CAJ | 학술논문

为了实现在无标记的情况下不对称地跟踪人手及其所有关节的位置和姿态，提出一种软硬件结合的混合跟踪计算框架，同时使用电磁跟踪器和无标记人手关节姿态分析算法提出基于 CUDA 的异步并行粒子群优化(PSO)加速方法。首先通过跟踪器测量人手手腕的位置姿态，使用Kinect数据作为输入，在三空间(双颜色空间和深度空间)下进行手部区域分割；然后使用PSO方法将手关节的23个自由度的跟踪问题转化为求解一个优化问题，使用不对称策略来提高部分手指的跟踪性能，寻找给定参数空间内能够最小化观测值和估计值之间偏差的手模型参数解。该方法不需要进行任何标记，可以对手部关节姿态进行连续跟踪，实验结果表明，其在实验的硬件平台上可以达到12帧/s的运行速度，平均误差稳定在10 mm以内。
위료실현재무표기적정황하불대칭지근종인수급기소유관절적위치화자태，제출일충연경건결합적혼합근종계산광가，동시사용전자근종기화무표기인수관절자태분석산법제출기우 CUDA 적이보병행입자군우화(PSO)가속방법。수선통과근종기측량인수수완적위치자태，사용Kinect수거작위수입，재삼공간(쌍안색공간화심도공간)하진행수부구역분할；연후사용PSO방법장수관절적23개자유도적근종문제전화위구해일개우화문제，사용불대칭책략래제고부분수지적근종성능，심조급정삼수공간내능구최소화관측치화고계치지간편차적수모형삼수해。해방법불수요진행임하표기，가이대수부관절자태진행련속근종，실험결과표명，기재실험적경건평태상가이체도12정/s적운행속도，평균오차은정재10 mm이내。
In order to track position and orientation of hand and its articulations, we implement a hybrid solution for markerless hand articulation tracking with electromagnetic tracker and visual observations from Kinect. The position tracking of hand uses an electromagnetic tracker, and tracking of 23 DOFs of articulations is treated as an optimization problem which is solved using Particle Swarm Optimization (PSO), and it is essentially the progress of seeking for the hand model parameters that minimize the discrepancy between the hypothesized hand model and actual hand observations. The system input is RGB and depth images from Kinect sensor. This system does not require special markers and is capable to track hand articulations frame by frame. Experiments show that ro-bust 3D tracking of hand articulations can be achieved in 12FPS, and average error is less than 10mm.