Moving Object Tracking with Smooth Pan-Tilt

Abstract

In conventional real-time object tracking systems with pan-tilt camera, a camera in stopped for simple frame subtraction and prosessing. However, in order to perform fast and stable tracking, a system should process images captured during its rotation. We propose a method for extracting a moving target and estimating its image velocity from a sequence of images captured by a moving camera. Assessing the effect due to images plane rotation, we have found that on position shift a background can be well eliminated by translation of a previous frame and frame subtraction. We investigate about improvement of the accuracy of the velocity of a moving object, used for predictive control of a camera, by compensating non-linearity between velocity and optical flow in large displacement. We implemented the above methods in a real-time tracking system and conducted an experiment. As a result, the system could stably track a walking human in a room with complex background by processing non-interlace images of 1/60 [s] interval.