Evaluation of the Effectiveness of Convolutional Neural Networks in the Task of Autonomous Piloting

Abstract

This study is devoted to the development of a system for autonomous movement of the GreenCamel AgroMul RC2400 unmanned trolley across the territory of an enterprise. The key task is to create a machine vision system capable of real-time recognition of obstacles, such as traffic cones, and taking actions to bypass them. As a solution for the task of object detection, the pre-trained YOLOv11 algorithm, which is the latest version of a popular family of real-time object detectors, was chosen. YOLOv11 has an improved architecture and training methods, which makes it an effective tool for solving a wide range of computer vision problems. To train the model, data from the Microsoft COCO dataset was used, supplemented by a dataset of traffic cones from the open source roboflow. Of the 80 classes presented in Microsoft COCO, the 9 most relevant to the task were selected, and the "traffic cone" class was also added. During the training, the YOLOv11 model demonstrated high accuracy and efficiency, achieving a MAP (average value of the average accuracy indicator for all classes) value of 91.2%, Precision (accuracy) - 90.1% and Recall (recall) - 86.5%. As a result of the study, it was confirmed that YOLOv11 is an effective tool for detecting traffic cones. The results demonstrate the possibility of using the algorithm to create an autonomous movement system for the GreenCamel AgroMul RC2400 unmanned cart. As limitations of using the algorithm, it can be noted that: collecting and marking data for detection tasks are labor-intensive processes that require high-quality preparation of annotations; the problem of class imbalance may arise when adding specialized small classes, as in the case of "traffic cone". Despite these limitations, YOLOv11 is an effective tool for detecting environmental objects and obtaining data that can be used to solve other problems, such as determining the distance to an object by its area in the image.