High Speed Target Tracking using Self Windowing

Summary

Self windowing is a simple target extraction algorithm that works by calculating the logical AND of the input image and the dilated image of the target calculated on the previous frame. It takes advantage of the high-speed nature of the processing: all the pixels of the new target image falls within the 1-pixel dilated image of the previous estimated target. Let the input image be f_k(i,j) at time k, and the target image be g_k(i,j), and we describe the algorithm as following ("|","&" are OR, AND operaration respectively).

g_k+1(i,j) = ( g_k(i,j) | g_k(i+1,j) | g_k(i-1,j) | g_k(i,j+1) | g_k(i,j-1)) & f_k+1(i,j)

We realized target tracking considering collision and separation between a target object and its environment using the algorithm on 1ms visual feedback system. Now target extraction by self windowing, collision detection and separation detection require 300us(23steps), 480us(1 summation calculation) and 1.7ms(3 summation calculations and 20 steps) respectively, and this combination requires 2.5ms(4 summation calculation and 43 steps). In future using high speed summation calculation circuits, such as a resistive network, the algorithm will be executed more quickly.

Movie

Reference

1. Idaku Ishii, Yoshihiro Nakabo, and Masatoshi Ishikawa: Target Tracking Algorithm for 1ms Visual Feedback System Using Massively Parallel Processing, IEEE Int. Conf. Robotics and Automation (Minneapolis, 1996.4.25)/Proc. IEEE Int. Conf. Robotics and Automation, pp.2309-2314