DSpace collection: 博碩士論文

小型語言模型在多目標摘要管線中的最佳化研究: 品質、效率、碳排約束下的設計探索;Small Models， Smarter Pipelines: Multi-Objective Optimization of SLM Summarization Systems under Quality， Efficiency， and Carbon Constraints

title: 小型語言模型在多目標摘要管線中的最佳化研究: 品質、效率、碳排約束下的設計探索;Small Models， Smarter Pipelines: Multi-Objective Optimization of SLM Summarization Systems under Quality， Efficiency， and Carbon Constraints abstract: 以大型語言模型（Large Language Model， LLM）為基礎的文本摘要已成為當前主流，然而既有研究多以單一品質指標為最佳化目標，鮮少關注摘要系統日益複雜後所衍生的效率與碳排成本；而以小型語言模型（Small Language Model， SLM）為核心的替代方案，在多目標限制下的表現亦缺乏系統性探討。本研究提出多目標最佳化摘要系統（Multi-Objective Optimization Summary System， MOOSS）—— 一個以 SLM 為基礎的模組化摘要平台，並建立三維度的品質–效率–碳排（Quality–Efficiency–Carbon， QEC）評估框架，配合漸進式 QEC 引導配置搜尋（Progressive QEC-Guided Configuration Search， PQCS）方法，逐階段收斂設計空間。本研究於四份基準資料集（BBC News、CNN/DailyMail、Multi-News、PubMed）上評估 700 組以上配置，涵蓋四種提示策略、兩個模型家族及四種規模、七種嵌入模型，以及多項連續參數。實驗整合出五項主要發現：第一，8B 以下的 SLM 在四份資料集中之三份接近或超越 API 模型上限（GPT5.1），API 模型僅在最簡單的 BBC News 之詞彙指標上領先，而回饋式檢索增強進一步縮小多文件摘要任務的差距。第二，回饋式檢索增強在半數以上的「模型 × 策略」格中帶來正向收益，且收益幅度隨任務複雜度與模型容量擴大。第三，系統因素之重要性具有明顯層級：提示策略與模型規模主導品質變異，而系統模式與切塊策略對品質貢獻可忽略。第四，如嵌入模型選擇等細部元件，雖對品質影響極小（七種嵌入間 ROUGE-1 差距僅 3.2%），卻造成顯著的效率與碳排差距（極端配置間能耗與碳排相差 1.27 倍、VRAM 相差 2.80 倍）。第五，連續參數調整在多數情境下為 Pareto 改善（Pareto-improving）：以最佳品質為調整目標時，效率變化 Δ?

STARTrack：基於方向一致性軌跡關聯之小目標多物件追蹤方法;STARTrack: Direction-Consistent Trajectory Association for Small Target Multi-Object Tracking

