Embedded AI System for Road Nail Detection and Transportation Safety
Mutia Delina (a)*, Haris Suhendar (a), Van-Huy Pham (b,c), and Immanuella Senja Dwi Febriani (a)

(a) Physics Department, Faculty of Mathematics and Sciences, Universitas Negeri Jakarta, Jl Rawamangun Muka, Jakarta Timur 13220, Indonesia

(b) Faculty of Information Technology, Ton Duc Thang University, Ho Chi Minh City, 700000, Vietnam

(c) Faculty of Technology, University of Management and Technology in Ho Chi Minh City, Ho Chi Minh City, Vietnam

*mutia_delina[at]unj.ac.id

Abstract

Road safety and infrastructure maintenance are critical aspects of modern transportation systems to support mobility and protect road users. One major challenge is presence of nails on roads, which can cause tire damage, traffic disruption, and accidents. This study proposes a road nail detection system using digital image processing based on the YOLOv4-tiny algorithm. The model demonstrated promising detection performance, with the loss value decreasing to 0.2876 and the mean Average Precision (mAP) reaching 70% at the 5400th iteration. Although a decline in mAP after this iteration indicated potential overfitting, the model was generally capable of recognizing nail objects effectively within the training dataset. Performance evaluation showed an Average Precision (AP) of 90.87% for the ^nail^ class, with 394 true positives and 32 false positives, indicating strong detection capability. Additional metrics, including 85% precision, 82% recall, 83% F1-score, and an average Intersection over Union (IoU) of 67.17%, indicate that the system performs reasonably well. The proposed system has potential applications in preventing tire punctures and improving road safety. Furthermore, this research potentially supports highway patrol officers and road maintenance authorities in monitoring road conditions more efficiently by enabling early detection and rapid removal of hazardous objects such as nails.

Keywords: Nail, YOLOv4-Tiny, safety road, AI

Topic: Instrumentation and Computational Physics