We present a new world-coordinate tracking algorithm for road users seen from static surveillance cameras, denoted GUTS. It is based upon the previously published UTS method but simplifies and replaces parts allowing association logic to work in world coordinates, by using a novel convolutional neural network denoted SAMHNet to convert every detection into world coordinates. Experimental evaluation on synthetic data shows a MOTA increase of 41 % or 153% depending on distance metric, compared to UTS. Furthermore, the system is verified to work on a real-world recording. We further introduce a synthetic dataset denoted UTOCS which is the first of its kind to be standardized and made publicly available, allowing fair comparison between methods.