Bise-ResNet: Combine Segmentation and Classification Networks for Road Following on Unmanned Aerial Vehicle

Unmanned Aerial Vehicles (UAVs) have a wide range of applications and long-term development prospects. Automatically following the road areas in aerial images plays an essential role in many applications, such as disaster monitoring, traffic surveillance and building panoramic maps. In this paper, we formulate the road following of UAV as a classification problem. We propose a new deep neural network, BiSe-ResNet, for introducing pixel-level features to a classification network by combining a segmentation network, and then it can decide the moving direction of UAV by directly operating on the image. We collect a new dataset of campus roads from UAV perspective and experiments are performed both on this new dataset and a public dataset. Comparison with existing technique on these two different datasets shows the generalization ability and effectiveness of our method.