ICRA2018点云相关论文汇总

1.Incremental Segment-Based Localization in 3D Point Clouds

• We propose an efficient method for localization based on 3D segment matching.
• A set of incremental algorithms for the normal estimation, segmentation and recognition steps is presented.
• Localization at 10Hz in urban driving environment is achieved (speedup of x7.1 over batch solution).
• The implementation is available open source:

 

2.SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud

• LiDAR point cloud segmentation is important for autonomous driving.
• We propose a CNN based model with high accuracy and real-time inference speed.
• Real-world data and simulated data are combined to train the model.
• Source code is released: https://github.com/BichenWuUCB/SqueezeSeg

3.Sampled-Point Network for Classification of Deformed Building Element Point Clouds

• Recognizing deformed objects is critical for disaster relief robots
• A deep learning method is proposed to classify deformed building elements from point clouds
• Synthetic deformations such as noise, bending, and truncation were applied to a CAD model database
• Robustness was achieved using improved regularization such as point sorting and resampling

 

4.GemSketch: Interactive Image-Guided Geometry Extraction from Point Clouds

• Interactive sketching for extracting geometries from point clouds
• Objects represented as generalized cylinders and cuboids
• Accurate within 5.66% Hausdorff distance for BigBIRD ground truth.
• Geometries produced used for object manipulation in clutter

5.Signature of Topologically Persistent Points for 3D Point Cloud Description

• We present the Signature of Topologically Persistent Points (STPP), a global descriptor for 3D point cloud data
• STPP uses persistent homology to encode a topological signature based on the birth-death pairing of the homology generators
• STPP can be computed quickly and efficiently, requires no preprocessing of the data, uses a single tuning parameter, and is competitive with state of the art global descriptors

 

6.AA-ICP: Iterative Closest Point with Anderson Acceleration

• We propose to apply Anderson acceleration (AA) for ICP instead of the commonly used Picard iteration procedure: xi+1 = ICP(xi)
• AA uses history of previous iterations to find a better guess for the next iteration
• AA-ICP was implemented as a part of Point Cloud Library (PCL)
• On real-world data AA-ICP converges significantly faster (up to 30-40%) compared to the unmodified ICP

 

7.Noise-Resistant Deep Learning for Object Classification in 3D Point Clouds Using a Point Pair Descriptor

• Novel 4D convolutional neural network architecture using descriptor values as input
• High robustness to noise and occlusion for realistic indoor point cloud data
• Superior performance for retrieval and classification on ScanNet and Stanford datasets
• More details at https://rebrand.ly/obj-desc

 

8.Real-Time Object Tracking in Sparse Point Clouds Based on 3D Interpolation

• New methods even to track sparse point cloud objects in real time is proposed.
• Proposed 3D interpolation for distributions, called EVD, augments information at unoccupied areas of target object.
• Through a coarse-to-fine grid search for real-time processing, the tracker find the optimal difference

9.Robust Generalized Point Cloud Registration Using Hybrid Mixture Model

• A novel point cloud registration method is proposed where the orientation information associated with each point is utilized.
• In the M-step of the algorithm, a closed-form solution to the scalar weighted rigid registration problem is proposed.
• The experiments demonstrate the proposed algorithm outperforms the other two under conditions of various noise levels, outliers percentages.

 

10.A General Framework for Flexible Multi-Cue Photometric Point Cloud Registration

• Register point clouds with no explicit correspondences
• We currently use intensity, range and normals for registration
• Easy to extend for using other cues
• Runs at framerate on TUM and KITTI data

 

11.Spatiotemporal Learning of Dynamic Gestures from 3D Point Cloud Data

• We demonstrate an end-to-end spatiotemporal gesture learning approach for 3D point cloud data using a new gestures dataset of point clouds acquired from a 3D sensor
• Nine classes of gestures were learned from gestures sample data through a 3D convolutional neural network that learns the spatiotemporal features in the data without explicit modeling of gesture dynamics.
• The developed model is able to classify gestures from the dataset with 84.44% accuracy.