论文速读 – Automatic Labeling to Generate Training Data for Online LiDAR-based Moving Object Segmentation
先赞后看，养成好习惯。有帮助的话，点波关注！我会坚持更新，感谢您的支持！
参考：
1. Auto-MOS
2. Mapless Online Detection of Dynamic Objects in 3D Lidar
3. Rigid scene flow for 3d lidar scans
4. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments
5. ERASOR
6. Multiple Dynamic Object Tracking
7. LMNet
8. KPConv: Flexible and Deformable Convolution for Point Clouds
1. Summary
Supervised networks for deep learning require a large amount of labeled data, which is expensive. This paper proposes to automatically generate labeled training data offline to improve the performance of the learning-based MOS (Motion Segmentation) system. It first utilizes occupancy-based dynamic object removal to roughly detect possible dynamic objects. Extract again possible objects and track them using Kalman. Based on the tracked trajectories, actual moving objects such as moving cars and pedestrians can be marked. Conversely, non-moving objects are marked as stationary. The neural network we trained to automatically generate annotations achieves the same performance as manual annotations.
2. Introduction
Main work:
1) Our method can generate better labels for point cloud motion segmentation tasks.
2) Based on the labels we generated, the network achieves similar performance; after adding manually labeled labels, the network performance will improve.
3) We can generate effective tags for radars with different patterns and different environments.
Related work:
基于几何的方法: Inspired geometric Mapless Online Detection of Dynamic Objects in 3D Lidar , estimated rigid body motion flow Rigid scene flow for 3d lidar scans深度学习方法: SuMa++ improves SLAM performance, LMNet离线标签生成: Generate a clean static map, mainly including Removert, ERASOR, and methods of occupying the map.
Our method不需要真实位姿和追踪方式distinguishes between moving and non-moving objects.
3. Model method
A total of 5 modules are included.
- First use the radar odometer or SLAM (
SuMa) to calculate the pose between adjacent frames. ERASORDense moving object detection using the clean map method ( ).- Clustering ( ) based on the dense moving target in the previous step
HDBSCANdivides the point cloud into disjoint instances, each instance includes center, orientation angle, length, width, and height, and confidence. - Then use multi-target tracking for instance association ( Multiple Dynamic Object Tracking ), and generate the final instance label based on the tracking trajectory. EKF will help a frame that was not detected and re-assign the instance.
- Training Neural Network LMNet
4. Experimental evaluation
Evaluation index : IoU
SementicKitti dataset :
- 00-07seq + 09-10seq ----> training
- 08 ----> Verification
- 11-21 ----> test
result
Generalization ability :
3 datasets testedApollo-ColumbiaParkMapData 、MulRan、IPB-Car