2D Human Pose Estimation Based On Boosted Regression Trees

Estimating human pose from still images is an challenging research subject in computervision area, this subject researching how to extraction and analysis image feature, estimate andrebuild human pose. The research has a very broad application prospects, currently has pre-liminary applications in image retrieval, advanced human-computer interaction and intelligentvideo surveillance.Current mainstream 2D human pose estimation algorithms are based on the Pictorial Stric-tures Models( PSM). Although the PSM is successful in some human pose estimation task,but it is difficult to solve the problem with the following : human pose estimation algorithmsbased on the PSM need to detect human parts in images, but in real world, detecting a singlemember of the human body is very difficult because of the background noise and the wide va-riety of human appearance; simple PSM is difficult to achieve a good enough estimate effects,and complex PSM costing too expensive. To reduce the computational cost in ensuring the es-timated effect, enhance the estimated speed, in this paper, we propose a human pose estimationalgorithms based on Boosted Regression Trees.Firstly, a mixtures of deformable part models are used to represents the target body parts(including head, shoulders and upper torso) to provide a richer representation of the human body,and using Latent Support Vector Machine to train model, effectively detecting human target partfrom the static image. Then, build a dependency graph showing the relationship between thehuman body reference point(such as human joints), the human body pose estimation problemis decomposed into some local pose regression problems on each dependent path by this lay-ered approach, dependency graph can reduce the body pose estimation difficulty. Finally, usingBoosted Regression Trees modeling the local pose regression, from the root of the dependencygraph( center of the detection frame), using trained Boosted Regression Trees to estimate localpose along the path, and eventually combined into human pose. On the end, experiment on theLSP image database verify the effectiveness of the proposed algorithm; experiment on crossdataset verify the generalization of the proposed algorithm.