Deteksi Penggunaan Ponsel pada Pengemudi Kendaraan Roda Empat Menggunakan Metode YOLOv8 secara Real-Time

Abstract

The behavior of using mobile phones while driving is one of the primary sources of distraction among four-wheeled vehicle drivers contributing to an increased risk of traffic accidents. According to the National Highway Traffic Safety Administration (NHTSA) in 2023, the United States as one of the countries with the largest number of four-wheeled vehicle users recorded 3,275 fatalities caused by distracted driving with mobile phone usage identified as a major contributing factor. To minimize such risks, this study developed a real-time detection system for identifying mobile phone usage behavior in drivers by implementing the You Only Look Once version 8 (YOLOv8) object detection method. The system was implemented in an Android-based application that utilizes the front camera to monitor the driver’s activity during driving. Data collection was conducted by creating a custom dataset consisting of driver images in both phone-usage and non-usage conditions. The visible presence of a phone being held directly by the driver’s hand was used as a key parameter in detecting phone usage behavior. Based on the evaluation results, the system was able to detect phone usage behavior with an accuracy of 83% and a recall of 0.84. However, the precision score of 0.68 indicated a significant rate of false alarms, mostly caused by the similarity between hand poses during non-usage and phone-holding conditions. This suggests a bias in the model and its failure to properly recognize the phone object itself. Additionally, the system was capable of operating in real-time with frame rate 5–8 FPS and an inference time of 65 ms on a POCO X7 Pro device.