Motion-Augmented Inference and Joint Kernels in Structured Learning for Object Tracking

Video object tracking is a fundamental task of continuously following an object of interest in a video sequence. It has attracted considerable attention in both academia and industry due to its diverse applications, such as in automated video surveillance, augmented and virtual reality, medical, automated vehicle navigation and tracking, and smart devices. Challenges in video object tracking arise from occlusion, deformation, background clutter, illumination variation, fast object motion, scale variation, low resolution, rotation, out-of-view, and motion blur. Object tracking remains, therefore, as an active research field. This thesis explores improving object tracking by employing 1) advanced techniques in machine learning theory to account for intrinsic changes in the object appearance under those challenging conditions, and 2) object segmentation. More specifically, we propose a fast and competitive method for object tracking by modeling target dynamics as a random stochastic process, and using structured support vector machines. First, we predict target dynamics by harmonic means and particle filter in which we exploit kernel machines to derive a new entropy based observation likelihood distribution. Second, we employ online structured support vector machines to model object appearance, where we analyze responses of several kernel functions for various feature descriptors and study how such kernels can be optimally combined to formulate a single joint kernel function. During learning, we develop a probability formulation to determine model updates and use sequential minimal optimization-step to solve the structured optimization problem. We gain efficiency improvements in the proposed object tracking by 1) exploiting particle filter for sampling the search space instead of commonly adopted dense sampling strategies, and 2) introducing a motion-augmented regularization term during inference to constrain the output search space. We then extend our baseline tracker to detect tracking failures or inaccuracies and reinitialize itself when needed. To that end, we integrate object segmentation into tracking. First, we use binary support vector machines to develop a technique to detect tracking failures (or inaccuracies) by monitoring internal variables of our baseline tracker. We leverage learned examples from our baseline tracker to train the employed binary support vector machines. Second, we propose an automated method to re-initialize the tracker to recover from tracking failures by integrating an active contour based object segmentation and using particle filter to sample bounding boxes for segmentation. Through extensive experiments on standard video datasets, we subjectively and objectively demonstrate that both our baseline and extended methods strongly compete against state-of-the-art object tracking methods on challenging video conditions.