**4.3 Lab-based evaluation**

To evaluate the developed vision algorithm, we carry out a series of ground tests. Fig. 7 shows the test results. First, eight dots in the pattern are extracted from the image (Fig. 7-a). And Fig. 7-b shows the selected images for camera auto-calibration, and extracted camera pose is shown in Fig. 7-c,d.

While evaluating the performance of proposed visual localization method, we mainly consider two points: one is the pattern recognition rate, and the other one is the accuracy of pose estimation. For pattern recognition rate, we arbitrarily move the camera and take 306 images. 150 of them are correctly recognized, 85 are failed to recognize, 61 are blurred because of camera movement, and 10 do not include the full pattern. Except 61 blurred images and 10 of missed-pattern images, the recognition rate is about 64%. However, consider the fact that about half of the non-recognized images are rotated more than 40 degrees from the pattern, the recognition rate is about 81%.

To evaluate the accuracy of pose estimation, we fix the camera and locate the pattern at 12 known positions between 400mm to 700mm distance. Calculated average pose estimation error is 0.155mm and the standard deviation is 1.798mm.

Fig. 7. Process of pose estimation.
