以手機與車輛等移動平台上架設相機進行多物件追蹤之需求興起，而相異於單物件追蹤，多物件追蹤需考慮物件之間彼此量測相互影響的問題，而當運用於移動平台上，更須隨著背景的變動，仍能準確估計正確的相機動量以達成高準確度之物件追蹤。因此，本論文提出結合粒子濾波器估測之量測與聯合機率關聯性(JPDA)之多物件演算法，以粒子濾波器(PF)的估測作為JPDA之量測，且在PF的更正階段中，加入條件篩選對候選粒子們的量測，計算物件和量測間之關聯機率，以基於卡爾曼濾波器計算聯合機率數據關聯性，計算聯合機率關聯性。此外，在運動模型方面，本論文採用SURF作特徵點匹配，再結合仿射模型(affine model)協助計算相機動量，以供做於運動補償模型以補償物件在二維影像中之動量。從實驗結果中顯示，我們提出的演算法用於移動式相機之多物件追蹤中，也可達到良好的追蹤效果。The necessity of multiple object tracking based on mobile platforms such as mobile phones and vehicles are growing. Different from single object tracking, multiple object tracking needs to consider the data association problem between objects and measurements. Moreover, for mobile platforms, how to estimate camera motion accurately plays a key role in the success of object tracking. Therefore, this paper proposes a multiple object tracking algorithm that combines the particle filter (PF) based measurement and Kalman filter (KF) based joint probability data association (JPDA). Before the correction stage of PF, several particles may be filtered out based on the validation region of each object. In addition, this paper combines SURF matching and affine model to estimate camera motion for motion compensated model. Experimental results show that the proposed multiple object tracking algorithm based on mobile cameras performs well.
