Deep Unsupervised Transfer Adversarial Network for Abnormal Driving Behavior Recognition Based on Smartphone Sensors

Abnormal driving has been widely recognized as one of the key factors highly related to traffic accidents. Smartphones mounted on vehicles can be leveraged to record a variety of vehicle motion-related data, and therefore, can serve as a platform for monitoring driver abnormal behavior. However, due to the data distribution shift and domain discrepancy, the abnormal driving behavior recognition (ADBR) model trained in one driving scenario probably fails in predicting the behavior data acquired from the other driving scenarios. We propose a groundbreaking unsupervised domain adaptation (UDA) approach that Provides a solution to transfer knowledge acquired from the tagged source domain (SD) to target domains (TDs) that do not have tagged data. Specifically, a dual-stream feature extraction module consisting of 1-D convolution and multihead attention is first established to extract transferable features from raw sensor data. Then, a confidence-based pseudolabeling self-training approach is developed to fully utilize the unlabeled target domain data. Furthermore, a joint adversarial domain adaptation (JADA) method is presented to reduce both marginal and conditional distribution discrepancy simultaneously. By doing this, source and TD data can be aligned well. The proposed method is tested on real-world driving behavior datasets and the results demonstrate the effectiveness and superiority of our model in cross-scenario ADBR.