Segmentation, Recognition And Tracking Of Humans In Surveillance Videos

With a very security conscious society, semi-automated video surveillance systems have become center of focus for many important applications for private, industrial, government and military needs. The research interest in video based object detection, identification, human tracking and action recognition has considerably increased recently, due to its importance in security applications and surveillance. It has therefore got tremendous focus from industry and the academia.The main idea of this research work has been to tackle some fundamental problems in background modeling for video based surveillance systems, segmentation of different moving objects, recognition of valid pedestrians and their tracking in dynamic environment in an efficient way as they walk around with partial occlusion. This work covers the development of an improved framework of surveillance system with optimized algorithms and its performance testing on various publically available surveillance video datasets. The algorithm has been divided into sub modules which are optimized to perform different tasks related to the overall application.Accurate segmentation of foreground objects in any surveillance video field from noisy dynamic background is necessary for efficiency and accuracy of subsequent enhancement, feature extraction and recognition algorithms. As a first contribution to this research, we have put forward a novel and robust method of detection and segmentation of moving object in surveillance videos with a two level hierarchical segmentation which uses moving average background model combined with a second refining model and multiple adaptive thresholds based on Gaussian distributions of the component intensities. The multiple threshold adaptation method is used to simultaneously update the proposed system to environmental changes hence making it sufficiently robust. This method is tested on various environments and experimental results show that proposed method is more robust and efficient than others popular video-based object extraction techniques.As a second contribution to this thesis, we have proposed a simple and cost effective algorithm for the detection of valid moving humans which is an essential process of automated video surveillance systems. Five body parts based human template have been used along with a simple, robust and computationally efficient feature set. The features used are concise, invariant to pose and transformation. Human recognition has been performed by FF Neural Network trained by feature vectors obtained from two different kind of human models. The proposed algorithm method is tested on various environments and experimental results show that this human recognition method is more robust and efficient than similar kind previous work using different feature set. The third contribution of the thesis addresses the development of part human body parts based tracking algorithm for efficient tracking of fully visible and partially occluded humans in video frames. Tracking has been performed by well known Kalman Filter to estimate the target pedestrian position in each next frame using a novel feature set.All these proposed algorithms formulate an effective and robust system for moving human/object detection, classification and visual tracking including partial occlusion. The overall proposed application is tested by detailed experiments with comparison of the results with existing methods.