Traffic sign extraction using deep hierarchical feature learning and mobile light detection and ranging (LiDAR) data on rural highways

Abstract

The application of deep learning techniques on point cloud data holds significant promise for efficient data segmentation and classification of traffic signs. This study proposes modifications to the PointNet++ neural network to improve performance on outdoor scenes. In addition, the method leverages the use of local geometric features in the training process. Several models with different combinations of geometric features and proposed changes were trained using labeled data from seven highway segments in Alberta, Canada. The results indicate that the proposed models have improved performance in accuracy and processing times compared to previous studies on sign detection using point cloud data. The overall per sign detection performance shows a 99.2% recall (98% per point) and a 98% F1-score (97% per point). Overall, the inclusion of z-gradient significantly increased sign detection in terms of precision, recall, and F1-score, by 9%, 4.9%, and 7.1%, respectively, allowing the model to yield notable performance improvements for outdoor scene recognition. Ablation tests were performed to validate the performed PointNet++ modifications. The modified PointNet++ was compared with SqueezeSegV2, a state-of-the-art neural network designed for road-object segmentation, and showed improved performance. A comparison was also made with existing sign detection methods on the Paris-Lille-3D benchmark, finding higher recall rates than existing studies. The proposed approach suggests that with adjustments, the PointNet++ neural network architecture can achieve remarkable results on large metric scale scenes for sign extraction using point cloud data.