A multispectral vision-based machine learning framework for non-contact vehicle weigh-in-motion

The present study proposes an automated non-contact weigh-in-motion framework using multispectral vision and machine learning method, which provides an all-weather monitoring of heavy vehicles on transportation infrastructures. Multispectral based computer vision techniques consist of two steps, one is acquiring the tire deformation parameters from thermal imaging camera and the latter is obtaining the sidewall characters from the ordinary one. A novel edge detection algorithm is designed for thermal image-based tire deformation estimation. A general recognition model with well-established tire characters database for universal tire sidewall markings, such as non-standard fonts or non-text characters, curved markings or inverted markings, and is developed based on Efficient neural network (E-NET) and YOLOV5S methods. A machine learning-based model for estimating the various vehicles tire-roadway contact forces is established based on eXtreme Gradient Boosting (XGBoost). A database of different vehicles' tires with a total of 1127 samples from 19 types of tires and 648 of laboratory datasets from common vehicles are collected to train the model. The Bayesian optimization is employed to capture the optimal model, and a hybrid training strategy including biased data is applied to update the model. An importance analysis for 11 features is conducted to validate the performance proposed model. The accuracy, feasibility, and effectiveness of the proposed method is carefully investigated by field experiments for trucks and sport utility vehicles under static and low-speed drive. A maximum of 5 % difference is witnessed by comparing the results with that of the traditional WIM system. The present method presents a cost-effective means with no sensors on structures for vehicle weighting, that demonstrates a great potential to implement in toll station, highway infrastructures as well as health monitoring of civil engineering fields.