Automatic Object Detection for the Testing of Displays

Abstract

Testing of car dashboard displays has traditionally been an expensive and involved process. This thesis proposes a method for detecting icons on a car dashboard's display for the purpose of quick, non-rigorous testing in iterative development using a low-cost consumer camera and minimal setup costs. The HSV Value channel of the images is thresholded using Otsu's method and Regions of Interest are identified using Normalized Cross Correlation. The Regions of Interest are then registered with the templates of the icons and a combination of the normalized root-mean squared error and structured similarity index is used to remove false positives.

The method was found to have almost real-time performance for single icons and acceptable performance when searching for up to 20 icons simultaneously. To simulate real conditions, the display was placed at various angles and distances from the camera and the system was found to be relatively robust to perspective transformations and changes of scale. In the tests, angle and distance were varied over a larger than expected range of conditions and still a zero false positive rate could always be achieved. However, false negatives do occur outside the expected range of variation.