具有相機感知能力之聯邦無監督行人重識別--均衡身份分布提升去中心化數據聚類效果

Abstract

隨著智慧城市的快速發展，行人重識別（pedestrian re-identification）已成為監控影像中個體追蹤的重要技術。然而，這些方法需要大量已標註的資料，而標註過程耗時且費工，使其難以廣泛應用於實務場域。因此，本研究採用無監督學習方法，以消除對標註資料的需求，提升模型於真實場景下的適用性。 此外，監控影像中涉及個人面貌與特徵等敏感資訊，使得集中式訓練模式在隱私保護上日益受到挑戰。為此，本研究設計一個具個別化能力的聯邦學習架構，使多個資料保留於本地的客戶端可在無需共享原始資料的前提下協同訓練模型，有效平衡隱私保障與辨識性能。本研究所提出之整合框架結合無監督學習與聯邦學習於行人重識別任務中。於無監督學習部分，導入一種考慮攝影機資訊的對比式損失函數，以提升模型跨視角區分身份的能力；於聯邦學習部分，針對傳統 FedAvg 方法中常見的資訊不平衡問題，設計一個改良的參數聚合策略。同時，我們亦提出一項前處理技術，以降低無監督學習中的聚類雜訊並提升各客戶端對聚合模型訓練的貢獻。 本方法於八組公開基準資料集上進行評估，實驗結果顯示其具備先進的準確率與隱私保護效能，驗證了所提方法於智慧監控系統中之應用潛力，並為未來智慧城市中的安全、可擴展與智慧化監控系統之發展提供重要貢獻。

With the rise of smart cities, pedestrian re-identification (re-ID) has become crucial for tracking individuals across surveillance footage. However, its reliance on large labeled datasets limits practicality due to the high cost of manual annotation. To address this, we adopt pure unsupervised learning, which eliminates the need for labeled data and enhances real-world applicability. Meanwhile, privacy concerns over sharing sensitive visual data make centralized training problematic. We propose a personalized federated learning (pFL) framework that enables decentralized clients to collaboratively train a model with out sharing raw data, preserving privacy while maintaining strong performance. Our framework combines unsupervised and federated learning for pedestrian re-ID. We introduce a contrastive loss leveraging camera-specific information and propose a refined aggregation strategy to address imbalances in traditional FedAvg. Additionally, a novel preprocessing method reduces clustering noise and improves each client’s contribution. Our method is evaluated on eight benchmark datasets, demonstrating state-of-the-art performance and validating its effectiveness in both preserving privacy and achieving high accuracy. By addressing the dual challenges of unsupervised learning and federated learning, this research significantly advances the development of secure, scalable, and intelligent surveillance systems for future smart cities.