| With the ever-increasing interaction between machines and humans, it has become imperative that machines have the ability to intelligently understand human behavior and modify their response accordingly. A primary aspect of human activity recognition is object tracking. While being a simple task for humans, object tracking is monumentally more challenging for computer vision systems. This work addresses two problems. We first propose a novel non-intrusive algorithm for human activity recognition. Our approach is monocular, fairly view independent, and can be applied to both frontal and lateral views of most activities. Experiments with short and long video sequences show robust recognition under conditions of varying view angles, zoom depths, backgrounds and frame rates. Secondly, we address the problem of robust object tracking in noisy environments. We adopt an analytical approach to formulating weighting functions for a feature-based tracking algorithm. The robustness of incorporating the weighting functions is evaluated over noisy synthetic as well as real test images. Performance comparisons of weighting functions are also made with image sequences where noise has been removed prior to testing. Simulation results in a noisy environment show a marked improvement in the performance of the feature-based tracking algorithm. |
