基于稀疏表示的LiDAR点云目标识别

LiDAR point cloud target recognition based on sparse representation

LiDAR获取数据的方式导致点云数据存在阴影、遮挡等现象,造成现有算法识别率低,鲁棒性差。稀疏表示理论表明过完备字典可通过少量重构系数重构样本,从而达到降噪目的。据此提出基于稀疏表示的LiDAR点云目标识别算法。首先,在由所有训练样本组成的过完备字典上重构样本;然后,计算每个测试样本在字典上的稀疏表示重构误差,并利用该重构误差判别测试样本的类别归属。实验表明,所提算法对点云目标的识别率较现有算法均有显著提升,并具有较高的鲁棒性。

The method of LiDAR data acquisition leads to shadow and occlusion of point cloud data, which results in lowrecognition rate and poor robustness of existing algorithms. The sparse representation theory indicates that the over-completeness ofthe dictionary enables the algorithm to reconstruct samples effectively with a small number of reconstruction coefficients, so as toachieve the purpose of noise reduction. Firstly, the samples were reconstructed on the over-complete dictionary composed of alltraining samples. Secondly, the sparse representation of each test sample in the dictionary indicates the reconstruction error, whichis used to identify the category of test samples. The experimental results show that the proposed algorithm has significantlyimproved the recognition rate of point cloud objects compared with the existing algorithms, and has high robustness.