Model-Based Viewpoint Invariant Human Activity Recognition from Uncalibrated Monocular Video Sequence

摘要

There is growing interest in human activity recognition systems, motivated by their numerous promising applications in many domains. Despite much progress, most researchers have narrowed the problem towards fixed camera viewpoint owing to inherent difficulty to train their systems across all possible viewpoints. Fixed viewpoint systems are impractical in real scenarios. Therefore, we attempt to relax the fixed viewpoint assumption and present a novel and simple framework to recognize and classify human activities from uncalibrated monocular video source from any viewpoint. The proposed framework comprises two stages: 3D human pose estimation and human activity recognition. In the pose estimation stage, we estimate 3D human pose by a simple search-based and tracking-based technique. In the activity recognition stage, we use Nearest Neighbor, with Dynamic Time Warping as a distance measure, to classify multivariate time series which emanate from streams of pose vectors from multiple video frames. We have performed some experiments to evaluate the accuracy of the two stages separately. The encouraging experimental results demonstrate the effectiveness of our framework.