Ampliação da Resolução de Sensores LiDAR Utilizando Redes Neurais Artificiais

Abstract

Environment perception with adequate resolution is essential for the safe navigation of autonomous robots. The LiDAR sensor is frequently used for precise distance measurements; however, the number of points in its scan is limited by its hardware design, which can compromise obstacle detection and mapping. This work addresses the problem of low point density in 2D LiDAR sensor, developing and evaluating low-complexity Artificial Neural Networks for resolution upsampling. The main objective is to double the angular resolution of a LiDAR sensor’s scan, with data acquired in a simulated environment using the Gazebo simulator and the TurtleBot3 robot. The measurements were preprocessed and split to train two architectures: a Multilayer Perceptron, using a windowing technique, and a One-Dimensional Convolutional Neural Network. The models were trained on a subset of data, simulating a lower-resolution sensor, with quantitative evaluation performed through the analysis of the Cumulative Distribution Function and the Kolmogorov-Smirnov statistical test. For the qualitative evaluation, a visual analysis of the reconstructed signal was conducted by plotting the results in polar coordinates. The results demonstrated that both models learned the spatial patters and were capable of reconstructing the missing measurements while maintaining the statistical characteristics of the original data. The Multilayer Perceptron architecture showed slightly superior performance compared to the One-Dimensional Convolutional Neural Network, with more stable training losses and less differences in the Cumulative Distribution Function analysis. We concluded that the use of low-complexity Artificial Neural Networks is a viable and effective approach for upsampling 2D LiDAR sensor data, offering a new method to enhance the perception of mobile robots with limited resources.