Dynamic multi-channel feature dictionaries for robust object tracking

Abstract

Object tracking is a fundamental problem in computer vision, with applications in the fields of video analytics and medical imaging. In this thesis, we present a novel approach to solve the object tracking problem in a particle filter framework based on sparse visual representation.

We experimentally validate our approach on standard object tracking benchmark datasets using several quantitative evaluation measures and robustness tests. Specifically, we use average center pixel location error, average success rate, success plots and precision plots as quantitative evaluation metrics. To test the robustness of our approach to position and scale initialization errors, we report results for both temporal and spatial robustness tests. Finally, we perform a comparative analysis of the results for each of the evaluation tests with several recently proposed tracking algorithms that report state-of-the-art performance, and show significant gains.

Current state-of-the-art trackers use low-resolution image intensity features as part of object appearance modeling. Such features often fail to capture sufficient visual information about the object, and ultimately drift away. In our work, we employ visually richer representation schemes to model the appearance of the object. Specifically, we construct multi-channel feature dictionaries using image intensity, normalized gradient magnitude, and quantized gradient orientation information. To further mitigate the tracking drift problem, we take into account the dynamics of the past state vectors of the object, and propose a novel dynamic adaptive state transition model. We also demonstrate the computational tractability of using richer appearance modeling schemes by adaptively pruning candidate particles during each sampling step, and using a fast augmented Lagrangian technique to solve the associated optimization problem.