Hand Pose Estimation Based On Neural Network And Particle Swarm Optimition

Depth-based hand pose estimation is a hotspot in human computer interaction.In the past few years,particle swarm optimization(PSO)has attracted increased attention for hand pose estimation,and PSO-based algorithm has made great progress.However,PSO-based algorithm is sensitive to initial pose and is easy to be trapped in local optimal.To address these problems,this paper presents a novel hand pose estimation method based on multi-view convolutional neural network(MV-CNN)and PSO algorithm.Experiments demonstrate the effectiveness of our approach.Firstly,we present a hand pose estimator based on MV-CNN.Our approach aims to obtain preferable initial poses from MV-CNN firstly,and then the outputs of MV-CNN are employed as the initialization for the following PSO stages to improve the performance of hand pose estimation.Secondly,we propose to apply a novel local pose sampling strategy during PSO stages.The local pose sampling strategy aims to expand the searching range during random walk stage of particle.Our strategy improves the searching ability of PSO algorithm,and prevents the local optima problem.Experiments on challenging NYU datasets demonstrate that our approach outperforms previous methods and greatly improves the robustness and accuracy to large pose variations in hand pose estimation.