Real time human action recognition and understanding

Human action recognition and interpretation constitutes an important part of the video understanding. In this study, a novel action recognition system is developed which can detect and understand human actions. Thus, a robust system that can understand actions automatically is produced, so that, a structure is put forward as a first step to a fully developed video sensing framework.The proposed system defines the action as a temporal pattern which is described by still image information combined with movement information. To realize that, firstly, foreground is extracted and the region containing no beneficial data is eliminated. Then, the location and coordinates of area of interest are obtained using blob analysis. Optical flow is used to determine the movement in the interest region and the output is used to construct the movement power shape. Descriptors are constituted by histogram of oriented gradient analysis of optical flow power shapes. Descriptors lined up in the time plane to form identifiers that contain the concept of motion, still image and time domain. Based on the power and reliability of these descriptors, a simple sliding window classifier is formed in the time domain.In this study, it is seen that human actions can be represented efficiently by using descriptors of motion, still image and time. As a result, an autonomous action recognition system is produced which can detect human actions in a way that is steady to work successfully under changing conditions, hence, a structure is put forward for video sensing.