Rigid Body Tracking

The recognition of non-deformable objects in still images and tracking in dynamic scenes is a demanding task. It requires the recognition of objects despite variations in their rotation scale and orientation. It also requires the detection of objects in cluttered and noisy scenes for most real world applications.

Invariant object detection has been the subject of much research due to the large number of useful applications such as robotics, automated manufacture and defence. One approach is using correlation methods; these can be implemented at the speed of light with the use of an optical correlator. A biologically inspired vision system is to use a logarithmic polar mapping to simulate the distribution of cells in the mammalian retina. This provides invariance to rotation and scale of the target object.

For object tracking we have been combining these techniques with background subtraction and Kalman filters for predictive tracking. It is hoped that this research will contribute towards a complete model of the mammalian vision system.

-Peter Bone.

This video shows an example of a rotating car mapped into log(r)-theta space making it rotation and scale invariant. The correlator then picks out the presence of the car, plus its scale and angle of rotation.