Performance analysis of YOLO versions for automatic vehicle detection from UAV images

Abstract

Automatic vehicle detection, one of the study areas in Remote Sensing facilities, has become widely used on several issues such as transportation, disaster management, highway management, parking lot management and real-time vehicle detection in smart cities. In recent years, deep learning methods have been widely preferred in vehicle detection. Although this method has advantages such as high accuracy and speed of detection, some problems such as not detecting vehicles, double detection and class confusion in detection from digital images caused by vapor and shadow in adverse weather conditions (i.e., rain, fog, sunlight) have been raised. Thus, vehicle detection is still a significant issue that should be studied. In this study, versions of You Only Look Once (YOLO), one of the deep learning (DL) architectures, have been investigated in terms of performance assessments of vehicle detection in parking lots. To perform the analysis, Unmanned Aerial Vehicle (UAV)-based images collected from Yildiz Technical University, Campus of Davutpasa (dated 2018) were used. The labeling process was performed for three classes (car, bus, and minibus) using the Visual Object Tagging Tool (VoTT). The labeled dataset has been trained via transfer learning in YOLOv4-CSP, YOLOv4-tiny, YOLOv4-P5, YOLOv4-P6, YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x architectures. The weights of YOLO versions have been implemented to the parking lots and results have been compared. To assess the performance of YOLO-based vehicle detection, mAP and F1-Score values were computed.