Lidar3dnet - Abordagem inteligente para classificar objetos 3d com base em nuvens de pontos reais e sintéticas

Abstract

Use point clouds of datasets from the literature and generated in simulators and avoid collection time costs. In this way, there is access to an unlimited and organized amount of point clouds, an ideal scenario for deep learning networks. However, models that use convolutional neural networks and synthetic data have limitations when applied to real-world data. This work proposes a multilayer perceptron (MLP)-based method for classifying 3D objects based on real point clouds, acquired with the Light Detection and Ranging (LiDAR) sensor, and a pre-processing step. In addition, two dataset are proposed, with real and synthetic data, and the research uses dataset ModelNet to encourage comparisons. The proposed neural network, Lidar3DNet, has an architecture based on MLP, as it supports a high amount of data as input. To provide the best performance of the selected neural networks, a pre-processing step is proposed with two treatments, normalization of points and adjustment of the point cloud in the 3D Cartesian plane, to overcome discrepancies in the distribution of points in the point clouds. The research consists of four steps: (1) the nine networks with the best performance in the dataset ModelNet40 are selected and all architectures are configured and executed. then (2) the use of the ModelNet40 dataset is replaced by the proposed dataset with point clouds, varying the application of the original and pre-processed point clouds. Next, (3) the dataset with real point clouds is replaced by the dataset with synthetic point clouds, also with original and pre-processed point clouds. In these stages, five evaluation metrics were applied together with two statistical evaluation methods. Finally, (4) a variation in the number of points of the point clouds is implemented to evaluate the interference of the number of points in the performance of the Lidar3DNet network, with CPU and GPU tests and with real and synthetic data. The pre-processing step provided an increase in the performance of the selected neural networks with gains of 25% in the best case. The proposed network reached 98.33% accuracy and a test time of 88 µs with synthetic data and with real data it reached 98.47% and 125 µs in accuracy and test time, respectively. Evaluation of the variation in the number of points of point clouds for the classification of 3D objects indicated that the use of GPU is valid from point clouds with more than 1024 points