A differential evolution based axle detector for robust vehicle classification

Video based vehicle classification is gaining huge grounds due to its low cost and satisfactory accuracy. This paper presents a robust vehicle classification system. The system in its essence, aims to classify a vehicle based on the number of circles (axles) in an image using Hough Transform which is a popular parameter based feature detection method. The system consists of four modules whereby the output of one module feeds the next in line. We test our system on single lane highway and street traffic. When the information about the problem at hand (changing weather conditions, camera calibration parameters etc.) is limited or is dynamic, determining the Hough Transform set-up parameters manually becomes time consuming, challenging, and may often lead to false detections. This calls for finding the appropriate parameter-set dynamically according to the situation, which inherently is a global optimization problem. Differential Evolution has emerged as a simple and efficient global optimizer, and we couple it with Hough Transform to improve the overall accuracy of the classification system. We test five different variants of DE on varied videos, and provide a performance profile of all the variants. Our results demonstrate that employing DE indeed improves the system´s classification accuracy (at the expense of extra compute cycles) making the system more reliable and robust.