• Menu
  • Introduction
  • System Structure
  • Experiment
  • Presentation
  • Reference
  • About Us
  • Menu
  • Introduction
  • System Structure
  • Experiment
  • Presentation
  • Reference
  • About Us

Implementation Plan 

Utilize and Collect our own data  

  • Using Kinect 
  • v2 Space translation

Optimization  

  • Adding temporal data 
  • Other adaptation measures

Visualization and Presentation  

  • In 2D/ 3D  
Real Scene

Real Scene

Detect

Detect

Output

Output

Experiment - Kinect V2

Original for XBOX  

  • More affordable compared to LiDAR 
  • Computer vision development communities (C++, C#)  

Structure  

  • RGB camera 
  • Depth camera 
  • Infrared camera  
Depth image

Depth image

RGB image

RGB image

Auto rotation

Auto rotation

Kinect V2

Kinect V2

Experiment - Camera Calibration

Intrinsic matrix  

  • From photo coordinate to camera coordinate 
  • Differ slightly with different Devise

Extrinsic matrix   

Instrinsic matrix

Instrinsic matrix

  • From camera coordinate to world coordinate 
  • Define world coordinates
  • Neglect shift
Extrinsic matrix

Extrinsic matrix

Mapping the relation with rotation  

  • Rightward +x, Forward +y, Upward +z
  • Roll(x)/Pitch(y)/Yaw(z) 
  • Differ with rotation angle

Experiment - optimization

Temporal data  

  • Time relationships between frames 
  • Duplicate output / Wrong classification/ Falsely detected

(1) Non-maximum Suppression  

  • Using Integration over Union(IoU) as threshold
  • Eliminate same class duplication 
  • Correct different class misclassification

(2)  Timeout measures

  • Assure existence 
  • Delete false misunderstood ones
without NMS

without NMS

with NMS

with NMS

Experiment - Other adaptations

Desk/Table merging   

  • High similarity 
  • Differ with furniture style Model misconstrue

Toilet/Bathtub deletion  

  • Not used in presentation scenes 

Page End

Previous Next

NTHUCS CVLAB 2020

Website Templates

created with

Website Builder Software .