Infrared video-based object recognition for agricultural research field uses

Kazan Federal University, Institute of Information Technology and Intelligent Systems, 420008, 35 Kremlevskaya st., Kazan, Russia.

^* Corresponding author: vlada.kugurakova@gmail.com

Abstract

The paper presents a study of an object recognition system based on infrared (IR) video stream using the YOLOv8 model. The aim of the work was to develop and evaluate the performance of a computer vision algorithm for analysing infrared images under various conditions including low light. While the use of RGB images can provide rich information in normal conditions, it becomes inefficient to use a visible spectrum camera in total darkness or in the presence of clutter. The study was based on an extensive dataset containing about 176,000 annotated frames from an infrared camera. The YOLOv8 model was adapted and trained to recognise 15 classes of objects, including people, animals and different types of vehicles. The results showed high performance of the model in recognising the most represented classes in the dataset, such as cars, people with F1-score up to 0.8. However, the problem of unbalanced classes was identified, leading to lower recognition accuracy for rarely seen objects. The study demonstrates the potential of using infrared cameras in computer vision tasks and opens new directions for further developments in the field of object recognition in limited visibility conditions. The results of the work may find application in a wide range of disciplines, including medicine, agriculture and environmental monitoring.