3Pod: Federated Learning-based 3 Dimensional Pothole Detection for Smart Transportation

Abstract

The roadway is the backbone of the country. Bad road conditions can cause vehicles damages and create hazardous driving conditions. Auto damages caused by potholes can add up to thousands of dollars per vehicle. The problem is supposed to be more significant to the auto-driving cars. Therefore, evaluating and maintaining roads of road defects plays an essential role in the economy. Most cities rely on residents to report those conditions. However, aside from being a burden on residents, it may not be an effective method. Automating road defect detection using computer vision techniques would significantly ensure roads remain safe and efficient. Smart cities leverage physical and virtual technologies that rely on sensors and cloud-based communication to improve urban environments. This paper proposes a federated deep learning-based 3 Dimensional (3D) pothole detection (3Pod), which is an intelligent real-time evaluation and reporting platform of road conditions and MRI (Maintenance Responsiveness Indicator) using IoT and Artificial Intelligence technologies. It detects road defects in 3D with size estimation to discern other road objects, including patched potholes, fake road bumps, etc. Furthermore, it provides an avoidability score for each defect to show its risk to commuters on the road and their vehicles. We also propose a crowd-voting technique to calculate MRI (Maintenance Responsiveness Indicator), which helps evaluate maintenance performance.