title: STARTrack：基於方向一致性軌跡關聯之小目標多物件追蹤方法;STARTrack: Direction-Consistent Trajectory Association for Small Target Multi-Object Tracking abstract: 視覺多物件追蹤技術廣泛應用於環境監測、無人機監控及國防安全等領域。在遠距離監控與高空觀測等應用場景中，目標物件通常僅占影像中極少數像素，且常伴隨低解析度、低對比度、邊界模糊、紋理資訊不足以及目標外觀相似等問題，使得小目標偵測與多物件追蹤成為極具挑戰性的研究課題。現有多物件追蹤方法多採用偵測式追蹤 (Tracking-by-Detection) 架構，先透過物件偵測模型取得目標位置，再進行跨幀的物件關聯。然而，傳統偵測模型在多層下採樣過程中容易造成小目標特徵流失，而依賴外觀特徵的關聯方法則容易受到目標尺度過小及外觀差異不明顯等因素影響，導致錯誤匹配、身分切換及軌跡中斷等問題。本研究提出 STARTrack(Small-object Trajectory-Aware Robust Tracking)，一套基於方向一致性軌跡關聯之小目標多物件追蹤方法，從目標偵測與軌跡追蹤兩個層面進行改進。在目標偵測方面，本研究以 YOLOv11 為基礎架構，引入高解析度特徵層並擴展多尺度偵測頭設計，以強化淺層空間資訊的保留能力，使模型能更有效擷取小尺度目標特徵，提升極小目標的辨識能力與定位精度。在目標關聯方面，本研究提出結合交集比 (Intersection over Union，IoU) 與歷史軌跡中心距離之匹配成本函數，使關聯過程同時考量目標間的空間重疊程度與運動連續性，降低因輕微位移或偵測誤差所造成的匹配不穩定。此外，本研究亦針對卡爾曼濾波之量測噪聲模型進行調整，透過重新配置量測噪聲權重，使狀態估測更符合小目標場景特性，進一步減少追蹤過程中的邊界框漂移現象。最後，為提升追蹤軌跡管理的可靠性，本研究提出基於方向一致性的重複軌跡移除機制，藉由分析軌跡運動方向與方向相似度，有效降低目標交叉或短暫重疊時所產生的誤刪與身分交換問題。實驗結果顯示，本研究所提出之小目標偵測模型於 Anti-UAV Challenge Track3 資料集上，在 Precision、Recall、mAP50 與 mAP50-95 等評估指標皆優於現有主流偵測模型，展現優異的小目標偵測能力。在多物件追蹤方面，STARTrack 於紅外線地面目標追蹤資料集與 Anti-UAV Challenge Track3 資料集上皆取得最佳的 MOTA、IDF1 與 HOTA 表現，有效降低身分切換次數，驗證 STARTrack 在小目標監控場景下兼具高偵測準確率、優異的追蹤穩定性與身分一致性，具備即時追蹤之應用潛力。;Visual multi-object tracking (MOT) techniques have been widely applied in fields such as environmental monitoring， unmanned aerial vehicle (UAV) surveillance， and national defense. In applications involving long-distance surveillance and high-altitude observation， targets typically occupy only a small number of pixels in an image and are often affected by low resolution， low contrast， blurred boundaries， insufficient texture information， and similar target appearances. These factors make small-object detection and multi-object tracking highly challenging research problems. Existing multi-object tracking methods commonly adopt a Tracking-by-Detection framework， in which an object detection model is first used to obtain target locations， followed by cross-frame object association. However， conventional detection models tend to lose small-object features during successive downsampling operations. Meanwhile， association methods that rely on appearance features are susceptible to factors such as extremely small target sizes and insignificant appearance differences， resulting in incorrect matching， identity switches， and fragmented trajectories. This study proposes STARTrack (Small-object Trajectory-Aware Robust Tracking)， a small-object multi-object tracking method based on direction-consistent trajectory association， with improvements made at both the object detection and trajectory tracking levels. For object detection， YOLOv11 is adopted as the base architecture， with high-resolution feature layers introduced and multi-scale detection heads extended to enhance the preservation of shallow spatial information. This design enables the model to more effectively extract features from small-scale targets， thereby improving the detection capability and localization accuracy of extremely small objects. For target association， a matching cost function that combines Intersection over Union (IoU) and the center distance of historical trajectories is proposed. The association process simultaneously considers the degree of spatial overlap and motion continuity between targets， thereby reducing matching instability caused by slight target displacement or detection errors. In addition， the measurement noise model of the Kalman Filter is adjusted by reconfiguring the measurement noise weights， enabling state estimation to better accommodate the characteristics of small-object scenarios and further reducing bounding-box drift during tracking. Finally， to improve the reliability of trajectory management， a direction-consistency-based duplicate trajectory removal mechanism is proposed. By analyzing the motion directions and directional similarities of trajectories， the proposed mechanism effectively reduces incorrect trajectory removal and identity switches caused by target crossing or temporary overlap. Experimental results demonstrate that the proposed small-object detection model outperforms existing mainstream object detection models on the Anti-UAV Challenge Track 3 dataset in terms of Precision， Recall， mAP50， and mAP50-95， demonstrating its excellent small-object detection capability. For multi-object tracking， STARTrack achieves the highest MOTA， IDF1， and HOTA performance on both the infrared ground-target tracking dataset and the Anti-UAV Challenge Track 3 dataset. The proposed method effectively reduces the number of identity switches， demonstrating that STARTrack provides high detection accuracy， excellent tracking stability， and strong identity consistency in small-object surveillance scenarios， while demonstrating its potential for real-time tracking applications.

