Road Segmentation from Satellite Images Using FCNN for Autonomous Driving Vehicles

Road detection from images is a complicated task, most of which is done manually. This method is not very efficient and not used in real-time scenarios. So the need for a better and more robust road detection method is necessary. For land transport, navigation is crucial in the present world as there could be more number and complex routes between two locations. With the applications, from annotating new roads to estimating the fastest possible routes, road detection can play a significant role. To overcome the difficulty in manually annotating roads, a Fully convolutional neural network (FCNN) based system is proposed which extracts and learns road features from satellite images. Two different models, FCNN-32 and FCNN-8 networks, were trained with the same dataset. The conventional skip connection in FCNN-8 network was re-arranged and optimised. This customisation helped the network preserve spatial information that could improve overall accuracy of the model. The proposed method has the advantage of processing speed, and it is cost-effective.