Detecting the shuttlecock for a badminton robot: A YOLO based approach

The ability to identify objects of interest from digital visual signals is critical for many applications of intelligent systems. For such object detection task, accuracy and computational efficiency are two important aspects, especially for applications with real-time requirement. In this paper, we study shuttlecock detection problem of a badminton robot, which is very challenging since the shuttlecock often moves fast in complex contexts, and must be detected precisely in real time so that the robot can plan and execute its following movements. To this end, we propose two novel variants of Tiny YOLOv2, a well-known deep learning based detector. We first modify the loss function to adaptively improve the detection speed for small objects such as shuttlecock. We then modify the architecture of Tiny YOLOv2 to retain more semantic information of small objects, so as to further improve the performance. Experimental results show that the proposed networks can achieve high detection accuracy with the fastest speed, compared with state-of-the-art deep detectors such as Faster R-CNN, SSD, Tiny YOLOv2, and YOLOv3. Our methods could be potentially applied to other tasks of detecting high-speed small objects.