Image motion analysis plays an important role in the everyday life of both humans and animals. Motion detectors based on gradients, correlations and energies have been proposed to account for the local motion processing in visual cortex of human. The motion energy model is one of the most widely accepted and biologically plausible models for modeling neurons in human brain. However, the computational costs limit the accuracy for image velocity extraction when utilizing a population of motion neurons to disambiguate the uncertainty of single neuronal responses....[ Read more ]

Image motion analysis plays an important role in the everyday life of both humans and animals. Motion detectors based on gradients, correlations and energies have been proposed to account for the local motion processing in visual cortex of human. The motion energy model is one of the most widely accepted and biologically plausible models for modeling neurons in human brain. However, the computational costs limit the accuracy for image velocity extraction when utilizing a population of motion neurons to disambiguate the uncertainty of single neuronal responses.

This thesis generalizes the motion energy model proposed by Adelson and Bergen and allows linear combinations of filters obtained from spatial and temporal impulse responses other than only sums and differences. The generalized motion filter is obtained by re-expressing the complex valued temporal filter as the real and imaginary parts and adding a phase shift on the imaginary path. This proposed phase shift motion energy model is approximately spatio-temporal orientation tuned where the orientation in space-time is adjusted by the newly introduced phase parameter. The phase shift model suggests an alternative way to construct the motion energy neuron population by varying phase values instead of varying temporal frequencies, which is referred to as the phase tuned population instead of frequency tuned population.

The phase tuned population greatly reduces the computational costs by replacing the spatio-temporal filtering bank with a single spatio-temporal filter connected to different linear combinations of responses on the real and imaginary paths. This novel phase shift mechanism for population construction also sheds light on the functional role for interpreting the general believed rule of biological systems that similar mechanisms are utilized when processing information from different sources. The connection between the phase tuned motion energy neuron and the phase tuned disparity neuron is established by an analogy between the disparity model of combining spatially Gabor-filtered versions of left and right images and the motion model of combing spatially Gabor-filtered versions of two delayed input images with different phase shifts.

A natural idea following the motion representation by phase shifts is to represent the spatial motion contrast with the phase difference between motion neuronal responses of different spatial regions. This idea facilitates the modeling of motion neurons directly sensitive to motion contrast instead of absolute motion. Based on the measurement of phase differences, the differential motion opponency is constructed to model the enhanced and suppressive interaction between center and surrounding regions and to detect motion popout.

We validate our phase shift model with experiments by using both synthetic and real images in both 1-D and 2-D. With a small amount of spatial pooling, the performance of velocity estimates when using the phase tuned population exceeds that when using the frequency tuned population. Utilizing a confidence measure given by the normalized peak response with the average response across the population, the erroneous motion estimates can be reliably rejected.

Finally, we demonstrate the efficiency and effectiveness of the motion estimation in the MultiMap vision system we developed. The system computes the population responses of phase tuned motion neuron population and the velocity map from inferred phase values.