Towards Evaluating the Built-Environment: A Hybridized Approach for Identifying Building Entrance Accessibility
Preprocessing and modeling scripts for Carleton University DATA 500 course project. Project objectives are to develop a binary classifier through deep learning (convolutional neural networks, or CNNs) for predicting whether an image is "accessible" or "inaccessible".
In 2019 the Accessible Canada Act was legislated to ensure a barrier-free Canada by 2040. A priority of this act is to evaluate the built environment to determine if it is accessible and inclusive for all, thus assuring there are no infrastructure barriers that impede those with mobility restrictions. However, currently there are data gaps on the infrastructure that make the built environment inaccessible for the 9.6% of people over 15 years-old who have mobility disabilities. To mitigate these data disparities, this project seeks to support this movement by using open source crowdsourced street-level images from Mapillary and convolutional neural networks (CNNs) to develop and test a binary image classifier that predicts whether a building’s entrance is accessible or inaccessible. Results indicate a model with 81.97% training accuracy and scored 85% accuracy when tested on 20 images. Though model should be improved through refining and expanding the dataset for training, results offer a promising direction towards evaluating the built environment for policymakers, city planners, and more.