Vehicle Occupancy Detection for HOV/HOT Lanes Enforcement

Abstract

High-Occupancy Vehicle (HOV) and High-Occupancy Toll (HOT) lanes have gained interest in recent years since they provide innovative solutions to roadway congestion and traffic safety in urban areas. Enforcement is one of the key components of HOV/HOT lane operations in order to maintain the efficiency and operational integrity of the facilities. The incorporation of video imagery to monitor HOV/HOT lanes has driven the development of automatic occupancy detection systems oriented to enforcement. However, seeing through the vehicle glass is a challenging task, independently of the type of sensor in use. In this paper we present a front-seat vehicle occupancy detection as part of a low-computational computer vision HOV/HOT lane enforcement system. Two main stages composes the proposed solution: (i) A novel vehicle front windshield image segmentation, using a Local Appearance Model (LAM) approach supported on multi-dimensional local features (HOGs); (ii) A machine learning image classifier to detect front-seat occupancy using local image representations of the occupant region-of-interest. Experimental evaluation was conducted on a dataset of frontal vehicle images obtained in a real highway scenario. A comparative evaluation against state-of-art solutions was performed and the results reported confirm the good performance of the proposed vehicle occupancy detection system and the robustness of the novel vehicle windshield image segmentation technique.