A comparative study of different yolo models for vehicle detection

Abstract

Development of automatic vehicle detection (AVD) system using either still images or videos would be quite beneficial to create an automated traffic management system. Based on AVD, there are numerous research articles that have been published in the literature. This thesis focuses on three major object detection algorithms under YOLO (You Only Look Once) family, namely YOLOv5, YOLOv7 and YOLOv8 for AVD. This thesis also discusses the architectural differences found in these variants of YOLO models. For experimental evaluation, we have used three AVD datasets developed for Indian subcontinent, namely JUVDsi v1 and IRUVD and one international dataset namely Udacity Self Driving Cars. We have achieved a satisfactory outcome on the IRUVD dataset with a mAP score of 0.96 using the YOLOv7 model, mAP score of 0.817 on JUVDsi v1 dataset using the YOLOv8 model and a mAP score of 0.753 on Udacity Self Driving Cars dataset. All codebase and detailed results can be found at: https://github.com/JUVDsi/YOLO-comparison.git.