A Research Of Human Pose Estimation Based On Deep Convolutional Neural Network

Pose estimation is a foundmental technology to image understanding and behavior recognition. However, occlusions, appearance variations, in-plane and out-plane rotations make human pose estimation long be a significant challenge in computer vision. Recent years, the development of deep leaning brings new way to pose estimation. Compared to shallow machine learning, deep model has deeper hierarchies and the ability to learn more complex patterns. Nonetheless, deep learning itself also remains many unsolved problems, e.g., explanations on its effectiveness in theory and hard to train a usable model in in practice. Meanwhile, it is remains to be further researched how to address pose estimation with deep learning.In this paper, The development of pose estimation and deep learning are reviewed.An Adaptive Holons Representation(ADORE) framework taking advantages of local and global cues is proposed to improve the pose estimation accuracy. In particular, ADORE is made up of two componets:(1) the holons part, global model is designed to first infer body joints location on gloabl level;(2) the adaptive part, local model is proposed to subsequently detect the joints in the potential resions for more accurate joints locating.Laying in the heart of ADORE, global model can be used in three ways:(1) locates body joints based on the whole image;(2) generates several potential regions to reduce the search spaces for local model;(3) be regarded as an image-dependent spatial model to replace the hand-crafted model. These three features are realized by a Convolutional Neural Network named Independent Losses Net(ILN). In ILN, pose estimation is formulated as a classification problem regard to joints’ locations. ILN consists of two independent output layers that respectively predicts the two dimensions of a joint’ location.Furthermore, two independent loss functions are set to instruct the learning of ILN.In ADORE, global model itself is a complete model, and local model is a compensation to global model. This flexibility make ADORE applicable to different scenarios.Experimental results on two challenging benchmark tasks demonstrate that our proposed framework is more efficient than other deep models while remains desirable performance.