Using the perceptional map as a foundation, along with additional geo information such as points of interest (POI) and mobility data, I developed a web application that enables tourists or night runners to select various pedestrian routes within the city. Furthermore, I leveraged the vast dataset of over 100,000,000 photos to calculate safety scores for the entire country. These scores were subsequently sold to Daimler Innovative Lab the following year, providing them with valuable insights and information for their initiatives.
The Idea
A personal experience of a terrible pedestrian encounter inspired this project. Motivated to improve pedestrian safety, I developed an algorithm that calculates safety scores by analyzing the visual context of street views. Using this algorithm, I created a pedestrian routing web application that determines routes based on the average safety score along each road. The aim was to provide users with safer and more secure paths for their journeys, enhancing the overall pedestrian experience in urban areas.
Back in 2016, I had just relocated to a new apartment and planned to meet a friend for a meal nearby. Unfortunately, Amap suggested a rather unpleasant walking route. The path was lined with crumbling roads, dimly lit areas, high walls on either side and even many slums concealed behind these walls, emitting an offensive odor.
Modeling the perceptional predictor
This experience sparked my project’s concept: how many unsafe streets still exist within the urban region of Shanghai? To address this question, I trained a safety perception classifier and developed software that could automatically evaluate the safety score of a street. This led to the creation of a comprehensive safety map of Shanghai.
Demo Video
Below is a demo video showing how the score is applied for pedestrian routing.
Revisit the Dataset and Algorithm
Place Pulse 2.0
A significant source of street perception data is the “Place Pulse” project. It collects public feedback on various urban areas using a pairwise labeling interface. Initiated in 2010, the project’s website showcases roughly 4,000 images from four cities: Boston, New York, Linz, and Salzburg. As shown in Figure 1, it gauges public perception across six dimensions. Participants are presented with pairs of street views and asked, “Which image looks safer?”
Over the years, the dataset was enhanced to version 2.0, which includes 110,998 random street-view images gathered from 56 cities in 28 countries across six continents. From 2013 to 2016, a whopping 81,630 online volunteers took part in the survey, providing over 1,170,000 pairwise comparisons. Given the immense diversity and volume of image samples, participants, and corresponding responses, this dataset closely aligns with general human perceptual preferences regarding urban scenes.
Binary SVM for Baseline Measurement
As previously discussed, the simplest and most effective method involves training a binary classification model to obtain a predictive model for safety score. Here, we display the accuracy results of binary SVM (Support Vector Machine) for all six models, across six dimensions.
The CAMs of street views
Below are the negative and positive samples after adopting the updated models.
Collaboration with Diamler
In 2018, our model delivered safety maps for more than 300 cities across China. This data was utilized to create safe pedestrian routes for an app developed specifically for Daimler users.
Initially, our clients were skeptical about our algorithm’s effectiveness. To challenge our model, they sent us a batch of photos that did not consist of street view images. Surprisingly, in a scoring system out of 100, the image deemed safest was a nighttime picture, despite our training dataset lacking any night images. This unexpected result successfully demonstrated our model’s robustness and subsequently sealed the deal.
