Bayes Filters with Improved Measurements for Visual Object Tracking

Abstract

Visual object tracking uses cameras to track target objects in the environment, which has many applications nowadays, such as intelligent surveillance, medical care, intelligent transportation and human-machine interaction. However, it is still a challenging task because of background noises, occlusions, illumination changes and fast motion. The goal of this dissertation is to improve measurements in Bayesian filtering frameworks for visual object tracking as follows:First, we combine multiple visual cues to improve the measurement for lane tracking. The lane is modeled by a linear-parabolic shape, which is a trade-off between accuracy of the fit and robustness with respect to image artifacts. In contrast to previous methods for linear-parabolic lane tracking, we use not only the color and edge information, but also the gradient orientation as visual cues. The lane tracking becomes a statistical reference problem when these local visual cues are available. The probabilistic distribution of lane parameters are estimated from the visual cues by multiple kernel density estimation, which is proved to be very robust to the image noise. Furthermore we use this probabilistic distribution function as the measurement model of the partitioned particle filter to update lane parameters. The experiments show that our novel lane tracking framework has its strength in a new combination and improvement of various advanced methods.Second, we use color invariant histograms to improve the measurement for rigid object tracking. Color histograms have become popular and important descriptors for object tracking, due to their simplicity, effectiveness and efficiency. However, they suffer from illumination changes, e.g., the RGB color histogram is the most prevalently used histogram, but it has no invariance properties to illumination changes. This paper addresses this problem by: a) studying the invariance properties and the distinctiveness of color histograms; b) evaluating the color histograms on large benchmark datasets; c) studying the effects of the kernel mask which adds the spatial information to the color histogram; d) investigating three state-of-the-art object tracking algorithms for evaluations: the integral histogram based exhaustive search, the kernel based mean shift and the particle filter. The results reveal that color histograms which have invariance properties can improve the performance of object tracking. If no prior knowledge about the environment of the dataset is available, the HSV, Spherical and nRGB histograms are recommended.Third, we employ multiple sensors to improve the measurement for moving object tracking. Tracking systems can be more accurate, complete and robust by using fused information from multiple sensors. Therefore, we develop a new filter called central difference information filter (CDIF) for nonlinear estimation and sensor fusion, which has fewer predefined parameters as compared to the unscented information filter (UIF) which was introduced in the literature recently. In addition, we introduce the square-root extensions of the CDIF and UIF to improve the numerical stability, e.g., improved numerical accuracy, double order precision and preservation of symmetry.In summary, we have proposed three methods to improve the measurement for robust object tracking, i.e., multiple visual cues combination, color invariant histograms and multiple sensor fusion. We believe that the new ideas and theoretical insights presented in this thesis, will open new ways of research for future algorithms and applications.