Abstract
The Global Positioning System (GPS) system is popularly used in an outdoor environment. A GPS tracking system usually displays the location of a moving object on a digital map. It is difficult to realize its actual situation without visual information. Therefore, cameras are incorporated with a GPS in order to provide a new GPS-based visual tracking service (GPS-VT). The tracking of an object with a GPS receiver can be performed not only on the digital map but also on the real-time camera image. GPS-VT was proposed in our previous study. It relies on the camera calibration for the transformation from a GPS coordinate to the coordinate on the camera image. Five point correspondences, i.e., a GPS coordinate and its corresponding image coordinate, must be established to generate the intrinsic and extrinsic parameters of the camera needed for coordinate transformation. There are two drawbacks in calibrating a camera. One is the difficulty in establishing five point correspondences manually for an uncalibrated camera. The other is the inconvenience of recalibrating a calibrated camera after adjusting its field-of-view (FOV). In this paper, an automatic calibration method and a rectification method are proposed to overcome separately the above-mentioned drawbacks. For an uncalibrated camera, the GPS moving vectors of the operator are matched with those moving vectors of objects in the camera image to locate the operator. The operator simply walks to five different locations and the point correspondences can be established automatically. For the rectification of a calibrated camera, two sets of feature points are extracted separately from the images before and after the adjustment of the FOV. The correlations of feature points of two images are analyzed by using a normal cross correlation (NCC) method to obtain the offsets of the adjustment in x and y directions. The offsets are used to rectify the image coordinate transformed from the GPS coordinate. Two experiments were also designed for both methods separately. For an uncalibrated camera, the locating errors are estimated and compared for the five point correspondences that are established manually or automatically. The difference in locating errors of manual and automatic calibration only totalled 2.6 pixels. The results show that the accuracy of the automatic calibration method is almost the same as that of the manual calibration. For the rectification of a calibrated camera, the differences in locating errors are estimated for various adjustment offsets in x and y directions. The errors of the proposed method totalled less than five pixels for the adjustment up to 200 pixels, without recalibrating the camera. The tests show that the result of the rectification method is close to that of manual calibration. Therefore, the two proposed methods overcome the problems of using the GPS-VT in an actual environment.