Recognizing Distant Vehicles on GMM by Extracting Far Road Area Based on Analyzing Trajectories of Nearby Vehicles

Abstract

In today’s motorized society, road accidents occur frequently, and their incidence continues to rise with the increasing number of car users worldwide. A significant proportion of these accidents occur at intersections, where one promising countermeasure is the use of multi-camera systems that assist pedestrians and drivers by detecting moving vehicles in the intersection area. However, conventional vehicle detection methods suffer from reduced accuracy as vehicles move farther from the camera, since distant vehicles appear smaller in images. To address this limitation, we propose a method that first identifies the distant region of the road in an image and then applies up-sampling to enhance the visibility of faraway road area for improved vehicle detection. In the proposed approach, nearby moving vehicles are roughly extracted using inter-frame subtraction across consecutive frames, and these subtractions are accumulated over time as trajectories. Based on these trajectories, we introduce a novel method to estimate the road’s vanishing point, which is then used to determine the distant road area. This region is subsequently up-sampled in consecutive frames, and vehicle detection is performed using a Gaussian Mixture Model (GMM) to identify distant vehicles. Extensive experiments confirm the effectiveness of the proposed method. The results demonstrated that, although detection accuracy naturally decreases with distance, our method achieves more than twice the accuracy of conventional approaches under both daytime and nighttime conditions.