Accurate Head Pose Estimation-Based SO(3) and Orientation Tokens for Driver Distraction Detection

Driver distraction is an important cause of traffic accidents. By identifying and analyzing the driver’s head posture through monitor images, the driver’s mental state can be effectively judged, and early warnings or reminders can be given to reduce traffic accidents. We propose a novel dual-branch network named TokenFOE that combines Convolutional Neural Networks (CNN) and Transformer. The CNN branch uses an Multilayer Perceptron (MLP) to infer the image features from the backbone, then generating a rotation matrix based on SO(3) to represent head posture. The Dimension Adaptive Transformer branch uses learnable tokens to represent the head orientation of 9 categories. Integrate the losses of both branches for training, ultimately obtaining accurate head pose estimation results. The training dataset uses 300W-LP, and the quantatitive testing datasets are AFLW-2000 and BIWI. The experiment results show that the Mean Absolute Error is improved by 21.2% and 9.4% compared to the original SOTA model on the two datasets, and the Mean Absolute Error of Vectors is improved by 19.2% and 10.2%, respectively. Based on the model output and calibrated through the camera adapter module, we present the qualitative results on the largest driver distraction detection dataset currently available, the 100-driver dataset, robust and accurate detection results were achieved for four different camera perspectives in two modalities, RGB and Near Infrared. Additionally, the ablation study shows that the model inference speed (21 to 75fps) can be used for real-time detection.

@article{Zhao2024,
  title     = {Accurate Head Pose Estimation-Based SO(3) and Orientation Tokens for Driver Distraction Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {10},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Xiong Zhao and Sarina Sulaiman and Wong Yee Leng},
  doi       = {10.14569/IJACSA.2024.01510105},
  url       = {https://doi.org/10.14569/IJACSA.2024.01510105}
}