TLDR: This research paper presents a novel method using Google Street View (GSV) images and deep learning (YOLOv4) to predict cycling and motorcycling mode shares in 185 global cities. By detecting vehicles in GSV images and combining these counts with population density in a beta regression model, the study achieved accurate predictions (R² ~0.61). The findings demonstrate that GSV imagery, coupled with computer vision, offers a valuable and efficient complementary data source for understanding global travel behaviors, especially in areas lacking traditional survey data, revealing distinct regional patterns in two-wheeled transport.
Understanding how people travel in cities is crucial for public health, road safety, and environmental sustainability. Specifically, data on cycling and motorcycling behaviors can inform policies that promote physical activity, reduce air pollution, and prevent injuries. However, comprehensive and consistent global data on these modes of transport have traditionally been scarce, often relying on time-consuming and inconsistent methods like household travel surveys or censuses.
A new research paper, titled “Vehicle detection from GSV imagery: Predicting travel behaviour for cycling and motorcycling using Computer Vision,” introduces a groundbreaking approach to address this data gap. Authored by Kyriaki [Kelly] Kokka, Rahul Goel, Ali Abbas, Kerry A. Nice, Luca Martial, SM Labib, Rihuan Ke, Carola Bibiane Schönlieb, and James Woodcock, this study leverages Google Street View (GSV) imagery combined with advanced computer vision techniques to estimate cycling and motorcycling levels across a wide range of global cities.
A Novel Approach to Data Collection
The researchers utilized data from 185 cities worldwide. For each city, they collected approximately 8,000 GSV images from sampled locations. To identify cycles and motorcycles within these images, they employed the YOLOv4 deep learning model. This model was specifically fine-tuned using images from six diverse cities (Rome, Bangkok, Kampala, Tokyo, Bogota, and Tel Aviv) to ensure its accuracy and generalizability across different urban environments. The fine-tuned model achieved a high mean average precision of 89% for detecting cycles and motorcycles.
After extracting vehicle counts from the GSV images, the team developed a global prediction model using a statistical method called beta regression. This model used the GSV-detected counts of cycles and motorcycles, along with population density, to predict city-level mode shares (the percentage of trips made by a specific mode). The inclusion of population density was important as it influences overall activity and travel patterns in a city.
Key Findings and Predictions
The study revealed strong correlations between the GSV-detected vehicle counts and the actual mode shares. Specifically, there was a strong correlation of 0.78 between GSV motorcycle counts and motorcycle mode share, and a moderate correlation of 0.51 for cycling. The beta regression models proved effective in predicting mode shares, achieving R² values of 0.614 for cycling and 0.612 for motorcycling, with low median absolute errors (1.3% for motorcycling and 1.4% for cycling).
The model’s predictions were consistent across the full range of data, although some cities like Utrecht and Cali showed larger discrepancies between observed and predicted values. Interestingly, the researchers also investigated whether the difference in dates between mode share data and GSV images impacted the predictions, but found that it did not significantly improve the model’s performance.
A significant application of this model was its use to estimate cycling and motorcycling mode shares in 60 cities for which recent mode share data was unavailable. The predictions highlighted clear regional patterns: motorcycling was generally more prevalent in Asia, while cycling was more common in Europe. For instance, cycling mode shares in demo cities ranged from 1% in Hsinchu to 43% in Copenhagen, while motorcycling ranged from 0.37% in Rotterdam to 65% in Hsinchu.
Also Read:
- Designing Efficient UAV Vertiport Networks with a Smart Recommendation System
- AI Framework Enhances Emergency Response and Resource Equity in African Communities
Implications and Future Directions
This research underscores the immense potential of combining GSV imagery with computer vision for predicting travel behaviors globally. It offers a valuable complementary data source to traditional methods, providing consistent insights through large-scale analysis. The automated data extraction process makes it a highly efficient tool for public health planning and policymaking.
While the study has several strengths, including its global scope and the development of open-source models, it also acknowledges limitations. These include the spatial and temporal constraints of GSV data, such as limited coverage of local roads and inconsistent update frequencies. Future work could involve implementing more advanced computer vision techniques like segmentation models to not only identify vehicles but also determine their area proportion and state (parked or moving), and to link these behaviors with specific environmental characteristics. Additionally, the model could be applied to even more cities as GSV coverage expands globally.
For more detailed information, you can access the full research paper here.
