A comparative analysis of lightweight deep learning models for driver distraction detection

Azizoğlu G., Toprak A. N.

VII. INTERNATIONAL TURKIC WORLD CONGRESS ON SCIENCE AND ENGINEERING, Priştine, Kosova, 13 - 15 Kasım 2025, ss.2-12, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Priştine
  • Basıldığı Ülke: Kosova
  • Sayfa Sayıları: ss.2-12
  • Erciyes Üniversitesi Adresli: Evet

Özet

Driver distraction is one of the leading causes of road accidents worldwide, making its accurate and efficient detection a critical component of intelligent transportation systems. While deep learning approaches have achieved strong performance in this domain, many state-of-the-art models are computationally expensive and unsuitable for deployment on embedded, real-time platforms. To address this challenge, this study presents a comparative analysis of lightweight deep learning models for driver distraction detection, focusing on the trade-offs between accuracy and computational efficiency. We evaluate ten popular lightweight architectures, which are MobileNetV2, MobileNetV3, MobileNetV4, ShuffleNetV2, SqueezeNet, EfficientNet-B0, ResNet18, MixNet, EfficientViT-B0, and GhostNetV2, using the AUC and the StateFarm distracted driver datasets. The models are assessed across key metrics, including accuracy, precision, recall, F1-score, inference speed (FPS), and parameter efficiency, with additional deployment experiments conducted on the NVIDIA Jetson Xavier platform. The results show that MixNet and ResNet18 achieve the highest accuracy on the more challenging AUC dataset, with 95.01% and 94.69%, respectively, while ShuffleNetV2 and MixNet reach near-perfect accuracy (99.96%) on the StateFarm dataset. In terms of efficiency, SqueezeNet delivers the fastest inference speed (116.61 FPS) with the smallest parameter count (0.74M), while MobileNetV4 provides a balanced trade-off between speed and accuracy. These findings highlight that the optimal model depends on deployment priorities: accuracy-focused applications may prefer MixNet, whereas resource-constrained, real-time systems benefit most from SqueezeNet or MobileNet variants. The study provides practical insights for integrating lightweight deep learning models into next-generation driver monitoring systems.