Abstract
Industrial fruit inspection systems must operate reliably under dense multi-object interactions and continuous motion, yet most existing works evaluate detection or classification at the image level without ensuring temporal stability in video streams. We present a two-stage detection-tracking framework for stable multi-apple quality inspection in conveyor-belt environments. An orchard-trained YOLOv8 model performs apple localization, followed by ByteTrack multi-object tracking to maintain persistent identities. A ResNet18 defect classifier, fine-tuned on a healthy-defective fruit dataset, is applied to cropped apple regions. Track-level aggregation is introduced to enforce temporal consistency and reduce prediction oscillation across frames. We define video-level industrial metrics such as track-level defect ratio and temporal consistency to evaluate system robustness under realistic processing conditions. Results demonstrate improved stability compared to frame-wise inference, suggesting that integrating tracking is essential for practical automated fruit grading systems.
Abstract (translated)
工业水果检测系统必须在密集的多对象交互和连续运动下可靠运行，然而大多数现有的研究工作仅通过图像层面来评估检测或分类效果，并未确保视频流中的时间稳定性。我们提出了一种两阶段的检测-跟踪框架，用于传送带环境中稳定地进行苹果质量检查。该框架首先使用经过果园数据训练的YOLOv8模型定位苹果位置，然后利用ByteTrack多对象跟踪技术保持持久的身份标识。接下来，一个基于健康与缺陷果实数据集微调后的ResNet18缺陷分类器被应用于裁剪出的苹果区域。此外，我们引入了轨迹级别的聚合方法来强制执行时间一致性，并减少帧间预测的波动。 为了评估系统在真实处理条件下的鲁棒性，我们定义了一些视频层面的工业指标，如轨迹级别上的缺陷比率和时间一致性等。实验结果表明，与基于单帧的推理相比，我们的方法显著提高了系统的稳定性，这说明对于实际的自动化水果分级系统而言，集成跟踪技术是至关重要的。
URL
https://arxiv.org/abs/2602.19278
