point cloud based object recognition in living...
TRANSCRIPT
HSR Pepper
3D sensor
The service robots
Decrease
Point Cloud Based Object Recognition in Living Environment○Takayuki Fujimoto, Takeshi Nishida, Kyushu Institute of Technology
Object Detection
BackgroundPoint Cloud
Processing method
Extraction of regionof interest
Landscape(Robot workplace)
Point Cloud Backgroundremoving
Clustering Calculation of surface normal vectors
Object Recognition
K-meansClustering
Test data Advance data
(a)Ball(b)Column(c)Cube
Evaluation function Result
Normal information
Original normal vectors
Clustering Reduced normalvectors
Matching
Cube・Normal vector templates
𝑈 = 𝒖𝑚 𝑚 = 1,⋯ ,𝑀 = 128
・Weight of direction
𝜂𝑚(𝑓): 𝑓 = Ball, Column, Cube
The service robots are equipping multi sensors, such as laser range finder and cameras, for the environment measurement. It can obtain data called point cloud by using 3D sensor. It is expected that it can be used for object recognition and object grasping.
In general, information size of the normal vectors of PC is large, and the calculation cost of the recognition procedure becomes a bottleneck of real time processing of the robot.
・Collection of characteristic vectors 𝐷(𝑗) = 𝒅𝑙(𝑗)|𝑙 = 1,⋯ ,𝐾
・Weight of number of surface normal vectors 𝜀𝑙(𝑗)
𝐸 𝑗 = 𝑙=1
𝐾
𝜀𝑙𝑗𝜂𝑚𝑓min𝒖𝑚cos−1 𝒅𝑙
𝑗𝒖𝑚
Search of the shape that minimizes 𝐸(𝑗)
Characteristic vectors
Correct answers Accuracy rate [%] Computation time per frame [s]
10 96 80.00 0.231
20 110 91.66 0.27
30 114 95.00 0.347
Table 1 Correct recognition ratio vs. the number of Characteristic vectors
Experiment
chips teacup box1 box2 tissue ball
Cylinder Cylinder Box Box Box Ball
Determined using evaluation function