Research On Mixture Part-based Pose Estimation Methods And Their Applications

The human pose estimation is a very popular problem in computer vision. In this work, we willconsider the problem of pose estimation on the static images. We propose an extension based on themixtures of part model to get a more reasonable and more effective pose estimation model to improvethe accuracy of pose estimation. The key idea of our method is to improve the accuracies for leafparts’ localizations-an issue that is largely ignored by previous works-by incorporating both localand non-local contextual information into the model. In particular, we use the local contextualinformation to reduce or eliminate the influences of the noises, while the non-local contextualinformation helping to improve the detection accuracies of the leaf parts. Since more accurate partlocalizations usually mean a more reasonable active set of spatial constraints, this potentiallyenhances the effectiveness of the subsequent optimization procedure. More importantly, we keep thebasic structure of tree-based model, hence taking the advantage of its conceptual simplicity andcomputational efficient inference.In addition, we will also briefly study the application of pose estimation in this work. We propose anovel action recognition approach combining the above pose estimation model. Unlike many othermethods, we directly take the pose obtained from the pose estimator as the input of the actionrecognizer. We first get the pose vector of the given image. Then we represent the pose with a newmodel-we use the spatial constraints exists between each pair of parts to describe the pose-whichcan models all the possible discriminative spatial constraints between different classes of actions intoit. Then we use a sparse group lasso method to learn these discriminative constraints from the posemodel and apply them to action classifications. This will help eliminate the noisy features and get acompact and robust action representation, which can simplify our classifiers and improve the speedand accuracy of classifications greatly. This is also helpful to improve the system’s capability totolerate unstable pose estimation and get the model a good generation performance.Our experiments on public datasets demonstrate the feasibility and effectiveness of the proposedmethods.