A Convolutional Neural Network-Based Noise Filtering Method for Photon-Counting LiDAR Data

Photon-counting LiDAR plays an important role in the surveying and mapping areas with its advantages of high sensitivity and low emission power. In recent years, LiDAR is used for a variety of scenarios including drones, unmanned vehicles, and even consumer electronics. The photon-counting detection system receives return points on the photon level, which retains rich detail than can be used for analysis. However, it is so sensitive that data are full of noise points. It is important to filter out noise in data preprocessing in order to prevent noise from affecting post-analysis. We proposed a convolutional neural network-based method to solve the noise filtering problem, which accepts varied input sizes and outputs noise labels for each point. Symmetry functions are designed to ensure consistent results on unordered data. Global features are extracted by the shared multilayer perceptron (MLP) and max pooling layer. Local features are extracted from each point and its spatial neighbors through the convolutional operator. We combine global and local features and build the network based on shared MLP to learn pointwise segmentation, which is the label of noise, from these combined features. The performance of our proposed method is evaluated by simulated data and MABEL data. It shows that our method can effectively eliminate most of noise points from data of photon-counting LiDAR.