Software Development of Autonomous Vehicle For Real-Time Object Detection

Folyóirat címe
Folyóirat ISSN
Kötet címe (évfolyam száma)

Object detection techniques help us classify objects in images and locate their exact position. This thesis aims to provide an in-depth analysis of the well-known YOLO family of object detection techniques. A detailed discussion is offered on the architecture of each algorithm, the approach of each algorithm to perceive an object in an image, and the techniques and novelties utilized to boost the performance of algorithms to make them feasible for real-time application. An extensive experimental analysis is also conducted on two algorithms from the YOLO family i.e. YOLOv5 and YOLOv8. These research investigations include evaluating the performance of models that have already been trained, developing these models using a unique dataset of traffic signs, and finally identifying traffic signs from real-time photos. The results revealed that the YOLOv8x model's pre-trained weights outperformed when evaluated on the COCO128 dataset with 0.829 mAP@50, but the YOLOv5l model achieved the highest mAP@50 of 0.987 when trained on the custom dataset for traffic signs. In addition, the detection results indicated that the YOLOv5 models outperformed the YOLOv8 models in terms of making accurate predictions for traffic signs in real-time images.

Object Detection, Deep Learning, YOLO Family, YOLOv5, YOLOv8