Research On Human Pose Estimation In Static Images

Human pose estimation in static images as the basis of subsequent behavior understanding was to locate body parts position after pedestrian detection in given images and to compute orientation and scale information. Human pose estimation was very challenging due to flexibility of human body and complexity of background. Shape feature and histogram of oriented gradient(HOG) was often used, but shape feature was easily affected by different pose and cloth and HOG feature was difficult to compute. Given an image, the position of body parts were unknown, pose search space was extremely big because of the pose human can take was changing over time. To extract body parts feature and reduce pose search space, two main contributions were made:1. In order to address the problem of low accuracy of pose estimation caused by insufficient feature expression of body part models, a pose estimation method based on Pictorial Structure and novel text feature is proposed. Better appearance model is adopted which prior and latent relationships of different body parts are learned from annotated images then help in estimating better appearance models on test images. Haar characteristics LBP text feature is used to extract the text information of body parts. Furthermore, an image is block processed and different weight is assigned to different block. Experimental results show that when compared with HLBP, normalized HLBP and color feature, weighted HLBP captures text feature more effectively and gains higher accuracy.2. Aiming at solving the problem of high dimension of pose search space during pose estimation, a pose search space reducing algorithm of Grab Cut based on Simple Linear Iterative Clustering(SLIC) superpixel approach was proposed. Treating pedestrian detection bounding box as an input and using SLIC to segment images into superpixels. Superpixels were used as nodes to build s-t graph. Foreground and background Gaussian mixture model(GMM) were respectively build and Gaussian parameters were updated using iterative processing. Image foreground extraction was achieved using Min Cut. Pose estimation was performed only in the foreground area obtained by foreground extraction. Experimental results showed that comparing to pose search space reduction method based only on Grab Cut, the algorithm of Grab Cut using SLIC has better performances both on running time and pose estimation accuracy.In words, two contributions discussed above had great capacity to estimate human pose in static images. Body parts feature extraction was committed by dividing color images into patches to abstract texture information then assigning different weights according to different feature values. Pose search space reduction was committed by using superpixel segmentation method to restrict pose inference in small area of foreground. Running speed and pose estimation accuracy can be improved at the same time.