Accurate yield estimation of passion fruits is essential for planning acreage and harvest timing. However, due to the complexity of the natural environment and tracking instability, the existing yield estimation methods suffer from excessively large models that are difficult to deploy or repetitive counting of fruit. Therefore, an improved approach for efficient passion fruit yield estimation was proposed using the lightweight YOLOv5s and improved DeepSORT. First, the video is fed into the proposed lightweight YOLOv5s called YOLOv5s-little to obtain coordinates and confidence information about the fruits within each frame. Then, the information obtained from the detection model is input into improved DeepSORT for continuous frame tracking of passion fruit. Considering the frequent error IDs (ID switching), two improvements based on DeepSORT are proposed: delaying the creation of tracks and adding a second round of IoU matching. Finally, to overcome the problem of repetitive counting, a specific tracking counting method based on the track information and state is used for accurate passion fruit counting. Our method achieved a competitive result in tests. YOLOv5s-little detector achieved precision of 98.9%, 98.3% recall, 99.5% mAP, and only 0.9MB model size. The improved DeepSORT algorithm achieved higher order tracking accuracy (HOTA) of 79.6%, multi-object tracking accuracy (MOTA) of 92.58%, identification F1 (IDF1) of 95.02%, and ID switch (IDSW) of 11 respectively. Compared with DeepSORT, it improved by 4.66%, 1.8%, and 9.16% in HOTA, MOTA and IDF1, respectively, and IDSW improved the most with 85%. Compared with FairMOT and TransTrack, the HOTA of YOLOv5s-little + improved DeepSORT achieved improvements of 11.56% and 25.24%, respectively. The statistical average counting accuracy of our proposed counting method reaches 95.1%, which is a 7.09% improvement over the maximum ID value counting method. The counting results from test videos are highly correlated with the manual counting results (\({\text{R}}^{2}\) = 0.96), indicating that the counting method has high accuracy and effectiveness. These results show that YOLOv5s-little + improved DeepSORT can meet the practical needs of passion fruit yield estimation in real scenarios.

准确估算百香果的产量对于规划种植面积和收获时间至关重要。然而，由于自然环境的复杂性和跟踪的不稳定性，现有的产量估算方法存在模型过大、难以部署或重复计算果实的问题。因此，提出了一种使用轻量级 YOLOv5 和改进的 DeepSORT 进行高效百香果产量估算的改进方法。首先，视频被输入到提出的轻量级 YOLOv5（称为 YOLOv5s-little）中，以获得每帧内水果的坐标和置信度信息。然后，将从检测模型获得的信息输入到改进的DeepSORT中，用于百香果的连续帧跟踪。考虑到频繁的错误ID（ID切换），基于DeepSORT提出了两个改进：延迟轨道的创建和添加第二轮IoU匹配。最后，为了克服重复计数的问题，采用基于跟踪信息和状态的特定跟踪计数方法来精确计数百香果。我们的方法在测试中取得了有竞争力的结果。 YOLOv5s-little detector实现了98.9%的精度、98.3%的召回率、99.5%的mAP，并且模型大小仅为0.9MB。改进的DeepSORT算法分别实现了79.6%的高阶跟踪精度（HOTA）、92.58%的多目标跟踪精度（MOTA）、95.02%的识别F1（IDF1）和11的ID开关（IDSW）。与DeepSORT相比，在HOTA、MOTA和IDF1上分别提升了4.66%、1.8%和9.16%，其中IDSW提升最多，达到了85%。与FairMOT和TransTrack相比，YOLOv5s-little+改进的DeepSORT的HOTA分别取得了11.56%和25.24%的提升。我们提出的计数方法的统计平均计数精度达到95.1%，比最大ID值计数方法提高了7.09%。测试视频的计数结果与手动计数结果高度相关（\({\text{R}}^{2}\)  = 0.96），表明该计数方法具有较高的准确性和有效性。这些结果表明YOLOv5s-little+改进的DeepSORT可以满足真实场景下百香果产量估算的实际需求。