3D Object Segmentation with Quasi Solid-state LiDAR
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Current, most 3D semantic segmentation models for autonomous driving are mainly trained on spinning Light Detection And Ranging (LiDAR) data because spinning LiDAR sensors have been one the most popular sensors for autonomous driving vehicles and there is an abundance of spinning LiDAR dataset available to the public. However, spinning LiDAR sensors are costly and requires large amounts of energy to operate. The newly emerged quasi solid-state LiDAR sensors are more cost efficient and require lower amount of energy to operate on autonomous driving vehicles. If we reuse the current models pretrained with spinning LiDAR data on quasi solid-state LiDAR data, its performance is below expectation. Currently there are not enough quasi solid-state LiDAR data to train 3D semantic segmentation deep learning models effectively, and the data pattern for quasi solid-state LiDAR is mostly different from the spinning LiDAR data.
This research will first develop a visualization tool and evaluate the existing 3D semantic segmentation models that are pretrained with spinning LiDAR data on some small scaled quasi solid-state LiDAR data. The performance of the model is under expectation, which calls for the retraining of the 3D semantic segmentation model on quasi solid-state LiDAR. The model chosen is SPVNAS. Since there are few publicly available large scaled quasi solid-state LiDAR datasets, several approaches are taken in parallel to synthesize the suitable dataset for training. The final approach decided was rich point subsampling, which takes a reconstructed scene of rich point cloud, filter out the point that are not in the vehicle’s field of view, and sample the points according to the data pattern for quasi solid-state LiDAR data. The processed data is fed into the SPVNAS model for training, validation, and testing. This final model for 3D Semantic Segmentation is the main product of the research.
Description
Keywords
3D Semantic Segmentation, quasi solid-state LiDAR, autonomous driving, SPVNAS, SPVCNN, rich point subsampling