Uma abordagem para Reconhecimento de Objetos em 3D a partir de nuvens de pontos capturadas com LiDAR utilizando Aprendizagem Profunda

Abstract

This work proposes the use of Deep Learning in recognition of 3D objects captured with Light Detection and Ranging (LiDAR), including a pre-processing step. In addition, it uses a real- world dataset that contains three classes. The pre-processing step is proposed to normalize the number of points of each captured cloud, bypassing the dimensionality discrepancies between the proposed dataset and the existing sets in the literature, such as ModelNet. Thus, cross-validation with three partitions is performed in the classification stage, aiming to measure the performance of the networks in different evaluation metrics from different arrangements of the set. In the first experiment with ten selected networks, the best performing network was RS-CNN, reaching 98.64% of average accuracy, being, however, one of the slowest, with an average time of 21.58 ms for each cloud, while the fastest network was LDGCNN reaching 25μs of average time with an accuracy of 92.50%. In the second experiment, the Lidar3DNetV1 network was optimized and compared with the original network, where a significant performance improvement was observed while maintaining similar evaluation metrics, reaching 84.80% accuracy in one of the tests subsets with 180 μ s average processing time.