Synthetic Data Meets Real-World Challenges: Transfer Learning for Driver Pose Estimation Using RAMSIS

Abstract

Driver posture recognition holds significant application value in intelligent driving and human-machine interaction. However, the difficulty in acquiring driving posture data and the high labeling costs result in large errors in posture recognition models, which severely limit the development of related algorithm research. This paper presents a driving posture transfer learning method based on the RAMSIS synthetic dataset, which can effectively reduce algorithm development costs and significantly improve model performance. First, we constructed a synthetic dataset based on RAMSIS, which includes images of 21 typical driving postures from 54 human body models and corresponding 3D keypoint ground truth labels. The data generation method efficiently acquires sample data, saving substantial labeling efforts, while also incorporating constraints from the automotive ergonomics environment to enhance the realism of the synthetic data—a crucial aspect in dataset construction. Next, we applied a pre-trained model to generate keypoint pseudo-labels for real vehicle driving images, which were then combined with the synthetic data to form a mixed dataset. Finally, we employed a fine-tuning strategy to perform transfer learning on the pre-trained model using the mixed dataset, and explored the impact of synthetic data proportion on model performance. Experimental results show that the proposed method achieves an optimal MPJPE of 30.4 mm, significantly outperforming other models. This provides new data support for driving posture estimation, with excellent model performance in posture recognition, thus offering potential for quantitative posture evaluation. It lays the foundation for broader applications in vehicle human-machine interaction and demonstrates promising engineering prospects.