Developing Video Surveillance Methods for Efficient Monitoring Analysis and Behavior Recognition

Abstract

Surveillance systems using camera for recording and monitoring activities are becoming ubiquitous in daily life. The increase in surveillance applications and rate at which videos are being generated have accelerated demand for automated analysis of videos to enable better identification of events. Analysis of human activity is one of the key aspects to recognize events in surveillance videos. In this thesis, we propose methods to monitor and analyze human activities in surveillance videos. newlineIn this thesis, we propose three major contributions towards the design of better surveillance systems. In the first approach, Motion History Image is used to extract information and augmentation of motion features is performed by moment and gradient based descriptors. Further, SVM is used to classify the activities in surveillance video. Effectiveness of the activity recognition system is mostly dependent on the efficiency of low level feature descriptors being used to encode the information representing any action. In the second approach, we propose a framework based on two novel feature descriptors to encode valuable information of motion image using segmented block and distance mean histogram of oriented gradients. The video is divided into frames and preprocessing is performed to aggregate information of three consecutive frames into one frame. The resultant motion rich frame is highly compressed which reflects the action in a dense representation. The proposed descriptors are calculated to transform resultant information of frame into quantitative values. We make use of random forest to predict the most likely class. In the third approach, we propose a framework for real time activity recognition based on root of sum of square for preprocessing and histogram of oriented gradients. The concept of parallel threading is used to achieve real time processing of framework. We have demonstrated our results on various video datasets such as ATM, HMDB and achieved good accuracy in comparison to the wellknown approach

Description

Keywords

Citation

item.page.endorsement

item.page.review

item.page.supplemented

item.page.referenced