Automatically detect crosswalks from satellite view images: A deep learning approach with ground truth verification

Abstract

Like roadway information is to motor vehicle safety, pedestrian facility information (e.g., sidewalk presence) is crucial towards improving the safety of these vulnerable road users. Yet unlike widely accessible roadway data, pedestrian facility data is unavailable for most state agencies. Without this information, data-driven problem identification, countermeasure analysis, project evaluation, and performance management will be heavily impeded. Thus, urgent need for this data was recognized by state departments of transportation (DOTs). To address this need, we developed an automated approach for the automated detection of crosswalks in satellite images. The most advanced deep learning methodology, transfer learning with a convolutional neural network (CNN) was used to handle real-world images. During the prediction process, a satellite image of a roadway pavement was analyzed by the satellite view model to predict the presence of a crosswalk. Then, the street view image of the same target was detected by the integrated street view model as a ground truth check. A total of 18 361 images from Bing maps in satellite view and street view were used to train and test the deep learning model. As a result, the satellite view model itself achieved 98.43% accuracy using testing data from the same region. When dealing with data from another region, using the satellite view detection with ground truth checking increased the accuracy by 49%. It is obvious that by integrating the ground truth checking model into the satellite view crosswalk detection, we can obtain a more robust model which can handle highly occluded, low quality satellite images.