Machine and Deep Learning Models for Vehicle Type Classification and Weight Prediction

Abstract

Roads and bridges form the backbone of a region's transportation network, playing a vital role in facilitating the movement of both people and goods. To ensure their long-lasting performance, these essential structures are carefully engineered and built with precise consideration of their load-bearing capacities. In practical terms, the acquisition of precise traffic volume and weight data has greatly benefited from the deployment of advanced technologies such as traffic counting devices and weigh-in-motion (WIM) systems. These innovations incorporate specialized sensors and equipment integrated into road surfaces or strategically placed at designated points.

In this research, we propose using Autogluon, a framework for automated machine learning (AutoML), to predict vehicle type classifications and ultimately estimate the gross vehicle weight (GVW) of vehicles. Specifically, we focus on vehicle classes ranging from class 4 to class 13, utilizing traffic counts and weigh-in-motion (WIM) data collected from transportation departments in New York, California, and Texas. This data serves as a valuable resource for training and evaluating machine learning and deep learning models within Autogluon.

Due to the high costs associated with installing traffic count sensors and weigh-in-motion (WIM) stations on every road, our approach leverages historical traffic count and WIM data from one location to accurately predict vehicle type classes and estimate vehicle weights on different roads in another location. We first investigate the use of New York's traffic count and WIM data to predict vehicle weights within the state and evaluate the results using its actual data. Subsequently, we transfer the model to predict the vehicle classes and estimate vehicle weights in California and Texas, using New York's WIM data as the reference, and evaluate the model's performance with California and Texas data, respectively.

Preliminary results suggest that, on average, the models successfully predicted the vehicle type classes and estimated the gross vehicle weight of the vehicles, demonstrating a dependable computational efficiency of Autogluon.