Unsupervised Cross-Scenario Abnormal Driving Behavior Recognition Using Smartphone Sensor Data

Accurately recognizing abnormal behavior of drivers (e.g., aggressive driving and fatigued driving) based on multivariate sensor data is vital for human-centric assistive driving systems. Existing data-driven deep learning models for abnormal driving behavior recognition (ADBR) achieve promising performance under specific driving scenes with sufficient labeled data. However, in the real world, dynamic driving scenes and unlabeled data pose a great challenge to the adaptability of models. In light of this, we put forward a novel unsupervised cross-scenario ADBR approach that can transfer domain knowledge in the source scenario with labeled data to the target scenario with only unlabeled data, thus considerably enhancing the adaptability of our model. Specifically, we first propose a feature extraction module that can obtain domain-shared and domain-specific features from raw sensor data derived from different driving scenes. Then, adversarial learning is presented to align the feature distribution of source and target domains to reduce the domain shift. A self-training strategy is further developed to boost the target domain classification performance by iteratively using the pseudo labels. Moreover, prediction uncertainty and ensemble classification are proposed to enhance the quality of pseudo labels. Extensive experiments on cross-scenario ADBR are conducted to evaluate the effectiveness of our model. The results manifest that our model significantly improves the recognition performance for the target domain and outperforms the competing algorithms.