多频飞行时间相机实时深度补偿算法研究  

Real-Time Artifact Compensation for Depth Images of Multi-Frequency ToF

作　　者：王森 王润孝[1] 左欣欣 于薇薇[1] WANG Sen;WANG Runxiao;ZUO Xinxin;YU Weiwei(School of Mechanical Engineering, Northwestern Polytechnical University, Xi′an 710072, China;School of Computer Science and Technology, Northwestern Polytechnical University, Xi′an 710072, China)

机构地区：[1]西北工业大学机电学院,陕西西安710072 [2]西北工业大学计算机学院,陕西西安710072

出　　处：《西北工业大学学报》2019年第1期152-159,共8页Journal of Northwestern Polytechnical University

基　　金：国家自然科学基金(51475373;61603302);陕西省科技统筹创新工程项目(2016KTZDGY06-01);陕西省自然科学基金(2016JQ6009);111引智计划(B13044)资助

摘　　要：针对多频飞行时间相机在曝光时间内物体运动以及相位融合中产生的偏差,提出一种基于光流以及空间核密度估计的实时补偿算法。首先获取原始的多频相位图像,计算出不同频率之间的相对光流,获得每种频率下的补偿相位;在此基础上针对不同频率的相位做出可信度假设并利用空间核密度估计对其进行排名,获得最终的相位融合图并生成对应的深度图。实验部分,使用Kinect V2相机获取的原始多频相位图对算法进行验证,同时利用GPU进行并行加速;结果表明该算法可以实时对物体平行于光轴运动的相位融合偏差进行补偿,有效地提高了深度图像的精度和稳定性。During the last few years, Time-of-Flight(TOF) sensor achieved a significant impact onto research and industrial fields due to that it can capture depth easily. For dynamic scenes and phase fusion, ToF sensor′s working principles can lead to significant artifacts, therefore an efficient method to combine motion compensation and kernel density estimate multi-frequency unwrapping is proposed. Firstly, the raw multi-phase images are captured, then calculate the optical flow between each frequency. Secondly, by generating multiple depth hypotheses, uses a spatial kernel density estimation is used to rank them with wrapped phase images. Finally, the accurate depth from fused phase image is gotten. The algorithm on Kinect V2 is validated and the pixel-wise part is optimized using GPU. The method shows its real time superior performance on real datasets.

关 键 词：飞行时间相机 多频 光流 空间核密度估计 KinectV2 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]