Master’s degree in applied geomatics, geographic environments program
Lynda Bellalite (co-advisor), Yacine Bouroubi (advisor), Samuel Foucher (external)
Field of research
Characterization of pedestrian corridors: automatic extraction of geographical features from urban panoramic images by convolutional neural networks
The condition of pedestrian corridors, such as sidewalks, has a significant impact on the mobility of people with reduced mobility, their participation in social life and their physical health. Automated monitoring of these corridors, similar to the one currently being developed for the road network in several major cities, would reduce the costs associated with manual surveys for regular maintenance purposes. This control project aims to improve the factors taken into account by a trip planning tool so that it better meets the mobility needs of this type of user.
Since the advent of convolutional neural networks (CNNs) in visual recognition, the applications related to these new machine learning tools have continued to multiply. CNNs are of major interest in the application of computer vision techniques in remote sensing. The main objective of the work presented is to develop a method for automatically characterizing sidewalks and other pedestrian corridors. To do this, a dataset was created from images of Google Street View urban scenes covering 25 kilometres of Sherbrooke’s road network. Then, a semantic segmentation per pixel and classification are applied to detect and characterize the sidewalks, depending on the condition and type of pavement. A transfer learning approach is used to take advantage of large annotated datasets, such as Cityscapes.