Research On Human Skeleton Extraction Based On Deep Learning

Human skeleton extraction is a fundamental research task in computer vision,which facilitates a broad range of applications such as activitity recognition,human-computer interaction and person re-identification.With the rise of deep learning,the research has made some progress.However,there still exist unresolved problems.The existing skeleton extraction approaches do not perform well in modeling the relations between joints and most of them focus on the model accuracy,ignoring the computational complexity,which limits the practicability of models.This paper presents two research findings which solve the above two problems described.The traditional skeleton extraction methods utilize graph models to deal with the relations between joints,but it is difficult to optimize the graph models.Besides,graph models ignore the difference among the joints.Therefore,graph-based methods cannot achieve good results.In order to solve this problem,this paper proposes a novel multitask curriculum context network,in which the different types of joints are regarded as different tasks and the mechanism of curriculum learning is introduced into the multi-task learning by constructing the relations between tasks.Specifically,we use multiple single joint regression networks to get the positions of the different types joints,and they form a multi-task network architecture.In our proposed approach,the internal data flow between the single joint regression networks represents the relations between tasks,which also help induce the dependent relationship between joints.The dependent relationship is similar to the process of human learning(from the easy to the difficult).The experimental results on the LSP dataset and FLIC dataset verify the effectiveness of the proposed method.To ensure the accuracy of the model while reducing model computational complexity,most of existing skeleton extraction methods either reduce the number of model parameters or utilize the parameter sharing technique.However,these methods hurt models' accuracy.In addition,many skeleton extraction models are multi-stage regression models,which ignore the differences between two neighbor tasks.A good multi-stage regression model does not require that the subnet at each stage has the same ability of regressing the accurate joint positions.In this paper,a novel multi-stage regression model based on multi-scale feature learning is proposed,which uses the multiple subnets at different depths to improve the algorithm accuracy and reduce the computational complexity.This paper adopts a Scale-Variable Hourglass Module for each subnet,which can realize multi-scale feature learning.In order to verify the feasibility and effectiveness of the proposed algorithm,the training and testing time,the number of parameters and the joint position accuracy on MPII and LSP datasets are showed in our experiments.