Real-Time Pothole Detection Based on Vision-Dominant Sensor Fusion and Dual ESP32 Architecture

This preprint presents the Intelligent Pothole Detection System (IPDS), a low-cost vehicle-mounted IoT platform for real-time pothole detection and severity assessment. The proposed system combines a custom-trained YOLOv8m computer vision model, dual ESP32 microcontrollers, MPU6050 inertial sensing, GPS geotagging, and SORT-based object tracking. A vision-dominant sensor fusion architecture is introduced, where computer vision provides pothole detection hypotheses and inertial sensing provides physical validation. The system employs geometric filtering and multi-stage false-positive suppression to improve detection reliability. Experimental evaluation achieved 81.68% mAP@0.5, 82.33% precision, and 74.42% recall while maintaining real-time performance on a CPU-only host. The platform generates geo-tagged, timestamped, and severity-scored pothole reports suitable for smart-city infrastructure monitoring, road-condition mapping, and maintenance planning. The complete hardware implementation is designed as a low-cost deployable solution using commercially available components.