Research on Truck Driver Behavior Recognition Based on Transfer Learning and Deep Learning

Truck driver behavior recognition is critical for road safety. This study proposes a recognition framework integrating dual-domain adaptive transfer learning and lightweight architecture optimization, constructing a specialized dataset covering normal, distracted, and fatigued driving behaviors. Enhanced ResNet34 and MobileNetV2 models achieved 96.65% F1-score with 18.7ms real-time inference speed, showing 4% accuracy improvement over baseline methods. Notably, through curriculum learning strategies, the model attained 93.6% accuracy using only 25% training data (5,606 samples), outperforming conventional data augmentation by 12.8%. This approach provides an efficient solution for driving safety early-warning systems.