YOLO-Rep: lightweight UAV visual tracking algorithm

In order to attain a more optimal trade-off between detection accuracy and detection speed of existing UAV tracking and detection algorithms, this paper designs a lightweight UAV tracking and detection algorithm YOLO-Rep based on the improved YOLO (You Only Look Once).The algorithm firstly takes the multi-scale fusion layer and the output detection layer of the YOLO as the neck network and the head network, respectively. The MobileNetV3 lightweight CNN is introduced to reconstruct the backbone network, and the MetaFormer architecture of lightweight ViTs (Vision in Transformer) is applied to split the token mixer and channel mixer of MobileNetV3 to achieve better lightweighting. Secondly, the block structure in the training backbone network is optimised to a multi-branch structure, which enables the kernel computation to run in parallel and improves the speed of model training. Finally, to simplify the inference model and further reduce the time used for model inference, the inference model is reconstructed using Structural Re-parameterization, which converts it to a single, more efficient convolutional operation during inference, improving the performance and training stability of the network. The model offers the benefits of both high detection accuracy and rapid processing speed, which is a great improvement over the original algorithm, and has a greater advantage in the UAV detection tracking and detection algorithm.