分散式即時多視角多人 3D 人體姿態估計系統;分散式即時多視角多人 3D 人體姿態估計系統

title: 分散式即時多視角多人 3D 人體姿態估計系統;分散式即時多視角多人 3D 人體姿態估計系統 abstract: 隨著虛實互動展演的發展，以人體姿態估計取代穿戴式設備，讓使用者直接以動作與設施互動，逐漸成為重要的應用需求。這類場域多為多人互動且容易產生遮擋，須以多台固定攝影機從多視角同時拍攝，並維持自影像擷取至 3D 姿態輸出全程的即時性。然而，現有多視角 3D 人體姿態估計研究多採循序式架構，難以滿足多視角處理的低延遲需求；且在即時性較佳的三角測量方法中，2D 偵測誤差會直接導致 3D 骨架變形。針對上述動機，本研究設計並實作一套無須專用硬體的多視角多人 3D 姿態估計端到端系統，具體貢獻包含：設計影像擷取、2D 推論、多視角資料處理、人物姿態重建與輸出五層分散式架構，以並行處理降低系統整體延遲；整合時間同步、時間差過濾與補幀機制，在視角延遲下維持運算所需的有效視角數量；並導入骨長修正方法，改善 2D 偵測誤差造成的骨架異常。實驗結果顯示，本系統能連續七天穩定運作，於有人活動時段內 99.986% 的時間維持 30 FPS 以上，且端到端延遲低於 60 毫秒。在 CMU Panoptic 資料集上，骨長修正機制將骨架長度誤差（Bone Length Error， BLE）由 26.28 mm 降至 22.11 mm；在模擬視角延遲情境下，時間同步與補幀機制將 BLE 由 22.28 mm 降至 21.34 mm。驗證了本系統在維持低延遲的同時，能有效維持 3D 骨架品質。;With the development of virtual-real interactive exhibitions， replacing wearable devices with human pose estimation (HPE) allows users to interact directly through body motion. Such scenarios involve multi-person interaction and frequent occlusion， requiring multiple fixed cameras to capture the scene from multi-view while maintaining real-time performance from image capture to 3D pose output. However， existing multi-view 3D HPE studies mostly adopt sequential architectures， struggling to meet low-latency demands. Furthermore， while triangulation offers low latency， 2D detection errors directly cause 3D skeleton deformation. To address these issues， this study designs and implements an end-to-end multi-view multiperson 3D pose estimation system without specialized hardware. The main contributions include: designing a five-layer distributed architecture (image capture， 2D inference， multi-view data processing， pose reconstruction， output) to reduce latency via parallel processing; integrating time synchronization， time-difference filtering， and frame fill-in mechanisms to maintain sufficient views under delays; and adopting bone length correction to fix 2D-induced skeletal anomalies. Experimental results show that the system operates stably for seven days. During human activity periods， it maintains 30 FPS or above and an end-to-end latency below 60 ms for 99.986% of the time. On the CMU Panoptic dataset， bone length correction reduces the bone length error (BLE) from 26.28 mm to 22.11 mm， while time synchronization and fill-in mechanisms reduce it from 22.28 mm to 21.34 mm under simulated delays， verifying both real-time performance and skeleton quality.

Design and Implementation of a Risk-Aware Dual-Track Retrieval-Augmented Generation Framework for Elderly Physiological and Psychological Intent Understanding;Design and Implementation of a Risk-Aware Dual-Track Retrieval-Augmented Generation Framework for Elderly Physiological and Psychological Intent Understanding

