Machine Learning for Real-Time Vehicle Detection in All-Electronic Tolling System

Abstract

All-Electronic Tolling Systems, a global market worth approximately $7B, have made significant contributions in toll collection, commuter convenience, traffic management and highway administration. The current infrastructure however, is multi-tiered and expensive to set up. Alternative ways of vehicle detection can help in significantly lowering costs on new toll infrastructure placement. In this paper, we apply a new perspective to the detection problem by evaluating the applicability of machine learning for detecting vehicle movement through toll gantries in real-time from a novel perpendicular overhead camera angle. We solve this multi-objective problem by incorporating object detection using You Only Look Once (YOLO), more specifically, YOLOv3 and a faster version with less memory requirements, Tiny YOLOv3 to detect vehicles passing through tolls from perpendicular overhead angles in real time and with high accuracy. Additionally, a classification is made between passenger vehicles and trucks/buses of the detected vehicles. Our Experimental results from training YOLOv3 on our data set indicate a recall of 100.0% and a precision of 98.0%. The results of training Tiny YOLOv3 on our data set show a recall of 100.0% and a precision of 98.5%. These results indicate that use of machine learning is not only effective for detecting vehicles in electronic tolling systems in real-time, but that it can be used on cameras positioned at a perpendicular angle despite insufficient annotations.