Access Type

Open Access Dissertation

Date of Award

January 2015

Degree Type


Degree Name



Computer Science

First Advisor

Ming Dong


Object tracking in real scenes is an important problem in computer vision due to increasing usage of tracking systems day in and day out in various applications such as surveillance, security, monitoring and robotic vision. Object tracking is the process of locating objects of interest in every frame of video frames. Many systems have been proposed to address the tracking problem where the major challenges come from handling appearance variation during tracking caused by changing scale, pose, rotation, illumination and occlusion.

In this dissertation, we address these challenges by introducing several novel tracking techniques. First, we developed a multiple object tracking system that deals specially with occlusion issues. The system depends on our improved KLT tracker for accurate and robust tracking during partial occlusion. In full occlusion, we applied a Kalman filter to predict the object's new location and connect the trajectory parts.

Many tracking methods depend on a rectangle or an ellipse mask to segment and track objects. Typically, using a larger or smaller mask will lead to loss of tracked objects. Second, we present an object tracking system (SegTrack) that deals with partial and full occlusions by employing improved segmentation methods: mixture of Gaussians and a silhouette segmentation algorithm. For re-identification, one or more feature vectors for each tracked object are used

after target reappearing.

Third, we propose a novel Bayesian Hierarchical Appearance Model (BHAM) for robust object tracking. Our idea is to model the appearance of a target as combination of multiple appearance models, each covering the target appearance changes under a certain situation (e.g. view angle). In addition, we built an object tracking system by integrating BHAM with background subtraction and the KLT tracker for static camera videos. For moving camera videos, we applied BHAM to cluster negative and positive target instances.

As tracking accuracy depends mainly on finding good discriminative features to estimate the target location, finally, we propose to learn good features for generic object tracking using online convolutional neural networks (OCNN). In order to learn discriminative and stable features for tracking, we propose a novel object function to train OCNN by penalizing the feature variations in consecutive frames, and the tracker is built by integrating OCNN with a

color-based multi-appearance model.

Our experimental results on real-world videos show that our tracking systems have superior performance when compared with several state-of-the-art trackers. In the feature, we plan to apply the Bayesian Hierarchical Appearance Model (BHAM) for multiple objects tracking.