A Multiscale Detection and Tracking Method for Oscillating Fruit

During the fruit harvesting and thinning processes, natural wind forces and robotic pull‐based harvesting or thinning methods often induce fruit oscillation, which subsequently complicates robot localization and reduces the success rate of the operation. To address this issue, this article proposes a solution that combines the CHD‐YOLOv8n detection model with an optimized DeepSORT tracking algorithm. CHD‐YOLOv8n is an improved version of the YOLOv8n model. First, in response to the computational limitations of mobile devices, we introduce the CAA‐HSFPN (Channel Attention Adaptive Hybrid Spatial Feature Pyramid Network) to replace the original head network in YOLOv8n. This modification dynamically optimizes feature weight distribution, reducing the number of parameters and improving feature transmission efficiency. Second, to address the degradation in recognition caused by leaf occlusion, the C2f‐DAttention (C2f Dual Attention) module is proposed, which utilizes both spatial and channel attention mechanisms to enhance feature extraction in the fruit region. Finally, to further improve fruit detection accuracy, the SPPF‐LSKA (Spatial Pyramid Pooling with Local Spatial Kernel Attention) module is incorporated, enabling the model to effectively fuse features at multiple scales. To improve the tracking performance of DeepSORT in orchard environments, the original detector with CHD‐YOLOv8n is replaced, and the ResNest50 network is introduced to enhance the discriminative power of appearance features. Meanwhile, CIoU is used to replace traditional IoU matching, optimizing the dynamic association of fruits. To reduce identity loss in oscillating fruit scenes, an adaptive noise‐scale Kalman filter is designed. The experimental results show that the CHD‐YOLOv8n model achieves mAP@0.5 of 95.23% and 96.18% for detecting young and mature peaches, respectively, with both precision and recall exceeding 91%. When combined with theoptimized DeepSORT algorithm, the tracking accuracy improves by 13.2–16.5% compared to traditional SORT, while the number of ID switches is reduced by 50–59.46%. These technical innovations provide an efficient and stable solution for intelligent harvesting and thinning robots.