基于正交匹配追踪及加速近端梯度的人体三维重建

发布时间：2018-02-01 16:38

本文关键词： 三维重建人体运动重建稀疏表示加速近端梯度　出处：《中国生物医学工程学报》2017年04期 　论文类型：期刊论文

【摘要】：为提高人体三维结构的重建精度,针对重建过程中字典中原子的最佳选择和结构矩阵的优化问题,结合稀疏表示和低秩约束,提出一种正交匹配追踪追踪及加速近端梯度(OMP-APG)算法,以此为医学领域提供丰富的信息,以辅助医生快速精确地制定出治疗方案。首先,对特征点观测矩阵进行奇异值分解(SVD)分解,利用列文伯格-马夸尔特(LM)算法得到唯一确定的相机旋转矩阵;其次,利用稀疏表示中"最大化逼近"思想,通过正交匹配追踪算法对轨迹基系数进行求解,结合预定义的轨迹基求解出人体三维结构矩阵;最后,根据结构矩阵是一个低秩矩阵,将其秩优化问题转化为核范数最小化问题,利用加速近端梯度算法对人体结构矩阵进一步优化处理。将该算法与稀疏逼近算法进行比较,对伸懒腰、瑜伽、拾物、喝水和跳舞等5组不同的人体运动模型进行三维重建,通过其三维重建效果图和三维重建误差的结果显示,其重建精度更高且稳定性更好。在该算法下喝水运动的重建效果最佳,其1 102帧图像序列41个特征点的重建误差为0.030 3,而在稀疏算法下的重建误差为0.017 8。因此,该算法可以有效地提高人体三维结构的重建精度,为医学领域辅助治疗提供相应的技术支持。
[Abstract]:In order to improve the reconstruction accuracy of human three-dimensional structure, the optimal selection of atoms and the optimization of structure matrix in the dictionary are combined with sparse representation and low rank constraints in the process of reconstruction. An OMP-APG algorithm for orthogonal matching tracking and accelerating proximal gradient tracking is proposed to provide rich information for the medical field to assist doctors to quickly and accurately work out the treatment plan. First of all. The singular value decomposition (SVD) decomposition is applied to the observation matrix of feature points, and the unique camera rotation matrix is obtained by using the Levinberg-Marquardt LM algorithm. Secondly, using the idea of "maximum approximation" in sparse representation, the trajectory base coefficient is solved by orthogonal matching tracing algorithm, and the three-dimensional structure matrix of human body is solved by combining the predefined trajectory base. Finally, according to the fact that the structure matrix is a low rank matrix, the rank optimization problem is transformed into a kernel norm minimization problem. Using the accelerated proximal gradient algorithm to further optimize the human structure matrix. The algorithm is compared with the sparse approximation algorithm to stretch, yoga, pick up objects. Five different groups of human motion models, such as drinking water and dancing, were reconstructed by 3D reconstruction. The results of 3D reconstruction and 3D reconstruction error showed. The reconstruction accuracy is higher and the stability is better. Under this algorithm, the reconstruction effect of water motion is the best, and the reconstruction error of 41 feature points of the 1 102 frame image sequence is 0.0303. The reconstruction error under sparse algorithm is 0.017 8. Therefore, the algorithm can effectively improve the reconstruction accuracy of human three-dimensional structure and provide the corresponding technical support for medical auxiliary therapy.
【作者单位】：浙江理工大学信号与信息处理研究所;
【基金】：国家自然科学基金项目(61672466) 浙江省自然科学基金(LZ15F020004,LY1720034) 浙江理工大学521项目
【分类号】：R318;TP391.41
【正文快照】： 引言利用二维图像序列来恢复人体三维结构成为近几年的研究热点,即寻找合适的算法从已知的二维序列观测点中恢复人体的三维结构[1-2]。在先前的研究中,Tomasz和Kanata提出利用观测矩阵的秩约束对观测矩阵进行因式分解(SVD分解)[3],即在正交投影模型下结合观测矩阵的秩约束条件

【相似文献】