物件追蹤在電腦視覺領域被廣泛應用，其中行人追蹤在視覺監控(visual surveillance)系統裡是相當重要的，然而環境中之其他靜態或動態物件常影響行人原本的運動行為，也可能造成遮蔽，而靠近或遠離相機會造成行人在畫面裡放大或縮小。因此，本論文提出結合交互多模型粒子濾波器與SURF特徵匹配作目標偵測之物件追蹤演算法。藉由粒子濾波器的更正階段計算出各運動模型之最大色彩權重對應的狀態，更新運動模式機率(mode probability)，提升運動模式機率(mode probability)的準確性。而交互多模型粒子濾波器以前一時刻的運動模式機率(mode probability)，於交互多模型粒子濾波器交互階段更新混合機率，使混合後之各運動模型對應之狀態分布，更趨近於行人目前時刻的狀態之事前機率分布，進而提升預測準確率。此外，本論文參考各運動模型經粒子濾波器估測的狀態，交互多模型粒子濾波器之整體估測狀態，與以SURF特徵匹配所得之目前畫面多個匹配特徵點為中心之不同大小的候選方框區域，計算與目標樣板的色彩相似度，增加追蹤的準確率。最後利用外觀相似度判斷，進行物件的外觀模型更新，可防止目標因外觀大小改變、光線變化影響色彩相似度的判斷。實驗結果顯示，相較於交互多模型粒子濾波器，我們提出的方案, 在行人的運動行為改變, 遮蔽, 光線變化, 放大縮小的情況下，相較於以色彩為基礎之IMMPF演算法, 皆有較好的追蹤效果。;Object tracking is widely used in applications of computer vision where pedestrian tracking is important in the visual surveillance system. However, either static or dynamic objects in the environment may frequently affect the motion model of the pedestrian or lead to occlusions. Moreover, scaling is inevitable. Therefore, this paper proposes to combine interacting multiple model particle filter (IMMPF) and SURF feature matching based target detection for pedestrian tracking. The mode probability of each motion model is updated by the state with the maximum color weight of the corresponding motion model, computed by the correction stage of the particle filter. And thus, accuracy of the mode probability is improved. The mixing probability is updated by the previous mode probability in the interaction stage of IIMMPF. It increases accuracy of approximation of the mixed a priori probability distribution of the pedestrian and thus improves prediction accuracy. To improve tracking accuracy, the proposed scheme refers to the estimated state of each motion model, the overall estimated state of IMMPF, and the neighborhood with varying size of SURF matched keypoints to compute the color similarity with the target template. Finally, target appearance model is optionally updated according to the similarity of appearance model. Experimental results show that the proposed scheme outperforms the color based IMMPF algorithm.
