Optical flow background estimation for real-time pan/tilt camera object tracking

As Computer Vision (CV) techniques develop, pan/tilt camera systems are able to enhance data capture capabilities over static camera systems. In order for these systems to be effective for metrology purposes, they will need to respond to the test article in real-time with a minimum of additional uncertainty. A methodology is presented here for obtaining high-resolution, high frame-rate images, of objects traveling at speeds ⩾1.2 m/s at 1 m from the camera by tracking the moving texture of an object. Strong corners are determined and used as flow points using implementations on a graphic processing unit (GPU), resulting in significant speed-up over central processing units (CPU). Based on directed pan/tilt motion, a pixel-to-pixel relationship is used to estimate whether optical flow points fit background motion, dynamic motion or noise. To smooth variation, a two-dimensional position and velocity vector is used with a Kalman filter to predict the next required position of the camera so the object stays centered in the image. High resolution images can be stored by a parallel process resulting in a high frame rate procession of images for post-processing. The results provide real-time tracking on a portable system using a pan/tilt unit for generic moving targets where no training is required and camera motion is observed from high accuracy encoders opposed to image correlation.