基于协同重排序的手势识别方法  

Collaborative Reranking: A Novel Approach for Hand Pose Estimation

作　　者：张芷君 钟胜 吴郢 王建辉 Zhang Zhijun;Zhong Sheng;Wu Ying;Wang Jianhui(School of Automation,Huazhong University of Science and Technology,Wuhan 430074;EECS Department,Northwestern University,Evanston,IL USA 60208)

机构地区：[1]华中科技大学自动化学院,武汉430074 [2]EECS Department, Northwestern University

出　　处：《计算机辅助设计与图形学学报》2018年第11期2182-2192,共11页Journal of Computer-Aided Design & Computer Graphics

摘　　要：手势识别是计算机视觉的一个非常具有挑战性的问题,可运用于人机交互、手语识别、虚拟角色控制等众多领域.然而,由于手本身具有极高的自由度,通过样本直接估计所有手势参数相当困难.为此,提出一种可分解手势参数的手势估计方法——协同重排序.首先将手根据指骨的关节角度划分为多个局部观测单元,并建立离线的局部估计数据库;然后利用此数据库,使用k-最邻近(k-NN)搜索算法对从深度图中获得的局部观测单元进行姿态估计;最后依据当前观测单元的k-NN搜索结果对姿态估计结果重新排序,收敛后得到最终估计结果.除了手势局部参数估计方法之外,还提出一种手的全局姿态估计的方法,使得整个方法可更好地适用于多种任务场景.对合成图像和真实深度图像数据集验证文中方法的性能:不用GPU加速的情况下,该方法可以在30 ms内完成手势识别(其中局部姿态估计17 ms,全局姿态估计12 ms),最大平均估计误差小于10°,具有很高的效率和有效性.The pose estimation of hand is a theoretically interesting and challenging problem in computer vision with many applications such as human-computer interaction,sign language recognition,virtual character control and so on.However,due to the high degree of freedom(DoF)of the hand pose,it is very difficult,if not impossible,to directly estimate the hand pose efficiently.In this paper,we proposed a hand pose estimation method,namely,collaborative reranking.It divides the observation of hand into many partial observations,which is related with a subset of the phalanges joint angles.And it builds a partial observation database for each partial estimator position off-line.At pose estimation stage,it extracts partial observations from the depth image,and estimate the parameters for each partial observation by k nearest neighbors(k-NN)searching.Then it reranks the k-NN of each partial observation according to the k-NN searching result of its neighbor partial estimators.We model this idea into a graph model,and obtain the collaborative reranking algorithm by systematically and rigorously mathematical inference.Although collaborative reranking mainly focuses on hand local pose estimation,we also proposed a method to estimate the hand global motion to make the system usable.Finally,we verify the performance of the proposed method by experimenting on synthetic and realistic depth image.The proposed method can estimate hand pose within 30 ms(17 ms for local pose estimation and 12 ms for global pose estimation)without GPU speedup,and the maximum average estimation error is less than 10°.Extensive experimental results demonstrated the efficiency and effectiveness of the proposed method.

关 键 词：协同重排序 手势识别 人机交互 数据库搜索 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]