This article presents a novel approach to incorporate visual cues from\nvideo-data from a wide-angle stereo camera system mounted at an urban\nintersection into the forecast of cyclist trajectories. We extract features\nfrom image and optical flow (OF) sequences using 3D convolutional neural\nnetworks (3D-ConvNet) and combine them with features extracted from the\ncyclist's past trajectory to forecast future cyclist positions. By the use of\nadditional information, we are able to improve positional accuracy by about 7.5\n% for our test dataset and by up to 22 % for specific motion types compared to\na method solely based on past trajectories. Furthermore, we compare the use of\nimage sequences to the use of OF sequences as additional information, showing\nthat OF alone leads to significant improvements in positional accuracy. By\ntraining and testing our methods using a real-world dataset recorded at a\nheavily frequented public intersection and evaluating the methods' runtimes, we\ndemonstrate the applicability in real traffic scenarios. Our code and parts of\nour dataset are made publicly available.\n