Depth And Skeleton Based View-invariant Human Action Recognition

Abstract

Recognition of human activity plays an important role in computer-human interaction, surveillance, reconnaissance, robotics for humans, and understanding interpersonal behaviour relationships. These activities can be recorded as a sequence of still images but only using vision to solve the HAR poses a major task due to problems like scale variation, wide change, in contrast, lighting, viewpoint and occlusions. Thus to address this our work is concentrated on developing and training two deep learning pipelines one Spatiotemporal based and the other being skeletal based on publicly available human activity classification datasets. Moreover, we merge the two pipelines using late fusion and provide a comparison between the three with the existing state of the art algorithms for various activities in the dataset. Finally, we present the future work for the same problem.