Self-Supervised Learning for Driving Maneuver Prediction from Multivariate Temporal Signals

Predicting driving maneuvers is a crucial task for advanced driver assistant system (ADAS). Accurately predicting the driving maneuvers from a driver in real-time is still a challenging mission. Good quality and well-annotated data recorded from drivers are required to train models using machine learning technique. However, data annotation obtaining work is a labor-costly, time consuming and sometimes unrealistic procedure. To deal with this problem, a Self-supervised Knowledge Distillation Network (SKDN) is proposed for driving maneuver prediction from multivariate temporal physiological signals. Unlike previous works that exploit architecture-specific cues for distillation, we try to explore a more distinct way for knowledge distillation from the pre-trained teacher model. By exploiting self-supervised learning as an auxiliary task, SKDN is able to transfer the hidden information from teacher to student effectively. Meanwhile, non-annotated data can be used through contrastive learning to release massive labeling work. To enhance the performance, a custom-built MTS-FReLU is adopted for SKDN to generate temporal contextual dependencies and spatial information from different signals. The model is evaluated on a data set collected from real-world driving scenarios, and the experimental results show that the proposed SKDN outperforms the other advanced knowledge distillation models.