Deep Urban Vision: Street View Composition, Depth, and Travel Mode Choice
Project Description
To explore the relationship between human-centered, eye-level built environments and travel mode choices in Hong Kong. By utilizing Google Street View (GSV) imagery and advanced computer vision, this project quantifies visual streetscape features (e.g., openness, complexity) to understand their impact on urban mobility and transportation system resilience.
Supervisor
ZHU, Pengyu
Quota
3
Course type
UROP1000
UROP1100
UROP2100
UROP3100
UROP3200
UROP4100
Applicant's Roles
1) Deep Image Recognition & Processing
1.1 Model Optimization: Implement and fine-tune deep learning algorithms (DeepLab V3+, Mask2Former, or YOLOv4) to segment street-view imagery into high-precision semantic categories.
1.2 Depth Estimation: Apply monocular depth scaling (MiDaS) or other methods to predict dense depth maps for streetscapes to evaluate spatial enclosure and openness.
2) Result Verification
2.1 Performance Validation: Identify and verify training results by calculating Mean Square Error (MSE) and comparing predicted values against ground truth datasets.
2.2 Algorithm Refinement: Detect poor partitioning or failure cases in image segmentation and perform iterative testing to switch or refine algorithms for higher accuracy.
3) Machine Learning & Data Analytics
3.1 Feature Engineering: Automate the calculation of eye-level environment indexes (OCEAN) across multiple buffer sizes (100m to 500m) for thousands of residential locations.
3.2 Model Interpretation: Assist in training machine learning models and utilizing Shapley Values (SHAP) to visualize the nonlinear impacts of visual features on various travel modes.
1.1 Model Optimization: Implement and fine-tune deep learning algorithms (DeepLab V3+, Mask2Former, or YOLOv4) to segment street-view imagery into high-precision semantic categories.
1.2 Depth Estimation: Apply monocular depth scaling (MiDaS) or other methods to predict dense depth maps for streetscapes to evaluate spatial enclosure and openness.
2) Result Verification
2.1 Performance Validation: Identify and verify training results by calculating Mean Square Error (MSE) and comparing predicted values against ground truth datasets.
2.2 Algorithm Refinement: Detect poor partitioning or failure cases in image segmentation and perform iterative testing to switch or refine algorithms for higher accuracy.
3) Machine Learning & Data Analytics
3.1 Feature Engineering: Automate the calculation of eye-level environment indexes (OCEAN) across multiple buffer sizes (100m to 500m) for thousands of residential locations.
3.2 Model Interpretation: Assist in training machine learning models and utilizing Shapley Values (SHAP) to visualize the nonlinear impacts of visual features on various travel modes.
Applicant's Learning Objectives
1. Computer Vision Mastery: Gain hands-on experience in state-of-the-art semantic segmentation and depth estimation using real-world urban big data.
2. Big Data Engineering: Learn to handle and process large-scale datasets retrieved via GSV APIs and integrate them with longitudinal survey structures.
3. Interdisciplinary Innovation: Understand how advanced computational techniques can be applied to solve critical urban planning and public policy challenges, such as traffic congestion and sustainable mobility.
2. Big Data Engineering: Learn to handle and process large-scale datasets retrieved via GSV APIs and integrate them with longitudinal survey structures.
3. Interdisciplinary Innovation: Understand how advanced computational techniques can be applied to solve critical urban planning and public policy challenges, such as traffic congestion and sustainable mobility.
Complexity of the project
Moderate