YOLOv5-based Vehicle Objects Detection Using UAV Images

Abstract

Traffic is the situation and movement of pedestrians, animals, and vehicles on highways. The regulation of these movements and situations is also a basic problem of traffic engineering. It is necessary to collect data about traffic in order to produce suitable solutions to problems by traffic engineers. Traffic data can be collected with equipment such as cameras and sensors. However, these data need to be analysed in order to transform them into meaningful information. For a difficult task such as calculating and optimizing traffic density, traffic engineers need information on the number of vehicles to be obtained from the image data they have collected. In this process, artificial intelligence-based computer systems can help researchers. This study proposes a deep learning-based system to detect vehicle objects using YOLOv5 model. A public dataset containing 15,474 high-resolution UAV images was used in the training of the model. Dataset samples were cropped to 640×640px sub-images, and sub-images that did not contain vehicle objects were filtered out. The filtered dataset samples were divided into 70% training, 20% validation, and 10% testing. The YOLOv5 model reached 99.66% precision, 99.44% recall, 99.66% mAP@0.5, and 89.35% mAP@0.5-0.95% during the training phase. When the determinations made by the model on the images reserved for the test phase are examined, it is seen that it has achieved quite successful results. By using the proposed approach in daily life, the detection of vehicle objects from high-resolution images can be automated with high success rates.