM and GMM classifiers have their advantages and demerits, an effective VAD system is designed by combing these classifiers using a weighted voting method. The performance of the VAD system is measured in terms of Rate of Detection (RoD), Equal Error rate (EER), and Area Under the Curve (AUC). Initially, the performance of VAD system is analyzed without feature selection by GMM classification newline