Object Tracking And Localization In Video

Object localization and tracking is the basis of video analysis. Although this topic has been studied for dozens of years. There are still various difficulties when it faces the complicated real world,which mainly results from pose variations,illumination variations, occlusion, fast motion et.al. In this thesis, we explore the topic on object tracking and localization, and propose three novel algorithms.We propose an adaptive compressive tracking via online vector boosting. First,the most discriminative feature template are selected from positive and negative samples via an online vector boosting method. And the object representation is updated in an effective online manner. Furthermore, a trajectory rectification approach is proposed to locate the object with the temporal information. Finally, a multiple scale adaptation mechanism is explored to estimate object size, which helps to relieve interference from background information when we update the appearance model. Extensive experiments on the CVPR2013 tracking benchmark and the VOT2014 challenges demonstrate the superior performance of our method.We present a robust facial landmark tracking based on pose estimation. Firstly,our system employs a pose-based cascade shape regression model to predict the facial landmark locations. Pose-based subset modeling decreases the shape variances in the model learning stage, making the learned regression model more robust to the large pose variances. In addition, we explore a pose tracking model to enhance the temporal consecutiveness between the adjacent frames, and leverage the Kalman filter to make the predicted shape more smooth and stable. Finally, we incorporate a reinitialization mechanism with the facial landmarks as the position priors into the system, which is able to effectively and accurately locate the face when it is misaligned or lost. Our method has ranked first in the 300-VW competition.We propose a general object localization in videos based on the convolutional neural network. Firstly, we cascade three region classification and regression networks by adaptively using the location-indexed convolutional features to refine the target bounding boxes. Besides, we explore the correlation filters on the convolutional feature maps to efficiently propagate the region proposals from the highly-confident detection results. Finally, we perform the object co-occurrence inference with an efficient Look-Up-Table method. This algorithm has ranked first in the ImageNet 2016 on the topic object detection in video..In this thesis, we propose three methods for object tracking and localization in video. Extensive experiments on public benchmarks have illustrated the superior performance of our methods.