Humanoid Robot Detection Using Deep Learning: A Speed-Accuracy Tradeoff

Recent advances in computer vision have made the detection of landmarks on the soccer field easier for teams. However, the detection of other robots is also a critical capability that has not garnered much attention in the RoboCup community so far. This problem is well represented in different RoboCup Soccer and Rescue Robot Leagues. In this paper, we compare several two-stage detection systems based on various Convolutional Neural Networks (CNN) and highlight their speed-accuracy trade off. The approach performs edge based image segmentation in order to reduce the search space and then a CNN validates the detection in the second stage. We use images of different humanoid robots to train and test three different CNN architectures. A part of these images was gathered by our team and will be publicly available. Our experiments demonstrate the strong adaptability of deeper CNNs. These models, trained on a limited set of robots, are able to successfully distinguish an unseen kind of humanoid robot from non-robot regions.