LiDAR is a key sensor commonly used in unmanned vehicles. Smog is a trouble for vehicle-mounted LiDAR when unmanned vehicles operates in actual road environments. It leads to a significant reduction in the ability of LiDAR-based scene understanding for them. Thus, it is essential to recognize the smog existing in the road scene quickly and accurately. This paper proposes a fine-grained point cloud smog segmentation network (SmogNet) for unmanned vehicles. We adopt an effective graph convolution kernel based on attention to extract features layer by layer. The key of SmogNet is two manual features we design specially to characterize the geometric features of smog in point cloud. We evaluate SmogNet in challenging real road scenes with simulated smog. It performs better than competitive methods and it can be effectively generalized. We use the Focal Loss during the training of the SmogNet and improve the problems caused by the imbalance of sample categories effectively.
SmogNet: A Point Cloud Smog Segmentation Network for Unmanned Vehicles
29.10.2021
3793485 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
An Improved Algorithm on Point Cloud Optimization for Unmanned Aerial Vehicles
Trans Tech Publications | 2014
|Online Contents | 1995
Online Contents | 2015
|FEATURES - Smog-eating technology
Online Contents | 2000