摘要
基于深度图像的手势识别技术是下一代数字媒体设备的主要交互手段,从深度图像中准确定位出"干净"的手部图像显得尤为重要.提出了一种同步进行左右手分割的改进方法,在传统SegNet算法的基础上,加入了类别权重、转置卷积、混合式空洞卷积组合和编解码器之间的拼接合并跳层连接,使左右手的F2-Score相较基准方法分别提高了7.6%和5.9%.推理速度在GPU上达到了20.5 ms/帧,可以实时处理深度图像序列.实验证明采用本方法对深度图像进行左右手同步分割时可以得到更加精准的分割结果.
Hand gesture recognition technology based on depth image,which relies on the accurate identification of"clean"hand in the captured depth image,is the primary interactive mode for digital media devices of future generation.We propose an improved approach to simultaneous left-and right-hand segmentation,extending the traditional SegNet algorithm by strategies including class weight,transposed convolution,hybrid dilated convolution,and skip-connection between the encoder and decoder performed by concatenation.Our approach achieves higher F2-Score than the existing baseline by 7.6%for the left and 5.9%for the right hand.The processing on the GPU reaches 20.5 ms per frame at inference time,making real-time hand tracking in depth image sequences feasible.The results of the experiment demonstrate that our approach can considerably improve the performance of simultaneous left-and right-hand segmentation from a single depth map.
作者
徐正则
张文俊
XU Zhengze;ZHANG Wenjun(Shanghai Film Academy,Shanghai University,Shanghai 200072,China;School of Communication,East China Normal University,Shanghai 200241,China)
出处
《上海大学学报（自然科学版）》
CAS
CSCD
北大核心
2021年第3期454-465,共12页
Journal of Shanghai University:Natural Science Edition
关键词
深度图像
手部分割
改进方法
depth image
hand segmentation
improved approach
