Towards Automatic Assessment of Perceived Walkability

We present a method for automatic assessment of perceived walkability by pedestrans, using a machine learning technique with deep convolutional neural networks (CNNs) trained on a dataset of georeferenced street-level images obtained from Google Street View. On a dataset of more than 17,000 human-assessed images used for training, validation and testing of CNN, out method yields an accuracy of 78% of correct and 99% of correct or 1-class-off predictions. These are quite promising, even encouraging results, paving the way for seamless large-scale applications of perceived walkability assessment on large metropolitan areas, and for a mass assessment and comparisons of walkability over many cities across regions.