ANALYSIS OF ACCURACY AND SPEED OF VEHICLE DETECTION USING NEURAL NETWORKS YOLOV8 AND YOLOV11

Abstract

The subject of the study is the application of convolutional neural networks (CNN) YOLOv8 and YOLOv11 for the task of detecting car images. For comparative analysis, medium-sized YOLO models were used, namely the YOLOv8m and YOLOv11m models. A specialized dataset was formed, which contains images of vehicles for different scales, acquisition conditions, lighting levels and quality. The dataset was annotated manually using the Roboflow tool. The size of the initial dataset was increased by data augmentation methods. The resulting dataset was divided into training, control and test datasets. The software for training the neural network and object detection was developed in Python using cloud technologies. To increase the speed of model training, parallel calculations were used, which are implemented on the GPU. The YOLOv8m and YOLOv11m models were further trained on the created dataset. Using the test dataset, the accuracy and speed of YOLO models of different versions before and after retraining were compared. The detection accuracy was compared using the metrics of Recall, Precision, IOU, mAP50, mAP60-95. The experimental results indicate that the retrained models demonstrate higher accuracy in detecting car images compared to models without retraining. The accuracy of the YOLOv8m and YOLOv11m models after retraining is practically the same. It is shown that the YOLOv11m model provides higher speed in detecting objects and requires less memory, which is a significant advantage in resource-constrained environments. Neural network models have been successfully tested in tracking cars on real video recordings from traffic surveillance cameras. The results obtained demonstrate the practical effectiveness of retraining YOLO neural networks, as well as the feasibility of further research into improving YOLO architectures.