影像在部分遮蔽下的一種改良Struck的追蹤設計

當追蹤器在進行物體跟蹤過程時，如果物體遭到部分遮擋就會很容易檢測到不正確的對象位置，而錯誤的檢測會讓追蹤過程出現錯誤。為了解決這個問題，本論文針對部分遮擋提出了一個有效的遮擋檢測來結合線上學習跟蹤檢測的方法。在選定第一幀的目標位置後，將目標對象以中心分成四個矩形塊，分別對目標對象和四個矩形塊訓練分類器，然後利用卡爾曼濾波器來考慮目標對象的運動並預測下一幀的對象位置，最後結合BC 係數相似性分析來進行目標對象的更新，從而改善Struck 跟蹤算法。在遮擋檢測方面，將選定候選對象以中心為原點分成四塊矩形塊，透過計算矩形塊的BC 相似性分析來確認候選對象是否有受到部分遮擋的影響，最後在有受到部分遮擋的矩形塊上將上一幀未受到部分遮擋的矩形塊進行替換，使得被部分遮擋的候選對象位置能夠更新。經過實驗結果得出，本論文提出的方法在解決部分遮擋的問題上確實優於其他跟蹤器。

關鍵字

影像處理；結構化支持向量機；卡爾曼濾波器；遮擋檢測；巴氏係數

並列摘要

When the tracker is tracking the object, the object is occluded. It will be easy to detect the incorrect position of the object, and the wrong detection will make the tracking process wrong. In order to solve this problem, this paper proposes an effective method of occlusion detection combined with online learning tracking detection for partial target occlusion. After selecting the target object in the first frame, and divide the target object into four rectangular blocks in the center. We train the classifier on the target object and the four rectangular blocks respectively, then use the Kalman filter to consider the motion of the target object, and predict the position of the object. Finally, the Bhattacharyya coefficient similarity analysis is used to update the target object to improve the Struck algorithm. In terms of occlusion detection, the selected sample is divided into four rectangular blocks with the center as the origin, and the BC similarity analysis of the rectangular blocks is calculated to confirm whether the object position is affected by the occlusion. Finally, replace the rectangular block that was not occluded in the previous frame on the rectangular block that was occluded, so that the position of the occluded object can be updated. The experimental results show that the method proposed in this paper is indeed superior to other trackers in solving the problem of partial target occlusion.