基于GPU的全波形并行LM分解算法

波形分解是机载激光雷达全波形数据处理的重要基础工作,通过求解波形函数模型的参数,将波形数据利用具体的函数模型拟合出来,实现对全波形及其中各个子波形函数表达。LM(Levenberg-Marquardt)算法及其改进的算法是波形分解中对参数进行拟合求解的常用方法。针对LM算法在参数拟合计算的过程中存在大量迭代和矩阵运算,提出了基于线程块组和线程两级并行粒度的并行计算方案。将串行多次循环迭代求解参数改为单次并行计算取最佳值实现对参数的选择,将矩阵运算进行线程块的协同并行计算,实现了LM算法在通用计算图形处理器上的并行计算。实验证明,在规定阈值条件下,并行LM降低了算法的迭代次数,提高了波形分解LM算法的计算效率,为提高波形分解的处理效率提供了研究思路。

GPU Based Parallel LM Algorithm for Full-Waveform Decomposition

As an important basic work of the processing of airborne LiDAR full-waveform data, the decomposition of waveform can calculate parameters of waveform function model which could simulate the waveform of LiDAR waveform data, thus, the full-waveform and the sub-waveform can be presented by specific parameterized function. LM ( Levenberg-Marquardt) algorithm and LM-based improved algorithms are frequently used to calculate the pa-rameter of function model in decomposition of waveform. In consideration of the iteration and matrix operation in LM, a parallel computational scheme based on block groups and blocks in GPU is presented. Serial iteration is turned into parallel calculating to choose the best parameter using parallel calculating the matrix by blocks in GPU. The experiment shows that this method can reduce the number of iteration, improve the computational efficiency of LM and provide a new idea on improving the efficiency of decomposition of waveform.