Implementation of deep learning network to detect distracted driving behaviors in night driving

Driver distraction during night driving is a leading cause of road accidents, creating the need for real-time intelligent monitoring systems. Currently, most existing distracted driving detection systems rely on visible (VIS) cameras and perform well under daylight conditions; however, their performance significantly degrades at night due to poor illumination, noise, and low contrast. Some studies have explored infrared (IR)-based systems to overcome low-light limitations, but limited comparative analysis between IR and VIS imaging under night driving conditions has been reported. In addition, many existing systems are developed and tested on high-performance computing platforms without considering deployment constraints on embedded systems. Existing models often fail to perform accurately in low-light conditions, offer limited comparison between infrared (IR) and visible (VIS) imaging, and face challenges when deployed on embedded systems due to restricted computational power. This research aims to implement convolutional neural network (CNN) models for detecting distracted driving behaviors at night using both IR and VIS image sequences. A deep learning framework was developed and trained using four CNN architectures InceptionV3, MobileNetV2, ResNet-50, and GoogLeNet on a custom dataset consisting of six distracted behaviors collected from 44 participants. The six distracted behaviors include normal driving, yawning, nodding, texting, calling, and talking. Data augmentation was applied to improve generalization. The models were evaluated across four types of images: DAY-VIS, NIGHT-VIS, NIGHT-VIS with CLAHE preprocessing, and NIGHT-IR. Accuracy and inference time were used as the main performance metrics. The final stage involved deploying the trained models on the NVIDIA Jetson Nano to assess real-time performance in an embedded system environment. InceptionV3 achieved the highest classification accuracy on the NIGHT-VIS dataset with 85.47 % accuracy and the NIGHT-IR dataset with 89.27 %, while ResNet-50 recorded the best accuracy of 90.29 % on the NIGHT-VIS-CLAHE dataset. However, ResNet-50 demonstrated more consistent performance across different datasets and maintained real-time capability above 10 fps on the NVIDIA Jetson Nano, making it the most suitable model for deployment in night-time distracted driving detection. Thus, CLAHE implementation achieved highest accuracy for distracted driving behavior in night driving. The findings indicate that ResNet-50 combined with CLAHE preprocessing provides the most reliable performance, achieving the highest accuracy in NIGHT-VIS-CLAHE conditions, and is therefore recommended as the proposed method for distracted driving detection at night. Although InceptionV3 recorded the highest accuracy in certain sessions, ResNet-50 demonstrated more consistent high performance across different datasets. MobileNetV2, while being the fastest model with stable inference times on the NVIDIA Jetson Nano, suffered from relatively low accuracy and thus cannot be suggested as a practical solution. On the other hand, ResNet-50 achieved more than 10 fps, thus ResNet-50 can be deployed on NVIDIA Jetson Nano with acceptable performance. The proposed solution has the potential to reduce night-time road accidents and improve safety for drivers and other road users through reliable and timely behavior detection.