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

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• The field is transitioning from 2D to 3D data, requiring advanced reconstruction techniques.
• Autoencoders are neural networks used for data reconstruction, but 3D data presents unique challenges.
• Reconstructing 3D data from laser sensors demands higher accuracy than 2D image processing.

Purpose of the Study:

• To investigate the effectiveness of 2D convolutional autoencoders for reconstructing 3D point cloud data.
• To evaluate various autoencoder architectures for 3D data reconstruction tasks.
• To enhance the accuracy and structural similarity of reconstructed 3D data.

Main Methods:

• Utilizing 2D convolutional autoencoders for processing and reconstructing 3D point cloud data.
• Implementing and testing diverse autoencoder architectures.
• Extracting Z-axis values and defining nominal X-Y coordinates for improved reconstruction.

Main Results:

• Achieved training accuracies ranging from 0.9447 to 0.9807.
• Obtained Mean Square Error (MSE) values between 0.059413 and 0.015829 mm, close to the laser sensor's Z-axis resolution (0.012 mm).
• Improved Structural Similarity Metric (SSIM) from 0.907864 to 0.993680 for validation data.

Conclusions:

• 2D convolutional autoencoders are applicable and effective for 3D data reconstruction.
• The proposed methods significantly enhance the quality and accuracy of 3D point cloud reconstruction.
• The results demonstrate the potential for high-fidelity 3D data representation using autoencoders.