Adaptive Self-Learning Framework for Resilient Vehicle Classification Through the Integration of Inductive Loops and LiDAR Sensors

Inductive loop sensors are widely deployed across the U.S. and can provide vehicle classification data with comparable accuracy to the current axle-based sensor systems when they are enhanced with the inductive signature technology and advanced machine learning models. However, the existing truck population is expected to turnover and be replaced with newer models that may generate distinct inductive signature characteristics. Consequently, legacy inductive signature-based models may not perform optimally in classifying newer trucks operating on the highways over time. To enhance the resilience of the signature-based classification system, this paper investigated a self-learning framework to address the classification system obsolescence through the integration of two complementary sensor technologies: Inductive loop sensors and Light Detection and Ranging (LiDAR) sensors. In this framework, the LiDAR-based Federal Highway Administration (FHWA) classification model served as a data labeling platform to generate class labels for validating and updating the legacy signature-based model. Next, an adaptive transfer learning framework was implemented to improve the performance of a legacy inductive signature-based classification model without compromising computation efficiency. This framework demonstrates the resilience enhancement of the inductive signature-based FHWA classification model with an intelligent system update to accommodate vehicle transition over time while retaining legacy knowledge of the pre-existing population using a methodology that significantly reduces the overall burden of periodic model calibration by utilizing the information stored in the legacy model. The experiment demonstrates that this adaptive self-learning framework achieves an overall correct classification rate of 0.89 on a dataset with distinctively different truck configurations.