Driver Maneuver Identification with Multi-Representation Learning and Meta Model Update Designs

Driver maneuver identification (DMI), i.e., the task of predicting the driver maneuver classes given sensor measurements (such as IMU sensors from mobile devices), can serve as a key enabler for many ubiquitous driver behavior analysis applications. Despite the prior studies, two major technical challenges remain before an effective DMI system can be deployed: (i) latent complex maneuver behavior feature relations for accuracy enhancement, and (ii) inconsistency and uncertainty in model adaptation given dynamic data collection settings (e.g., given different urban environments, drivers, and mobile devices). To address the aforementioned challenges, we propose MetaDMI, a novel and adaptive driver maneuver identification framework based on multi-representation learning and meta model update, with the case study on the inertial measurement unit (IMU) sensor measurements (i.e., accelerometer and gyroscope). Specifically, we first extract multiple feature representations for each driver maneuver record in the forms of graph, spectral, and time sequence. Next, MetaDMI processes them with a novel multi-representation learning network, extracting complex patterns and feature relations from the driver maneuvers. Finally, to further enhance the adaptivity of our DMI model to external impacts with dynamic data collection, we have designed a regularized meta learning-based training method to regularize the knowledge transfers across the source and target datasets (e.g., across cities/devices, and few-shot initialization) for consistent and robust identification performance. We have conducted extensive experimental studies upon our MetaDMI prototype based on two datasets (one is collected on our own) and shown that our approach outperforms other baseline approaches for DMI in terms of accuracy and adaptivity.