Human activity detection via pose estimation in constrained videos

Abstract

Behaviour recognition is one of the most promising areas of research in the recent past. Automatic detection of human behaviour will facilitate in understanding more about the context involved. Vision-based human behaviour recognition poses a sub-problem of Human Action Recognition (HAR). Solutions to these problems need to be robust rather than accurate. To state precisely, given a sequence of images or a video, the idea is to devise solutions for automatically recognizing the activity performed by the humans. This thesis has attempted at deriving solutions for detecting human activities from a public action dataset. The process probably starts with image processing techniques including noise removal, followed by (low-level) feature extraction to locate lines, regions and alignment areas with certain textures. Activity recognition involves interpretation of human actions using a series of image observations with respect to environmental conditions. To be more precise, the respective environment and the objects, over which the actions happen, also have a defined role in interpretation of the human actions concerned. This work concentrates only on full-body movements excluding the environments, context and other objects in the scene. This work proposes methodologies to detect human interactions from University of Texas Interaction dataset through human pose estimation followed by activity detection. The assumption is that the background, objects and other items are excluded and the skeleton based human body modeling is performed for determining effective human poses. newline