以YOLO-DSC模型最佳視角進行果梗定位應用於視覺引導型果萼保留小蕃茄採收機

Abstract

機器人採收技術雖已快速發展，但在植株冠層密集的環境中，小型果實的選擇性採收仍具挑戰。對臺灣鮮食市場的小番茄而言，保留果萼對維持果實新鮮度與外觀品質至關重要。其主要感知挑戰在於辨識完全成熟的果實、定位細小且經常受遮蔽的果梗，並在溫室雜亂環境與視角變化下取得可靠的切割參考。既有小番茄機器人採收研究多著重於整串採收、固定視角估測、僅進行感知層級驗證，或採用計算量較高的三維與多階段果梗定位方法。相較之下，本論文提出一套可信賴感知架構，整合直接二維 YOLO-DSC（You Only Look Once - Dynamic Snake Convolution）果梗定位、主動式最佳視角選擇、目標連續性維持與雷射深度確認。 本論文主要有三項貢獻。第一，為抑制溫室環境中的干擾物，同時維持對細長果梗的感知能力，本研究結合無目標資料（null data）訓練與骨幹網路層級的 Dynamic Snake Convolution (DSC)。加入無目標資料後，成熟小番茄的 precision 由 0.8958 提升至 0.9622，False Discovery Rate 由 10.42% 降低至 2.82%（p < 0.00001）。在 15 次單果試驗中，YOLO-DSC 未出現誤檢或漏檢，Mean Absolute Error 為 5 pixels，Root Mean Square Error 為 7 pixels；相較於 baseline YOLO，其改善達統計顯著（p < 0.05），相較於加入無目標資料的 YOLO，則呈現接近統計顯著的改善趨勢（p = 0.077）。YOLO-DSC 每次推論時間為 63.52 ± 5.60 ms，達到 15.7 FPS，具備即時偵測所需的運算效能；在跨研究的數值比較中，其推論速度亦高於 MTA-YOLACT 所報告的 13.3 FPS 與 R101-FPN YOLACT++ 的 9.2 FPS。 第二，為克服固定視角切割點估測的限制，本研究採用動態 yaw-pitch 視角搜尋，並以影像中可見的 K1-K2 果梗長度作為具可解釋性的可靠度指標，再進行雷射深度確認。 兩階段實驗的代表性分析顯示，在成功採收案例中，完成 yaw 對準後，pitch 調整可進一步增加可見果梗長度；相較之下，可靠度較低的案例則未呈現 pitch 階段的改善。 第三，為使所選目標不侷限於單一感知階段，本研究結合目標記憶、追蹤、跨相機投影與果梗配對，以在整體機器人作業流程中維持目標特定的感知參考。在多果實際部署中，11 個目標中有 10 個仍可取得有效果梗候選（90.9%），其 Mean Absolute Error 為 2.50 pixels，Root Mean Square Error 為 4.81 pixels。其餘限制主要涉及目標連續性、跨相機投影與切割後處理。研究結果顯示，YOLO-DSC 最佳視角感知可為保萼小番茄採收提供可靠的切割參考，而較長的連續採收流程仍仰賴穩定的目標追蹤與系統操作。

Robotic harvesting has advanced rapidly, yet selective harvesting of small fruits remains difficult in dense canopies. For Taiwanese fresh-market cherry tomatoes, preserving the calyx is essential for freshness and visual quality. The key perception challenge is to distinguish fully ripe fruit, localize the thin and often occluded pedicel, and obtain a reliable cutting reference under greenhouse clutter and changing viewpoints. Prior cherry-tomato harvesting studies commonly rely on bunch-level removal, fixed-view estimation, perception-only validation, or computationally intensive 3D and multi-stage pedicel localization. This thesis instead develops a trustworthy perception framework combining direct 2D YOLO-DSC (You Only Look Once - Dynamic Snake Convolution) pedicel localization, active best-view selection, target continuity, and laser depth confirmation. This thesis makes three main contributions. First, to suppress greenhouse distractors while preserving sensitivity to slender pedicels, null-data training is combined with backbone-level Dynamic Snake Convolution (DSC). Null data increased ripe-tomato precision from 0.8958 to 0.9622 and reduced the False Discovery Rate from 10.42% to 2.82% (p < 0.00001). Across 15 single-fruit trials, YOLO-DSC achieved no false positives or missed detections, with a Mean Absolute Error of 5 pixels and Root Mean Square Error of 7 pixels, significantly outperforming baseline YOLO (p < 0.05) and showing a near-significant improvement over YOLO with null data (p = 0.077). At 63.52 "±" 5.60 ms per inference, it achieved 15.7 FPS, competitive to real-time detection performance and numerically faster than the reported cross-study comparisons: 13.3 FPS of MTA-YOLACT and 9.2 FPS of R101-FPN YOLACT++. Second, to overcome fixed-view cutting-point estimation, dynamic yaw-pitch searching uses visible K1-K2 pedicel length as an interpretable reliability cue before laser depth confirmation. Representative analyses from both experiments showed pitch-dominant improvement in successful harvesting cases, with visible pedicel length increasing after yaw alignment, whereas less reliable cases showed no pitch improvement. Third, to maintain the selected target beyond isolated perception, target memory, tracking, cross-camera projection, and pedicel association preserve the target-specific reference through the robotic workflow. In multi-fruit deployment, valid pedicels remained available for 10 of 11 targets (90.9%), with a Mean Absolute Error of 2.50 pixels and Root Mean Square Error of 4.81 pixels. The remaining limitations involved target continuity, cross-camera projection, and post-cut handling. These findings show that YOLO-DSC best-view perception provided a reliable cutting reference for calyx-preserving harvesting, while longer sequential harvesting still depended on stable target tracking and system handling.