Cleaning robot

Robots: the more the merrier

Object detection or SURFing

This is the first attempt in detecting objects. In this experiment we’ll focus on PET bottles only. A future experiment should emphasize the feasibility of glass/cup detection.

We’re using SURF (Speeded Up Robust Features) features extracted from training images taken from various angles.

Our object detector is feeded with frames captured with an old Creative WebCam Live! device, which focuses on a scene with 3 objects, noted from left to right as follows:

  • – object X: a 2.0L bottle, very similar to others, but with a slightly different shape, which should remain undetected throughout the experiment;
  • – object A: a 2.5L empty bottle; this should be detected;
  • – object B: a 2.5L half filled bottle; this should also be detected.

Test 1 – horizontal angles

Training: 4 angles per object

  • 1 x Frontal
  • 2 x Profile
  • 1 x Rear

Description: tests from few horizontal angles

Result: no detection, except when angles are very closely matched

 

Test 2 – horizontal angles

Training: 8 angles per object

  • 1 x Frontal
  • 2 x Three-quarter front
  • 2 x Profile
  • 2 x Three-quarter rear
  • 1 x Rear

Description: tests from few horizontal angles

Result: both A & B detected; overall a good detection in horizontal angles;

 

Test 2.1 – vertical angles

Training: same as in Test 2

Description: vertical angles test

Result: almost no detection in vertical angles (need more vertical angles in training images)

 

Test 3 – scalar invariance

Training: same as in Test 2

Description: testing scalar invariance

Result: 75% success, this could definitely be improved

 
Conclusion: SURF is a top candidate for our robot’s object detection capabilities in real time.

Trackback from your site.

Leave a comment