Real-Time On-Device Vision-Based Road Anomaly Detection with Frame Span Interval (FSI) Event Counting

This study aims to improve transportation safety by addressing a persistent source of road accidents: pavement anomalies such as potholes and cracks. We develop a comprehensive system that detects potholes and multiple crack types (e.g., alligator, transverse, longitudinal), estimates their size/severity, and is architected to transmit evidence and metadata to a cloud service to support timely action by authorities. Central to our approach is the Frame Span Interval (FSI), which imposes a temporal window (with spatial tolerance) so that consecutive detections of the same physical anomaly at a specific location are consolidated into a single event-preventing duplicate counts, enabling accurate real-time tallying, and stabilizing downstream tracking metrics. The prototype, built with a Raspberry Pi, camera module, and CNN-based model with a laptop host, demonstrates on-device inference and local logging; backend integration and notifications are future work. This integrated approach targets the mitigation of accident risk arising from roadway damage and supports safer road conditions for the broader community.