利用景深資訊降低背景干擾以減緩飄移問題

Abstract

使用適應性物體模型是一種的物體追蹤方法。這種追蹤法具有演算法簡單與執行快速的優點，但也容易因為背景的干擾問題，出現飄移（Drifting）問題，而影響追蹤結果的正確性。飄移問題的發生主因來自於1) 物體的適應能力，以及2) 背景的干擾。在這篇論文中，我們以Online Boosting for Tracking (OBT) 演算法為基礎，引入了景深、多尺度的追蹤器和動態更新的追蹤器生命值等資訊，設計了一套新的物體追蹤演算法，稱為Enhanced OBT（簡稱EOBT）。在EOBT 中，景深資訊可用來濾除背景、多尺度的追蹤器可以改善追蹤的準確度，而動態更新的追蹤器生命值則可用以判斷物體是否被短暫的遮蔽，進而降低追蹤器因物體被短暫遮蔽所造成的準確度影響。此外，由於現有的準確度評估方法無法完全反映出追錯目標物體的問題，在本論文中，我們另外提出了新的評估方法，設計新的比率（Ratio in Object 及Ratio in Tracker）來評估追蹤的準確度。其中，Ratio In Object 反映了有多少比率的物體被成功地追蹤到；而Ratio In Tracker 則反映出待追物體落在追蹤器內的面積比率。我們也設計了不同的實驗，證明本論文所提出之EOBT 演算法能成功地減緩飄移問題，並提高物體追蹤的準確度。

Recently, tracking using adaptive appearance models is popular. Tracking algorithms adopting an adaptive appearance model are simple and fast, but suffer from drifting problems caused by background interference. The drifting problem, resulting in inaccuracy, comes from the accumulation of slight labeling errors occur in updating model in each tracking iteration. Taking online boosting for tracking (OBT) as the basis, we introduce depth, multiple scales and lifetimer to our algorithm (named Enhanced OBT; also abbreviate to EOBT) and eliminate drifting problems induced by background interference. In EOBT, depth can be used to filter out the background data, the racker with multiple scales can be used to improve the accuracy, and dynamically adjusted lifetimer can be used to determine whether the object is temporarily occluded. Since conventional evaluation method of accuracy may derive a high accuracy when an algorithm tracks a wrong target, we additionally design two ratios (`Ratio in Object' and `Ratio in Tracker') to avoid such a problem and precisely evaluate the accuracy. In our method, `Ratio in Object' shows the percentage of an object caught by a tracker, while the `Ratio in Tracker' reflects the percentage of a tracker occupied by the object to be tracked. In this thesis, we conduct several experiments to show that EOBT can effectively reduce drifting problems and improve the accuracy of object tracking.