Measuring visual walkability perception using panoramic street view images, virtual reality, and deep learning

Abstract

Measuring perceptions of visual walkability in urban streets and exploring the associations between the visual features of the street built environment that make walking attractive to humans are both theoretically and practically important. Previous studies have used either environmental audits and subjective evaluations that have limitations in terms of cost, time, and measurement scale, or computer-aided audits based on natural street view images (SVIs) but with gaps in real perception. In this study, a virtual reality panoramic image-based deep learning framework is proposed for measuring visual walkability perception (VWP) and then quantifying and visualizing the contributing visual features. A VWP classification deep multitask learning (VWPCL) model was first developed and trained on human ratings of panoramic SVIs in virtual reality to predict VWP in six categories. Second, a regression model was used to determine the degree of correlation of various objects with one of the six VWP categories based on semantic segmentation. Furthermore, an interpretable deep learning model was used to assist in identifying and visualizing elements that contribute to VWP. The experiment validated the accuracy of the VWPCL model for predicting VWP. The results represent a further step in understanding the interplay of VWP and street-level semantics and features.