A High-Speed Object Tracker from Off-the-Shelf Components

1 Background For many interesting robotics applications it is necessary to process the visual input of one or more cameras is real-time, in order to allow interaction between the robot and its environment, e.g. for visual servoing, basketball dribbling, or catching balls. Previous solutions for tracking object in cameras images either require markers on the object (e.g . Vicon setups), or they are not capable of running at high framerate of 100 Hz or more (e.g . Tyzx, OpenCV ). Fast markerless systems have been developed, but they are too expensive for most robotics research projects, because they rely on specialized hardware, such as FPGAs [1, 2, 3] or even custom-built chips [4]. To overcome these limitation, we introduce RTblob, a vision system for tracking simple colored objects that is at the same time very fast, runs on inexpensive hardware, and is available free of charge in open source form. RTblob uses a modular feed-forward architecture that makes it simple to use also for non-experts, and that allows easy adaption e.g ., to different camera setups and tracking tasks.