Study on Vehicle Re-Identification Methods for Advancing Automobile Traffic Volume Surveys

In recent years, artificial intelligence (AI) has been increasingly applied to streamline traffic volume surveys. Conventional approaches can estimate traffic counts at a single location but often fail to reliably identify the same vehicle across multiple camera sites, limiting their ability to capture inter-site vehicle flows. To address this challenge, we propose a novel method that integrates license plate recognition, vehicle passage time, and visual appearance features to achieve robust cross-camera vehicle identification. Specifically, to overcome the limitations of standard object detectors in recognizing small-scale license plate characters under low-resolution and motion-blurred surveillance conditions, we develop YOLOv8-LPR by incorporating Cross-Scale Attention (CSA), a Bi-directional Feature Pyramid Network (BiFPN), TokenBridge, Feature-wise Linear Modulation (FiLM), and a Boundary Head into YOLOv8. To construct the training dataset, we leverage the Segment Anything Model (SAM) to automatically extract license plate characters, significantly reducing manual labeling effort and enabling efficient dataset creation. Experimental results demonstrate that YOLOv8-LPR achieves over 95% recognition accuracy and reliably identifies all vehicles across different videos. This advancement enables the estimation of inter-site vehicle flows and directional traffic volumes at intersections—tasks traditionally difficult to accomplish—and is expected to significantly improve traffic survey methodologies.