Pathfinder’s Tech X DSTA challenge - elevator control vision module : elevator button detection & activation

Abstract

This report presents a computer vision system use to identify and press targeted button(s) on an elevator control panel based upon contour detection. In additional, there is no prior knowledge about the size, shape, colour, texture and images of the targeted button(s). But strong specular reflections present on the polished surfaces of the button(s) hamper reliability and performance of the contour detection. To overcome this, techniques such as Hough transform, morphological Closing and utilising information from preceding frames were experimented upon. Geometric relationship between detected buttons are then used to infer their respectively storey levels which they represent. A novel technique called Geometric Relationship Button Clustering level for clustering grid-like pattern of the buttons was implemented and discussed. Test shows that the system can successfully detect up to 82.66% of the actual buttons with 1.52% false positives. At an execution time of 200.413 ms per frame. More notably, the system is able to tolerate control panels that are not perfectly upright and/or some degree of perspective distortion.