Gaussian processes based transfer learning for online multiple-person tracking and building blocks for deep learning

Abstract

This thesis mainly study the application of the Gaussian Processes (GPs), especially the application of GPs regression for multiple-person tracking. Furthermore, we explore the scalability of the GPs model. Most existing multiple-person tracking algorithms are still limited by abrupt human pose change, scale change and lighting condition tend to drifts. We introduce a GPs regression based observation model to deal with these challenges. During the tracking process, background information is taken into account to cope with the dynamic background, the GPs regression based observation model fuses prior information to make a tracking decision, which can cope with short term occlusion. Another benefit is that the information of the target in the current frame can be extracted to re-weight the target information in the previous frame, after that a tracking decision is made, this can be seen as a transfer learning strategy. Recently, neural networks have made breakthroughs in various fields. Especially convolutional neural networks (CNNs) have made significant progress in image processing area. The key to the success of CNNs in image processing is that it has a high model complexity and can process high-dimensional features layer-by-layer. After that, extracting different levels of abstract information for further use. We explore the scalability of GPs model, by stacking multiple GPs together, we can construct a deep architecture with building blocks of GPs. This deep hierarchy Gaussian process model is capable of processing high-dimensional input and extracts different levels of abstract information, which is suitable for image processing. We further explore the GPs model equipped with the convolutional kernel, making it benefit from the non-local generalization of convolutional structure and achieving better performance for image data.