1. For each bus stop, street view images were collected around the intersections of roads within the 10-minute-walk distance (800m buffer area).

2. All street views were randomly paired and each pair was compared by an artificial neural netwoork trained on PlacePulsee2.0 dataset to determine which view is safer in terms of visual safety perception.

3. Based on the pairwise comparison of street views, ELO rating system, a ranking system to measure the skills of players competing at a sport, was employed to quantify the ranking score of safety perceptions.

4. To study the distribution of perceived safety in the proximity of bus stops, the mean, median, and stadard deviation of safety ranking scores of street views for each bus stop were calcualted.

To explore the existence of social inequalities in perceived safety, tract-level demographic data (such as average median income and poverty population%) was aggreagated with the 10-minute-walk distance of each bus stop. Regression model - spatial lag (Y = βX + α + ρWY) - was used to examine the relationship between perceived safety with each demographic variable with the nearest neighbor distance matrix. K-means was employed to cluster bus stops based on perceived safety and all demographic variables.