Research on AR target recognition algorithm for mobile based on YOLOv8

In this paper, the latest YOLO series algorithm YOLOv8 is studied in depth, and based on it, an improved target detection algorithm, designed for mobile, is proposed to realize an AR system for real-time recognition of industrial elements. Aiming at the problems of large number of model parameters, large computation, not easy to deploy, as well as recognition errors and instability due to light changes and viewpoint movement that occur in the deployment of target detection models on mobile, YOLOv8 is improved and YOLOv8-PDS is proposed. In this paper, a new structure SPC is proposed based on Partial Conv, which is used to replace the C2f in the network. module, this innovation significantly reduces the amount of computation and the number of parameters, and at the same time effectively improves the speed of the network as an efficient inference of the convolution. In addition, the article introduces an ultra-lightweight and efficient dynamic upsampling technique, DySample, and incorporates a spatial context-aware module, SCAM, to further improve the detection accuracy. These improvements ensure a significant reduction in the number of parameters, computation, and model size without compromising accuracy. With the lightweight improvements, the YOLOv8-PDS model reduces the computation by 22.2%, the number of parameters by 17.8%, and the model size by 17.5% compared to YOLOv8, making it easier to be deployed on AR devices, while achieving $\mathbf{9 0. 0 \%}$ accuracy and $\mathbf{0. 8 \%}$ recall improvement. In order to verify the ease of deployment and stability of YOLOv8-PDS, this paper designs an AR interactive system that converts YOLOv8-PDS models to ONNX format and implements the interactive mode through the Unity3D platform, so as to develop a mobile AR application with integrated target detection function.