A Semantic Segmentation Method of In-Vehicle Small Targets Point CloudBased on Improved RangeNet++ Loss Function

Abstract

<p indent="0mm">Pedestrians and other moving small-targets plays an important role in road scene point cloud. However, they perform poorly for real-time semantic segmentation. Therefore, a small-target semantic segmentation method based on RangeNet++ and optimization of loss function is proposed. Firstly, the Focal Loss function is used to improve the original RangeNet++ which uses cross entropy as loss function. Then, the weight of some rare, important moving small-targets in Focal Loss function is adjusted. The proposed method can improve the detection and segmentation accuracy of important moving small-targets categories such as pedestrians with fewer training sessions under the DarkNet21 which has fewer convolutional layers. Experiments on the SemanticKITTI dataset show that compared with the original RangeNet++, the improved method uses fewer convolutional layers backbone and fewer training sessions, but achieves higher accuracy and precision in semantic segmentation of moving small-targets.