Wavelet–k Nearest Neighbor Vehicle Classification Approach with Inductive Loop Signatures

Abstract

In this study, a new vehicle classification algorithm was developed with inductive loop signature technology. There were two steps to the proposed algorithm. The first step was to use the Haar wavelet to transform and reconstruct inductive vehicle signatures, and the second step was to group vehicles into FHWA vehicle types through the use of the k nearest neighbor (KNN) approach with a Euclidean distance classifier. To determine the proper proportion of the wavelet to apply for reconstruction and feature extraction, transformed signatures were examined with percentages of large components of their corresponding wavelets. To implement the KNN approach, a library of vehicle signature templates for each FHWA vehicle class was composed. The proposed vehicle classification algorithm demonstrated promising classification results, with a 92.4% overall accuracy. The algorithm can be applied to the real world without the concerns about recalibration and transferability that arise with the use of signature data from single loops. Two additional vehicle classification schemes were applied for performance evaluation. For the inductive signature performance evaluation classification scheme, which aimed to facilitate emission analysis and easy interpretation, the overall accuracy was 94.1%. For the axle-based vehicle classification scheme proposed in this project, which aimed to group vehicles by use and the number of axles, the overall accuracy was 93.8%. Future research will focus on refinement of the signature template library for each FHWA vehicle type to further improve the performance of the proposed vehicle classification algorithm. The selection of the value of k for the KNN approach will be investigated also.