Autoencoders Based on 2D Convolution Implemented for Reconstruction Point Clouds from Line Laser Sensors

Abstract

Gradual development is moving from standard visual content in the form of 2D data to the area of 3D data, such as points scanned by laser sensors on various surfaces. An effort in the field of autoencoders is to reconstruct the input data based on a trained neural network. For 3D data, this task is more complicated due to the demands for more accurate point reconstruction than for standard 2D data. The main difference is in shifting from discrete values in the form of pixels to continuous values obtained by highly accurate laser sensors. This work describes the applicability of autoencoders based on 2D convolutions for 3D data reconstruction. The described work demonstrates various autoencoder architectures. The reached training accuracies are in the range from 0.9447 to 0.9807. The obtained values of the mean square error (MSE) are in the range from 0.059413 to 0.015829 mm. They are close to resolution in the Z axis of the laser sensor, which is 0.012 mm. The improvement of reconstruction abilities is reached by extracting values in the Z axis and defining nominal coordinates of points for the X and Y axes, where the structural similarity metric value is improved from 0.907864 to 0.993680 for validation data.