Classification method for imbalanced LiDAR point cloud based on stack autoencoder

Abstract

The existing classification methods of LiDAR point cloud are almost based on the assumption that each class is balanced, without considering the imbalanced class problem. Moreover, from the perspective of data volume, the LiDAR point cloud classification should be a typical big data classification problem. Therefore, by studying the existing deep network structure and imbalanced sampling methods, this paper proposes an oversampling method based on stack autoencoder. The method realizes automatic generation of synthetic samples by learning the distribution characteristics of the positive class, which solves the problem of imbalance training data well. It only takes the geometric coordinates and intensity information of the point clouds as the input layer and does not need feature construction or fusion, which reduces the computational complexity. This paper also discusses the influence of sampling number, oversampling method and classifier on the classification results, and evaluates the performance from three aspects: true positive rate, positive predictive value and accuracy. The results show that the oversampling method based on stack autoencoder is suitable for imbalanced LiDAR point cloud classification, and has a good ability to improve the effect of positive class. If it is combined with optimized classifier, the classification performance of imbalanced point cloud is greatly improved