Classification

Classification (=Object Recognition)

The Udacity RoboND use an RGB-D sensor , eg. color. It does not fit to our problems.

• Pick list.yaml : label, color

• outpul.yaml : position(x,y,z)

  RoboND-Perception-Exercises/Exercise-3/ : Object Recognition with Python, ROS and PCL

1. SVM

1.1 Capture Object Features

1.2 Training

5. Classification

군집화된 포인트들을 사람, 자동차, 가로수 등으로 구분

5.1 Machine learning based approach

RBF (Radial Basis Function) 커널을 사용한 SVM (Support Vector Machine)

GM-PHD 필터를 이용한 보행자 탐지 성능 향상 방법(2015)

5.2 Deep learning based approach

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

Classification

1. List

2. Paper

• [추천] SqueezeSeg demo: CNN for LiDAR point cloud segmentation : Youtube, [깃허브]
• Tracking People in 3D Using a Bottom-Up Top-Down Detector: ICRA2011
• DepthCN: Vehicle Detection Using 3D-LIDAR and ConvNet
• Deep Semantic Classification for 3D LiDAR Data, Ayush Dewan
• 3D Object Recognition based on Correspondence Grouping: This tutorial aims at explaining how to perform 3D Object Recognition based on the pcl_recognition module.
• LiDAR 센서 정보를 활용한 데이터 마이닝 기법 기반의 수상함정 표적 식별기법 제안: 2017, 엄준호
• A General Purpose Feature Extractor for Light Detection and Ranging Data: 2010
• Multiple Sensor Fusion and Classification for Moving Object Detection and Tracking: 2015
• Multisensor Online Transfer Learning for 3D LiDAR-based Human Detection with a Mobile Robot: 2018, L-CAS
• LIDAR-based 3D Object Perception: 2015
• Instant Object Detection in Lidar Point Clouds: 2017
• Lidar Based Object Detection Near Vehicle: 2017
• A Fully Convolutional Network for Semantic Labeling of 3D Point Clouds: 2017
• Self-supervised online learning of appearance for 3D tracking: 2017
• 3D Lidar-based Static and Moving Obstacle Detection in Driving Environments:an approach based on voxels and multi-region ground planes: 2016
• A real-time LIDAR and vision based pedestrian detection system for unmanned ground vehicles: 2015
• Object detection in 3D point clouds: 2016, 3장 참고
  • Downsampling
  • ground Segmentation
  • Clustering
  • Tracking / Kalman filter
• Object Classification using CNN-Based Fusion of Vision and LIDAR in Autonomous Vehicle Environment: 2018
• Laser-Based Detection and Tracking of Dynamic Objects: 2014, 옥스포드, 183p
• No Blind Spots: Full-Surround Multi-Object Tracking for Autonomous Vehicles using Cameras & LiDARs: 2018
• [추천]3D-LIDAR Multi Object Tracking for Autonomous Driving: 2017, 140page, 석사학위 논문
• Object tracking and state estimation in outdoor scenes based on 3D laser scanner: 2016
• Real-Time Deep ConvNet-Based Vehicle Detection Using 3D-LIDAR Reflection Intensity Data : 2017
• 3D Scanning: A Comprehensive Survey: 2018
• 원(Raw) LiDAR 자료 구조를 이용한 LiDAR 자료의 분리(Segmentation)에 관한 연구: 2005, 서울대 유기윤, 지구환경시스템공학부
• 깊이 영상 기반의 보행자 검증을 적용한 보행자 검출 성능 분석: 2014, DGIST 임영철, 강민
• GM-PHD 필터를 이용한 보행자 탐지 성능 향상 방법: 2015, 서울대 서승우
  • 기반 논문 : Towards 3D Object Recognition via Classification of Arbitrary Object Tracks: 추천,Alex Teichman 2011
  • Alex Teichman 박사 학위 논문 : USER-TRAINABLE OBJECT RECOGNITION SYSTEMS VIA GROUP INDUCTION

[3] D. Prokhorov, “A Convolutional Learning System for Object Classification in 3-D Lidar Data,” IEEE Trans. Neural Networks, Vol. 21, No. 5, pp. 858-863, May 2010.
[4] S. Awan, M. Muhamad, K.Kusevic, P. Mrstik and M. Greenspan, “Object Class Recognition in Mobile Urban Lidar Data Using Global Shape
Descriptors,” 2013 International Conference on 3D Vision, Seattle, WA, USA, June 2013.

Premebida[5]의 연구에서는 불완전한 3D 센서인 4채널 LIDAR를 이용하여 도로 환경에서 보행자 탐지 알고리즘을 제안하였다. 데이터의 선형성과 원형성을 바 탕으로 한 특징을 사용하여 보행자 분류를 하였지만, 포인트 수가 적어 가까운 지역에서만 인식 가능하며, 완전한 3차원 데이터에 이 방법을 적용시키기에는 계산시간이 오래 걸린다는 한계가 있다.

[5] C. Premebida, O. Ludwig and U. Nunes, “Exploiting LIDAR-based Features on Pedestrian Detection in Urban Scenarios,” 12th International
IEEE Conference on Intelligent Transportation Systems, St. Louis, MO, USA, Oct. 2009.

Navaro-Serment[6]는 각 물체에 해당하는 3D 포인트 클라우드(Pointcloud)를 두 다리와 몸통에 해당하는 3가지 부분으로 나눠 각 부분의 분산행렬을 특징으로 분류하는 방법을 제시했다. 하지만 먼 거리의 사람데이터는 포인트 수가 부족하여 다리와 몸통을 구분하기 어렵다는 문제로 인해 인식 거리가 짧고 정확하지 않다는 문제점이 있다.

[6] L. E. Navarro-Serment, C. Mertz, and M. Hebert, “Pedestrian Detection and Tracking Using Three-Dimensional LADAR Data,” International Conference on Field and Service Robotics, 2009.

3. Article (Post, blog, etc.)

3. Tutorial (Series, )

4. Youtube

6. Material (Pdf, ppt)

7. Implementation (Project)

• ROS obstacle detection for 3D point clouds using a height map algorithm
  • ROS velodyne_height_map
  • 깃허브

8. Research Group / Conference