物件分析與追蹤於許多視覺影像處理的問題中,扮演著極重要的角色,當然亦可做為許多延伸應用的前處理部份,例如:行為分析於大型監控環境系統下、行人異常事件抽取之監控分析系統…等。 於多監視器之監控環境下,可以利用監視器監控環境彼此間重疊的區域,給予校正並獲得彼此特徵對應關係,藉此即可進行移動物的追蹤與特徵校正,然而於非重疊之多相機監控環境下,並無法套用上述方法順利進行移動物於多監視器之間的追蹤,因此,本篇論文將針對上述問題設計出一套系統,藉以達成於非重疊多相機之監控環境下進行移動物的特徵分析與追蹤,並會針對提出的演算法與系統進行驗證與系統的測試實驗。 由實驗結果得知,我們的系統可以有效的適用於不同環境底下,並且有很好的表現。
Object tracking and identification play an important role in many multimedia processes. This paper proposes a novel approach for pedestrians analyzing and tracking between multiple non-overlapping cameras. Traditional methods tried to analyze and track pedestrians between multiple cameras using their color transformation. However, the color feature is unstable under different lighting conditions and especially will change when the view is changed to another one. In addition, lots of training data are required for training the color transformation between views. To tackle the above problems, this paper proposes a framework which includes not only the spatial model but also the temporal features for well analyzing pedestrians even though they are observed under two non-overlapping cameras. In the spatial model, the appearance and geometry feature of pedestrians are included for extracting their invariant properties among different camera views. To reduce the effects of illumination change, instead of modeling the whole body, a component-based scheme is proposed for modeling a pedestrian’s appearances up to his body parts. In temporal model, we use speed and probability information between views as our measuring features. Experimental results reveal the performances of our system in several different conditions.