Infrastructure-based sensor fusion for acquiring gross vehicle weight rating classifications

Abstract

Gross Vehicle Weight Rating (GVWR)-based vehicle activity data are widely used in freight planning, fuel efficiency evaluation, and on-road emission estimation. However, existing data sources rely on either surveys or mapping from other classification schemes. GVWR classification data acquisition directly using existing highway sensor infrastructure remains challenging. To address this challenge, this paper developed an approach to acquire GVWR-based classification data through the aggregation of two complementary infrastructure-based sensing technologies: inductive loop sensors and side-fire cameras. An open-source intelligence (OSINT) method was initially adopted to establish a GVWR-based vehicle dictionary to overcome mapping challenges with classes that cannot be directly associated with singular GVWR-based classes. A dataset comprising 9,154 vehicle inductive loop signatures paired with images was then collected and annotated according to the pre-defined dictionary. Next, signature-based and image-based classification models were developed for GVWR classification, with model designed to function independently. The signature-based GVWR classification model was trained with a multi-layer perceptron (MLP) architecture and optimized through the implementation of a weighted cross-entropy loss function. The image-based GVWR classification framework was designed to extract vehicle objects in a two-stage process and classify them based on the GVWR scheme. Finally, a linear integration model was implemented to combine the output of the signature- and image-based models to achieve an improvement over each standalone classification model. The sensor integration framework significantly outperformed each individual sensing technology, achieving an average correct classification rate of 0.97 and an $F_1$ score of 0.96, which surpasses state-of-the-art methods.