Model-free Viewpoint Invariant Human Activity Recognition

Abstract —The viewpoint assumption is becoming an obstacle in human activity recognition systems. There is increasing interest in the problem of human activity recognition, motivated by promising applications in many domains. Since camera position is arbitrary in many domains, human activity recognition systems have to be viewpoint invariant. The viewpoint invariance aspect has been ignored by a vast majority of computer vision researchers owing to inherent difficulty to train systems to recognize activities across all possible viewpoints. Fixed viewpoint systems are impractical in real scenarios. Therefore, we attempt to relax the infamous fixed viewpoint assumption by presenting a framework to recognize human activities from monocular video source from arbitrary viewpoint. The proposed system makes use of invariant human pose recognition. An ensemble of pose models performs inference on each video frame. Each pose model employs an expectation-maximization algorithm to estimate the probability that the given frame contains the corresponding pose. Over a sequence of frames, all the pose models collectively produce a multivariate Viewpoint invariantime series. In the activity recognition stage, we use nearest neighbor, with dynamic time warping as a distance measure, to classify pose time series. We have performed some experiments on a publicly available dataset and the results are found to be promising.