Lidar Point Cloud Semantic Segmentation Using SqueezeSegV2 Deep Learning Network

Abstract

LiDAR point cloud semantic segmentation has become a crucial research area, particularly for applications in autonomous driving, robotics, and environmental mapping. However, challenges such as data sparsity, noise, occlusions, and varying point densities hinder the performance of segmentation models. This paper explores the use of the SqueezeSegV2 deep learning network to enhance LiDAR point cloud segmentation. SqueezeSegV2 improves upon its predecessor by incorporating domain adaptation techniques, advanced loss functions, and optimized convolutional operations to enhance robustness against noise and class imbalance. The model is trained and evaluated using the PandaSet LiDAR dataset, demonstrating significant improvements in segmentation accuracy. Key challenges such as scalability, domain adaptation, and dynamic scene understanding are discussed, along with potential strategies for further advancements. The global accuracy, mean accuracy, mean IoU, weighted IoU, and mean boundary F1 score for the entire Pandaset dataset are 0.9084,0.6268,0.5596,0.8362, and 0.7496, respectively. In addition, the accuracy, IoU, and the mean boundary F1 Score of the “Car” class are 0.9193,0.8064, and 0.9549, respectively. The results indicate that SqueezeSegV2 is an efficient and effective solution for real-time LiDAR segmentation, making it a valuable tool for autonomous systems.