在電腦視覺與機器人領域中,影像追蹤是一個很重要的議題。全方位攝影機提供了較寬廣的視野,但是利用全方位攝影機進行影像追蹤,必須克服變形及低解析度的問題。這篇論文利用循序權重取樣粒子濾波器(Sequential Importance Sampling Particle Filter)提出了一個使用全方位攝影機追蹤多目標物的方法,為了快速偵測到目標物的分布,我們提出以前景偵測為基礎的權重取樣演算法(Foreground-Based Importance Sampling Mechanism),使粒子能快速有效的散佈在目標物附近,此外融合顏色與輪廓的資訊估測影像相似度(Likelihood Measurement),此方法可以更準確追蹤目標物,並藉由整合兩個空間的影像相似度增強系統的穩定性,克服全方位攝影機造成的變形問題。最後,透過實驗來驗證此系統整體的效能及可靠性。
Visual tracking is an important topic in computer vision and robotics fields. The omnidirectional cameras provide a wider filed of view, but tracking with an omnidirectional camera must overcome the warping and low resolution drawback. This thesis presents an approach based on the sequential importance sampling particle filter framework to track multiple humans using an omnidirectional camera. In order to efficiently converge to the target distribution, a foreground-based importance sampling mechanism using foreground segmentation algorithm is proposed to draw particles from currently observed image. The fusion of color and contour features to evaluate the likelihood measurement makes human tracking more accurate. Likelihood evaluation by integrating two-space enhances the robustness of the system to the warping effect. The overall performance is validated using several videos in the experiments.