Abstract
In this study, we propose an architecture for estimating motion vectors by searching for one neighbor pixel in high-speed images and a machine learning algorithm that uses the estimated motion vectors. In high-speed imaging, the motion of pixels between frames is considerably small. Our architecture estimates motion vectors by assuming that the pixels move less than one pixel between frames. We verified that our method could classify images into two classes, i.e., dangerous (something is approaching) or safe (others), by employing a simple perceptron after extracting the features of the estimated motion vectors using a method based on Poggio's HMAX (Hierarchical Model and X) model. We used the target images captured by an in-vehicle camera for learning and verified that another set of images could be classified using our method. We confirmed that the proposed architecture can estimate motion vectors using a small number of operations and perform classification based on machine learning.
