Robust Real-time Junction Detection Under Various Conditions Using Dark Channel Maps

The study in this paper aims to demonstrate the performance of a junction detection architecture based on a deep learning paradigm, included with dark channel transformation. We must take into account the hostile conditions when detecting such features. Many of previous papers proposed models for junction detection with hand-crafted logic, which works well under normal conditions but not under hostile conditions. This necessitates a data-driven approach for junction detection. We attempt to do so using two recently proposed deep neural networks: ResNet50 and EfficientNet-B0. Here, given a set of input images of roads with junctions or no junctions, dark channel transformation is applied to better inform the networks about the road regions prominent in the images. According to our experiments on the Oxford RobotCar Dataset, using the dark channel transformation on the ResNet50 trained from scratch can achieve a junction classification accuracy of over 94%. This numerical figure is 8% greater than when the ResNet50 pre-trained on ImageNet is directly trained on RGB Oxford dataset images. When using pre-trained weights of the deep networks, junction classification accuracy rises to 95%, and the precision increases to 99%.