title: Design and Implementation of a Risk-Aware Dual-Track Retrieval-Augmented Generation Framework for Elderly Physiological and Psychological Intent Understanding;Design and Implementation of a Risk-Aware Dual-Track Retrieval-Augmented Generation Framework for Elderly Physiological and Psychological Intent Understanding abstract: 高齡者在表達身體不適、情緒困擾與社會孤立時，經常使用簡短、間接或同時涉及多個面向的語句，使傳統關鍵字式對話系統難以選擇適當的知識來源與回應策略。此外，語音語調與臉部表情雖可提供額外的情感脈絡，但此類模型輸出具有不確定性，不應被直接視為臨床診斷或風險判定依據。為處理上述問題，本論文設計並實作一套用於高齡者生理與心理意圖理解之風險感知雙軌檢索增強生成框架，目標是在非臨床情境下提供具證據條件、可檢查且較為審慎的對話支援。所提出的原型整合網頁式互動、語音轉文字、音訊情緒辨識、音訊與影像之憂鬱相關情感偵測、確定性意圖路由、密集向量檢索、大型語言模型回應生成，以及執行軌跡紀錄。系統依目前文字訊息將操作性意圖分類為生理、心理、混合或未知，並據此搜尋生理知識庫、心理知識庫或兩者。多模態情緒輸出僅作為調整同理程度與語氣的輔助脈絡，不參與知識庫選擇或確定性自我傷害規則。一般訊息會將目前訊息、篩選後的使用者對話歷史、情緒狀態與檢索證據共同提供給語言模型，並透過提示限制避免診斷、藥物指示與臨床決策。當文字符合預先定義的明確自我傷害片語時，系統則略過一般生成流程，搜尋輔助安全知識庫以供追溯，並回傳固定的求助建議。本研究以控制式工程評估檢驗原型，共執行五十六個單輪測試案例與六組、共十二輪的多輪對話。所有單輪請求均成功完成。意圖路由準確率為百分之六十二點五，巨集平均 F1 為零點六四零。生理與心理檢索的 Hit@5 分別為百分之七十五點零與百分之八十六點七，混合訊息同時取得兩軌相關證據的比例為百分之八十七點五。針對目前規則所涵蓋的自我傷害片語，召回率為百分之八十七點五；然而，其精確率僅為百分之五十點零，困難負例的特異度為百分之十二點五，且對未包含既定片語的語意高風險案例之觸發率為百分之零。人工結構化評分顯示，回應相關性為百分之六十點七，證據基礎性為百分之三十七點五，正確回應行動為百分之四十六點四；另一方面，所有受評回應皆避免明確診斷與直接治療指示。平均端到端延遲為九百三十五點六毫秒，第九十五百分位延遲為一千七百三十八點一毫秒。多輪測試中，第二輪回應正確聚焦目前訊息與避免前一主題延續的比例皆為百分之五十點零。結果顯示，本研究成功建立一套可執行、模組化且具執行層級可追溯性的高齡者對話支援原型，雙軌密集檢索亦能在多數可回答案例中取得相關證據。然而，確定性路由對間接或模糊表達的處理能力有限，語言模型未能穩定使用已檢索證據，片語式安全規則容易受到否定、引用、第三人稱與語意改寫影響，多輪主題隔離亦仍不足。因此，本系統僅應被視為技術研究原型，不可取代醫師、心理專業人員、照護者或緊急支援服務。未來研究應優先改善語意與多語言路由、獨立檢索評估、證據重排序與生成後查核、情境感知安全分類、多輪歷史選擇，以及由專家與真實高齡使用者參與的倫理審查研究。;Older adults may express physical discomfort， emotional distress， and social isolation through brief， indirect， or overlapping statements， making it difficult for conventional keyword-based conversational systems to select appropriate knowledge and response strategies. Vocal and facial behavior may provide additional affective context， but such model outputs are uncertain and should not be treated as clinical diagnoses or independent risk decisions. To address these challenges， this thesis designs and implements a risk-aware Dual-Track Retrieval-Augmented Generation framework for elderly physiological and psychological intent understanding. The framework is intended to provide evidence-conditioned， inspectable， and cautious conversational support in a non-clinical setting. The prototype integrates web-based interaction， speech-to-text processing， audio emotion recognition， audio-video depression-related affect detection， deterministic intent routing， dense vector retrieval， Large Language Model response generation， and execution tracing. The current textual message is operationally classified as physiological， psychological， mixed， or unknown， and this result determines whether the physiological repository， psychological repository， or both are searched. Multimodal emotion outputs are used only as auxiliary context for empathy and tone; they do not select a repository or activate the deterministic self-harm rule. For normal messages， the prompt combines the current message， filtered user-side history， emotional-state context， and retrieved evidence while prohibiting diagnosis， medication instructions， and clinical decisions. When the text matches a predefined explicit self-harm phrase， the system bypasses normal generation， searches an auxiliary safety repository for traceability， and returns a fixed recommendation to seek immediate human or professional support. The prototype was examined through a controlled engineering evaluation containing 56 single-turn cases and six multi-turn conversations with 12 total turns. All single-turn requests completed successfully. Operational intent routing achieved 62.5\% accuracy and a macro-F1 score of 0.640. Physiological and psychological retrieval achieved Hit@5 values of 75.0\% and 86.7\%， respectively， while 87.5\% of mixed cases retrieved relevant evidence from both tracks. For self-harm phrases covered by the implemented rule， recall was 87.5\%; however， precision was 50.0\%， hard-negative specificity was 12.5\%， and none of the semantic high-risk challenge cases without the predefined wording activated the high-risk branch. Structured response assessment produced pass rates of 60.7\% for relevance， 37.5\% for evidence grounding， and 46.4\% for correct response action. In contrast， all assessed responses avoided explicit diagnosis and direct treatment instructions. Mean end-to-end latency was 935.6~ms， with a 95th-percentile latency of 1，738.1~ms. In the multi-turn evaluation， only 50.0\% of second-turn responses remained focused on the current message and avoided carryover from the previous topic. These findings demonstrate the technical feasibility of an executable， modular， and execution-traceable elderly-support prototype， and show that Dual-Track dense retrieval can return relevant evidence for most answerable cases. Nevertheless， deterministic routing remained weak for indirect or ambiguous language， the language model did not consistently use retrieved evidence， phrase-based safety control was vulnerable to negation， quotation， third-person references， and semantic paraphrases， and multi-turn topic isolation was insufficient. The system must therefore be regarded only as a research prototype and not as a replacement for physicians， mental-health professionals， caregivers， or emergency-support services. Future work should prioritize semantic and multilingual routing， independent retrieval evaluation， reranking and evidence thresholds， post-generation grounding checks， context-aware safety classification， improved history selection， and ethically reviewed evaluation with experts and older adult users.