A Comparison of Point Cloud Segmentation Models for Road Unevenness Detection and Smooth Autonomous Driving

Detecting uneven and imperfect road segments is crucial for self-driving vehicles to ensure a safe and comfortable autonomous driving experience. This paper aims to explore the PointNet architecture across its many evolutions and examine how each generation performs when applied to road irregularity point cloud segmentation. The architectural differences between PointNet, PointNet++, and PointNeXt are explored, highlighting how each improves upon its predecessor. Customized versions of PointNet and PointNeXt are then developed and discussed, each obtaining overall training accuracies of 91% and 99.1%, respectively. These tailored implementations of PointNet and PointNeXt are then compared with an existing PointNet++ model for road unevenness segmentation, with our custom PointNeXt model demonstrating a dramatic 76% decrease in model size and a 0.7% jump in accuracy. Our findings highlight the significant improvements across PointNet generations, especially when used for LiDAR point cloud segmentation tasks. Moreover, this study demonstrates the superiority of PointNeXt for real-time road unevenness detection, opening avenues for further research into better autonomous driving solutions.