The Research Of Pose Estimation On Pictorial Structure

The technology of human pose estimation is widely used in the fields of video surveillance, human-computer interaction. It is considered as the basis of target classification and behavior understanding, as well one of important issues in the field of computer vision. In pictures, people may appear at any scale, illumination varies diversely, and contrast may be poor. A person’s appearance is unconstrained and his pose is various. This thesis estimates human pose in such conditions, and the main contents are as follows:1) As human pose estimation is susceptible to the background, this thesis proposed a pre-process stage to reduce the influence from background. It first used a deformable part model to detect approximate location of the human body by a bounding box, and then took use of a face detector to optimize the detection. In order to make the detection window adapt to the subsequent position estimates, this thesis adjusted the size of the box according to proportion of human in the detection window. This procedure reduced background effects while maintained human information in the foreground. Experiments showed that this pretreatment gained good result while there are multi-people in image or the proportion of people in image is small.2) The drawback of existing methods of pose estimation is that the search space is too large, which results in slow reasoning, low accuracy. In order to solve this problem, this thesis used the detection box obtained in the pre-process stage as the input of the GrabCut, segmented image in bounding box into foreground/background, removed the background in the bounding box, and restricted the search space determined by GrabCut. Then the method combined edge template with the region template to parse human pose to receive a final estimation results. Experimental results showed that this method reduced search space effectively, while at the same time kept reliable body module candidates for pose inference. Therefore, the pose estimation is more accurate and the time performance of the algorithm is much better.3) Since the existing methods did not consider the location relation between human body parts. Taking use of the detection box obtained from the pretreatment, this paper proposed a concept named location prior on the basis of the improved method in chapter four. Then this paper combined the location prior and the image histogram to obtain the specific location of human body part in test image. Experiment results showed that this approach could get more accurate pose estimation.