Abstract:A lightweight litchi detection model FEB-YOLO was proposed to detect litchi with small target, overlap and occlusion in large natural environment. Based on YOLOv8, the model introduced PConv to replace part of the conventional convolution in the C2f module to achieve lightweight improvement, and integrated EMA attention mechanism to improve the feature extraction capability of the algorithm. The neck network was replaced by BiFPN with P2 feature layer to enhance the cross-scale feature fusion of different sizes. The NWD measure was introduced into the regression loss function to improve the model's learning ability for litchi small targets and reduce the rate of missing detection. The experimental results show that the P, R and mAP of FEB-YOLO model are increased by 1.4%, 1.6% and 1.7%, respectively, compared with the original model, and the number of Params and FLOPs are reduced by 47.3% and 27.1%, respectively. The improved model has less computing resources, and can improve the recognition accuracy under complex conditions. In order to realize real-time estimation of litchi yield in orchard scene, an efficient dynamic recognition and counting method of litchi fruit is proposed. By using FEB-YOLO as the target detector of BoT-SORT tracker and the recognition output of FEB-YOLO as the input of BoT-SORT, dynamic video sequence tracking and counting is realized. Finally, an example is given to verify the effectiveness and feasibility of the proposed method. The improved model has good robustness and small size, and can be embedded in the edge equipment, which can not only be used for real-time estimation of litchi yield, but also for planning picking and storage, providing reliable support for orchard resource allocation.
