Lidar Point Cloud Quality Assessment in Autonomous Vehicles Using Deep Learning Techniques

Deep learning has been widely used in many aspects of computer vision, including point cloud (PC) quality assessment. It is important to determine the quality of a PC since it will affect the reliability and accuracy of perception in autonomous vehicles. A novel method based on deep learning is proposed for rating the quality of point cloud images using the PandaSet dataset. Instead of conventional methods that focus on geometric and attribute distortions relative to an ideal reference, this approach employs deep neural networks to forecast quality scores by directly exploiting features extracted from the inputs. Initially, three basic low-level characteristics—geometric distance, mean curvature, and gray level—are extracted from local patches of the point cloud data. These characteristics are then fed into a VGG16-inspired GQI_VGG16 custom-designed architecture neural network that learns how to predict each patch’s quality. The paper demonstrates the model’s performance through experimental results obtained using PandaSet data, showing high GQI scores and strong correlation coefficients with human ground truth assessments of image quality. The research provides an overview of recent progress made in assessing this type’s worthiness as well as its potential applications within self-driving cars or other similar systems.