RWDA

Real World Data AnalysisFinal Project

Object Classification on3D Environments

Abraham Monrroy Cano2014/01/29

Hypothesis

Objects in three-dimensional space can be classified accurately through the analysis of Point Cloud data

acquired by LiDARs

What is a LiDAR?

A LiDAR (Light RaDAR) is a sensor that measures distance by

illuminating a target with a laser and analyzing the reflected light.

X Y Z R

77.465 12.068 2.860 0.00

77.502 12.323 2.863 0.00

45.525 7.448 1.769 0.00

45.517 7.594 1.770 0.00

77.819 13.253 2.878 0.00

77.885 13.516 2.882 0.00

77.945 13.779 2.886 0.00

What is a PointCloud?

• A point cloud is a set of data points in some coordinate system.

• Are intended to represent the external surface of an object.

Analyzed Dataset

KITTI Dataset is a collection of:

• Velodyne Points (Point Clouds)

• Stereo Images

• GPS localization points

Is free and available for academic use only

Workflow

Data structure

Analyze Train Test Result

Data structure

Input• Velodyne Points in binary format• JPEG Images (used only for reference visualization)• Labels in XML format for each frame in the image coordinate

system

My Processed Output• Velodyne Points in Matlab’s structure array and labeled.• Extracted Velodyne points for each labeled object (extracted

from 2D to 3D)• Generated 2D image from 3D points

Analysis

Approaches:1. Using the 3D projection in an image. SURF, ORB, HOG features

• Result: Failed. Since the image is not complex enough, there were not enough features to perform data analysis.

2. Using the 3D points and try to obtain descriptive features.

• Result: Success, there are a few feature descriptors for 3D point clouds.

Features

• Extract local features from the now segmented velodyne points.

Feature NameSupports Texture / Color

Local / Global / Regional

Best Use Case

PFH No L

FPFH No L 2.5D Scans (Pseudo single position range images)

VFH No G Object detection with basic pose estimation

CVFH No RObject detection with basic pose estimation, detection of partial objects

RIFT Yes LReal world 3D-Scans with no mirror effects. RIFT is vulnerable against flipping.

RSD No L

NARF No L 2.5D (Range Images)

FPFH (Fast Point Feature Histogram)

Input Format

• A point cloud consisting of a set of oriented points P. Oriented means

that all points have a normal n.

• This feature does not make use of color information.

Output

• A 33 bin histogram stores in a vector for each point.

Radu Bogdan Rusu, Nico Blodow, Michael BeetzTechnische Universitat Munchen

PointCloud LibraryC++ API

MatlabMAT

PCD Files PCL CSV

SVM Training / Tests / Results

SVM Type Accuracy

Multiclass Model 72.8627%Car binary Model 85.8886%Pedestrian binary Model 96.3387%Van binary Model 90.1602%

Training set #Objects

1 572

2 56

3 705

4 1754

Todo Work

• Test with other features extractors.

• Train with more observations

• The development of a detector in the PointCloud would be useful for other

purposes such as perception systems.

ご清聴ありがとうございます