A Multi-Vehicle Speed Estimation Framework Using Line-Based Homography Transformation in Static Traffic Videos

Abstract

This study focuses on enforcing speed limits and detecting speed violations in an effort to reduce traffic accidents. Vehicle speed detection systems have become crucial in traffic law enforcement. Video-based camera technology is becoming increasingly popular due to its lower initial costs and rich visual capabilities. However, single-camera-based vehicle speed detection systems are limited, especially in accurately measuring vehicle distances. To address this, researchers aim to enhance vehicle speed detection systems using a single camera with homography transformation and intersection point definition algorithms, which can provide significant benefits in traffic law enforcement, vehicle speed monitoring, and overall traffic safety improvement. This study explores the use of the Segformer segmentation model, integrating homography transformation and the Bentley-Ottmann algorithm to obtain four points of the homography matrix. The results of Segformer show that it outperforms all compared segmentation models with an mIoU of 51.8 and 183.3 GFLOPS. For the vehicle detection task, YOLOv9-c outperformed all other variants examined with mAP50 and mAP50-95 of 0.694 and 0.493, respectively. YOLOv9-c also achieved the highest precision and F1-score with 0.740 and 0.675, respectively. In addition, vehicle speed evaluation in terms of MSE, RMSE, and MAE showed that the proposed framework outperformed state-of-the-art methods, achieving 2.01, 1.42, and 0.71, respectively.