Detection of Road Potholes by Applying Convolutional Neural Network Method Based on Road Vibration Data

Ozoglu F., Gökgöz T.

SENSORS, vol.23, no.22, pp.1-19, 2023 (SCI-Expanded)

  • Publication Type: Article / Article
  • Volume: 23 Issue: 22
  • Publication Date: 2023
  • Doi Number: 10.3390/s23229023
  • Journal Name: SENSORS
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), Biotechnology Research Abstracts, CAB Abstracts, Communication Abstracts, Compendex, INSPEC, MEDLINE, Metadex, Veterinary Science Database, Directory of Open Access Journals, Civil Engineering Abstracts
  • Page Numbers: pp.1-19
  • Yıldız Technical University Affiliated: Yes


In the context of road transportation, detecting road surface irregularities, particularly potholes, is of paramount importance due to their implications for driving comfort, transportation costs, and potential accidents. This study presents the development of a system for pothole detection using vibration sensors and the Global Positioning System (GPS) integrated within smartphones, without the need for additional onboard devices in vehicles incurring extra costs. In the realm of vibration-based road anomaly detection, a novel approach employing convolutional neural networks (CNNs) is introduced, breaking new ground in this field. An iOS-based application was designed for the acquisition and transmission of road vibration data using the built-in three-axis accelerometer and gyroscope of smartphones. Analog road data were transformed into pixel-based visuals, and various CNN models with different layer configurations were developed. The CNN models achieved a commendable accuracy rate of 93.24% and a low loss value of 0.2948 during validation, demonstrating their effectiveness in pothole detection. To evaluate the performance further, a two-stage validation process was conducted. In the first stage, the potholes along predefined routes were classified based on the labeled results generated by the CNN model. In the second stage, observations and detections during the field study were used to identify road potholes along the same routes. Supported by the field study results, the proposed method successfully detected road potholes with an accuracy ranging from 80% to 87%, depending on the specific route.