Robust Real-time Moving Object Segmentation in LiDAR Point Clouds
2024 (English)Independent thesis Advanced level (degree of Master (Two Years)), 28 HE credits
Student thesis
Abstract [en]
Robust real-time moving object segmentation (MOS) in LiDAR point clouds is crucial for generating high-quality 3D maps and gaining vital information for path-planning and object avoidance. This thesis focuses on MOS within forest environments, particularly for drone applications, where the available computa- tional resources are limited. The study evaluates state-of-the-art methods with a focus on mean Intersection over Union (mIoU), the real-time capabilities when run on an Nvidia AGX Orin, and their generalization across different environ- ments. It also investigates the effectiveness of synthetic data generation and trans- fer learning for adaptation to forest scenarios, where labeled data is scarce.
A variety of methods for MOS in LiDAR point clouds are compared, encom- passing the most common data representations, including range-view (LMNet, MotionSeg3D), bird’s-eye-view (MotionBEV, M-Detector), and voxel-based (Map- MOS, 4DMOS, Dynablox) representations. Both learning and non-learning meth- ods are considered. The non-learning-based methods, Dynablox and M-Detector, meet the real-time threshold of 100 ms. Furthermore, the learning-based meth- ods, LMNet and MotionBEV, achieve real-time performance using bird’s-eye and range-view approaches, respectively. In contrast, the learning-based volumetric methods, 4DMOS and MapMOS, are an order of magnitude away from the real- time threshold, while MotionSeg3D is marginally close.
The results indicate that range-view and bird’s-eye-based methods exhibit poor generalization capabilities from urban to forest environments, while volumetric and non-learning-based methods adapt significantly better. The non-learning- based method Dynablox shows the best generalization performance, with an mIoU of 72.4% over all the test sequences from a forest environment.
The study underscores the significant improvements obtained through synthetic data generation and fine-tuning. These techniques substantially enhance the per- formance of the fine-tuned models, demonstrating the value of synthetic datasets for model training in data-scarce environments like forests. For instance, fine- tuned LMNet improves its mIoU from 12.9% to 61.3%, MotionBEV from 2.3% to 73.2%, and 4DMOS from 45.6% to 58.6%.
Place, publisher, year, edition, pages
2024. , p. 84
Keywords [en]
moving object segmentation, lidar, point cloud, real-time, forest, drone, jetson, synthetic, simulator
National Category
Computer graphics and computer vision
Identifiers
URN: urn:nbn:se:liu:diva-206867ISRN: LiTH-ISY-EX--24/5699--SEOAI: oai:DiVA.org:liu-206867DiVA, id: diva2:1892088
External cooperation
Deep Forestry AB; AI Sweden
Subject / course
Computer Engineering
Presentation
2024-06-13, ISY Visionen, Linköping, 08:15 (English)
Supervisors
Examiners
2024-08-262024-08-252025-02-07Bibliographically approved