Human Motion Analysis for Efficient Action Recognition

Abstract

Automatic understanding of human actions is at the core of several application domains, such as content-based indexing, human-computer interaction, surveillance, and sports video analysis. The recent advances in digital platforms and the exponential growth of video and image data have brought an urgent quest for intelligent frameworks to automatically analyze human motion and predict their corresponding action based on visual data and sensor signals. <br /> This thesis presents a collection of methods that targets human action recognition using different action modalities. The first method uses the appearance modality and classifies human actions based on heterogeneous global- and local-based features of scene and humanbody appearances. The second method harnesses 2D and 3D articulated human poses and analyizes the body motion using a discriminative combination of the parts’ velocities, locations, and correlations histograms for action recognition. The third method presents an optimal scheme for combining the probabilistic predictions from different action modalities by solving a constrained quadratic optimization problem. <br /> In addition to the action classification task, we present a study that compares the utility of different pose variants in motion analysis for human action recognition. In particular, we compare the recognition performance when 2D and 3D poses are used. Finally, we demonstrate the efficiency of our pose-based method for action recognition in spotting and segmenting motion gestures in real time from a continuous stream of an input video for the recognition of the Italian sign gesture language.