Human pose estimation in image sequences

Abstract

Human action recognition (HAR) has been a popular research topic and received a huge attention for several decades due to its wide range of applications such as security and surveillance, human computer interaction, health care and video indexing. However, most research focus on either video or image sequence but very few work is done on still images. The process of estimating pose configuration in a still image is called as human pose estimation (HPE). One of the problems dealing with still image for human action recognition is that there exist many articulated human points which are difficult to be captured within a single image. Moreover, more often than not, the ability to obtain the posture adds as an extra cue to the contextual information for recognizing human action. Furthermore, excessive background elements are unnecessary and often contribute to false detection of pose estimation algorithm. The objective of this project is firstly to design an effective model in estimating human pose or structure in still images by showing skeleton line of different size depicting different parts of the human body. In order to analyze posture in still image, the low resolution video is separated into several frames and each frame is enhanced by subtracting the background for accurate detection. Then, the frame is parsed into pose estimation algorithm to capture the human structure. From the result of performance evaluation, background subtraction successfully increases the true positive rate (TPR) but not the precision. On the other hand, the introduction of region of interest (ROI) successfully increases the accuracy of HPE detection by 2.16 % in the positive rate and 16.46 % in the negative rate for proposed evaluation when threshold is equal to 25. However, the TPR of ROI enhancement (88.84 %) shows slightly lower than the original algorithm (93.39 %) due to certain frames that were unable to be detected. As a conclusion, the proposed method performed at least as good as those of the state-of-art methods in estimating the human post and subsequently in classifying the human actions.