Tensor Representation Based Visual Tracking

Visual tracking is one of the fundamental areas in computer vision. It plays a critical role in numerous applications, including surveillance, human-computer interaction, video compression and action recognition. Generally, a tracker can be broken down into several constituent parts, namely, motion model, feature representation, observation model and model updater. The feature extractor and observation model are the most important factors that could affect the tracking performance. In the thesis, we first present a structured way of organizing local descriptors to build a global representation, which we call tensor pooling, and cast visual object tracking as an online tensor learning and prediction task. Secondly, we present a tensor sparse coding based feature representation, and the observation model is constructed using generalized tensor regression model. Finally, based on the observations and analysis on the impacts of local noise and object deformation, we present a probabilistic part-based visual tracking algorithm. Details of our works are described as follows.We have proposed a tensor pooling based visual tracking algorithm. We argue that, compared with traditional pooling operators, the tensor pooling could deliver more intrinsic structural information for the target appearance, and can also avoid high dimensionality learning and overfitting problems suffered in concatenation-based pooling methods. Therefore, we propose to represent target templates and candidates directly with pooling tensors, and cast the visual tracking as an incremental tensor learning task. To further improve the robustness of our method against drifting and background noise, we propose to incorporate the tracking algorithm in a discriminative framework. Experiments on a benchmark indicate the effectiveness of our tracker, while the comparison results between different pooling operators demonstrate the superior performance of the tensor pooling operator against others.We have presented a tensor sparse representation based visual tracking algorithm. It applies the tensor sparse coding to local patches on a learned tensor dictionary to obtain local sparse tensor representations. Sparse representation in tensor form can explore the local spatial information of an image, which contributes to the discriminability of the tracking model. The observation model is constructed by generalized tensor regression model with a Bernoulli distribution assumption on the responses, where we assume that the candidates sampled nearby the target are positive ones and those far away from target are negative ones. Experimental results on several challenging sequences indicate the effectiveness of our method.We have proposed a part-based visual tracking algorithm which represents target with a probabilistic part space model. Part-based models are skilled at tackling partial occlusions and objection deformation occurred in visual tracking. Different from other part-based methods, in our proposed method, we consider the part-based tracking as sampling the best parts for tracking from part space under a learnt probabilistic distribution. The target state is estimated using tracked parts with votes. Experimental results on benchmark indicate the superiority of our tracker in handling object deformation and occlusions.