A Novel FHWA-Compliant Dataset for Granular Vehicle Detection and Classification

Abstract

This study addresses the critical challenge of limited annotated datasets for vehicle classification compliant with Federal Highway Administration (FHWA) guidelines. To overcome this challenge, we introduce a novel benchmark annotated dataset meticulously curated from diverse sources, capturing variations in time, resolution, camera position, and weather conditions, with images all aligned with FHWA standards. Building on this dataset, we conduct a comprehensive comparative analysis of several state-of-the-art object detection models including the YOLO series and Detection Transformers. Our analysis reveals the need for an enhanced YOLO model performance in detecting vehicle subcategories within the FHWA classification scheme. To achieve this, we propose improvements to YOLO models by incorporating the Convolutional Block Attention Module (CBAM). These enhancements lead to superior performance, with YOLOv9+CBAM achieving high precision (0.987), recall (0.973), mAP@50 (0.984), and mAP@50-95 (0.892). However, our comparative analysis yielded the Collaborative DEtection TRansformer (Co-DETR) as the best performing model with a precision of 0.993, the recall being 0.985, mAP@50 of 0.988 and an mAP@50-95 of 0.901. Our results underscore the effectiveness of integrating CBAM into YOLO architectures, significantly boosting feature representation and detection accuracy across multiple YOLO versions and the potential for using Detection Transformers in addressing vehicle classification problems.

Keywords

• Improved YOLOv5
• vehicle classification
• FHWA
• attention modules