基于组合矩的激光雷达距离像目标识别

激光成像雷达距离像与目标表面物理结构特性密切相关,体现目标的本质特征,是目标识别的主要研究方向。采用组合矩的神经网络方法进行了相干激光雷达距离像目标识别仿真研究。用Hu不变矩和仿射不变矩两者的低阶矩组合表示距离像目标区域特征,利用反向传播(BP)神经网络识别不同方位角的车辆。当视场角不变时,训练10个目标,每个目标取3～19个样本,在不同载噪比(CNR)情况下,分析Hu不变矩、仿射不变矩和两者组合矩的识别率。理论分析和仿真实验表明利用组合不变矩进行距离像目标识别性能优于单独利用其中一种不变矩。

Object Recognition of Ladar Range Image Using Combined Moment Invariants

Ladar range image is closely relative to the physical structure property of object surface, and reflects essential characteristics of objects, thus it has received considerable attention in object recognition field. Image moments are significant methods for object recognition. The low order moments of Hu moment invariants and affine moment invariants are combined to describe the features of three-dimensional objects, and back-propagation (BP) neural network is applied to recognize objects on different azimuth angles. When the view angle is constant, the case that training 10 objects and sampling from 3 to 19 per object, under different carrier-to-noise rates (CNRs), is analyzed about the correct rate of recognition for the Hu moment invariants, the affine moment invariants and the combination of both. The theoretical and experimental analyses prove that using the combination of the Hu moment invariants and the affine moment invariants as features to classify objects can achieve better recognition performance than only using either the Hu moment invariants or the affine moment invariants.