Car Recognition

1
Car Recognition
Fourth Quarter Report
Drew Stebbins, Pd. 6
2
Background

• License plate tracking and recognition a solved
  problem
• Need for complex vehicle recognition technology
• Object detection techniques
• Hough transform
• Neural networks
• SIFT keypoint matching

3
Problem statement

• Car recognition
• fast moving objects
• multi-coloured
• complex shapes with many features
• must identify in real-time, or near real-time

4
Hough transform

• Detecting simple shapes
• Region of interest identification
• multiple passes

5
(No Transcript)
6
Neural networks

• Neural Nets
• good at pattern matching, object categorization
• simple OCR with MNIST database
• backpropagation learning algorithm

7
Graphical user interface

• Streaming Input
• v4l/GTK viewing application
• button with imageprocessing callback

8
SIFT keypoint matching

• Keypoint detection
• scale-space extrema detection
• keypoint localization
• magnitude and orientation assignment
• keypoint descriptor

9
(No Transcript)
10
(No Transcript)
11
SIFT keypoint matching

• Matching to database images
• nearest-neighbour identification based on
  keypoint description

12
Review of progress

• Region of interest identification
• Hough transform circle detector
• Graphical user interface
• Neural networks
• optical character recognition
• SIFT keypoint identification

13
Conclusions

• Hough transform is fast, yet not suitable for the
  detection of complex objects
• Conventional neural networks are only at good at
  object categorization as their training set
  allows
• Local image descriptor methods such as SIFT
  keypoint matching are currently the best
  general-purpose algorithms for the detection of
  complex objects in images