Ton Peijnenburg, Fellow HTSC and Deputy Director VDL-ETG

AI, deep learning and algorithms such as Convolutional Neural Networks are outperforming other techniques in object detection in images or image streams. When we look at the detection of soccer balls in our soccer robots2, traditional machine vision techniques (including color segmentation and contour detection) are expensive in terms of computing power and not very robust to changes like different illumination. The Robocup initiative is struggling to move its soccer games outdoors, one of the reasons being poor sensing performance in outdoor daylight conditions. Driver assist systems for cars deal with outdoor conditions much better. When we use an off-the-shelf, less traditional neural network like YOLO3 in our lab, we can reliably and robustly detect all the balls in our field independently of their color, paint pattern, distance and illumination.