LANA-YOLO: Road defect detection algorithm optimized for embedded solutions

Abstract

Poor pavement condition leads to increased risk of accidents, vehicle damage, and reduced transportation efficiency. The author points out that traditional methods of monitoring road conditions are time-consuming and costly, so a modern approach based on the use of developed neural network model is presented. The main aim of this paper is to create a model that can infer in real time, with less computing power and maintaining or improving the metrics of the base model, YOLOv8. Based on this assumption, the architecture of the LANA-YOLOv8 (Large Kernel Attention Involution Asymptotic Feature Pyramid) is proposed. The model's architecture is tailored to operate in environments with limited resources, including single-board minicomputers. In addition, the article presents Basic Involution Block (BIB) that uses the involution layer to provide better performance at a lower cost than convolution layers. The model was compared with other architectures on a public dataset as well as on a dataset specially created for these purposes. The developed solution has lower computing power requirements, which translates into faster inference times. At the same time, the developed model achieved better results in validation tests against the base model.