基于轻量化YOLO 11n‑CLL和BoT‑SORT的百香果计数方法

Passion Fruit Counting Method Based on Lightweight YOLO 11n‑CLL and BoT‑SORT

摘要:

百香果轻量化检测、跟踪与计数是智慧农业中实现智能化采摘的关键技术。在真实果园场景中，百香果与枝叶的重叠遮挡与环境光照变化会导致目标出现漏检、误检及重复计数等问题。针对上述问题，本文提出了一种YOLO 11n?CLL融合BoT?SORT的轻量化计数方法。在目标检测中，YOLO 11n?CLL利用YOLO 11n作为基线模型，通过引入轻量化上下文引导模块，增强上下文感知能力;融合大核可分离注意力机制，提升多尺度特征建模能力;使用细节增强检测头，提升遮挡果实的检测能力。在目标跟踪部分，BoT?SORT跟踪器利用相机运动补偿模块，抑制抖动干扰，提升跟踪精度。在果实计数任务中，选择最佳的区域计数方法，实现果实准确计数。试验结果表明，YOLO 11n?CLL平均准确率为87.0%，模型内存占用量为5.0 MB;BoT?SORT跟踪器HOTA（Higher order tracking accuracy）和MOTA（Multiple object tracking accuracy）分别为62.4%和68.4%;采用区域计数法对果实进行计数，其平均计数准确率达到92.8%，比划线计数法和ID累加计数法分别高8.8、29.9个百分点。试验结果表明本文方法能实现百香果检测、跟踪与计数，为百香果园智能化采摘提供技术支撑。

Abstract:

Lightweight detection, tracking, and counting passion fruit are crucial technologies for achieving intelligent harvesting in smart agriculture. In real orchard environments, factors such as overlapping occlusion with leaves and branches, along with lighting variations, often result in missed detections, false detections and repeated counting. To address these challenges, a lightweight counting framework was proposed based on YOLO 11n?CLL and BoT?SORT. In object detection, YOLO 11n?CLL was built upon YOLO 11n as its baseline model. Firstly, a lightweight context guided block was introduced to enhance contextual perception capability. Nextly, the large separable kernel attention mechanism was incorporated to improve multi?scale feature modeling. Finally, the lightweight shared detail?enhanced convolutional detection head was employed to effectively reconstruct the texture features of occluded fruits. For object tracking, the BoT?SORT tracker was adopted, which incorporated a camera?motion?compensation module to mitigate camera?shake interference and improve tracking accuracy. In fruit counting task, a region?based counting method was adopted to achieve accurate counting of passion fruits. Results showed that YOLO 11n?CLL achieved mAP@0.5 of 87.0% with a compact 5.0 MB model size. The BoT?SORT tracker achieved a HOTA of 62.4% and a MOTA of 68.4%. The region?based counting method yielded an average counting accuracy of 92.8%, outperforming the line?based counting and ID?based counting method by 8.8 percentage points and 29.9 percentage points. The results demonstrated that the proposed framework enabled detection, tracking, and counting of passion fruits and provided technical support for intelligent harvesting in passion fruit orchards.