Video Object Tracking Using Particle Filter

Abstract

In this thesis, the problem of video object tracking is addressed in real world scenario. newlineThe problem is challenging because of various complexities of the scene such as illumination newlinevariation, dynamic entities in background, bad weather conditions, static and dynamic newlineshadows. The problem is further compounded because of camouflage, partial occlusion newlineof the object and camera jitter. In this research, attempts have been made to take care newlineof some of the above issues while tracking moving object in a scene. Three particle filter newlinebased novel schemes are proposed for tracking single as well as multiple objects in the real newlineworld scene with varying degree of complexities. newlineThe first particle filter based scheme employs the proposed notion of Feature Fusion newlineusing Template Partition (FFUTP) which focuses on target object modeling. The object newlinetemplate is partitioned into different parts and, in each partition Local Binary Pattern newline(LBP) and mean RGB color features are fused in probabilistic framework to generate newlinethe appropriate fused feature distribution to model the scene in the given partition. The newlineweights for feature fusion are determined by taking care of local scene dynamics. The newlinemodel of the entire template is the mean of the fused distributions of individual template newlinepartitions. In the tracking process, the initial target model is used throughout the video newlinesequences. A set of particles is generated using this template and these particles are newlinepropagated through a dynamic state model to take care of the nonlinear movements and newlinescene uncertainties and in turn help in tracking the object in the subsequent frame. It is newlinefound that appropriate target model improves the tracking accuracy. The proposed scheme newlineis successfully tested on different videos from LASIESTA, DAVIS 2016, CDnet 2012 and newlineCDnet 2014 datasets. The performance of the proposed scheme is found to be superior as newlinecompared to Color model, LBP-Color model, Discriminating Coefficient (DC) model and newlineEdge Oriented Histogram and HSV (EOH-HSV) target model based approa