物體偵測與辨識是許多電腦視覺應用中重要的處理工作之一。當影像中存在著多個不同物體的情況下，要使電腦選擇某一特定物體的確切位置並持續追蹤更是一個值得探討的議題。本文提出了一種多物體辨識及目標物追蹤方法，有效地解決這個議題。所提出的演算法首先會由一張輸入影像與一張目標物參考影像，透過特徵點描述子提取及匹配演算法求得特徵點描述子以及特徵匹配點。接著，於參考影像上訂定一個控制點，再使用候選點運算方法將參考影像與輸入影像之間的特徵匹配點計算出候選點在輸入影像中的位置，其會分佈於設定的目標物控制點周圍。此時可使用mean shift演算法將先前算出的控制點進行分群的運算，藉此找到多個候選點群集的中心，其不同的群集中心即代表不同物體的中心點。最後，利用template matching演算法 對參考影像的中心區塊進行視覺追蹤模型訓練，並使用其所訓練出來的模型對物體的中心區塊進行追蹤。在實驗測試中，本文所提出之演算法不僅能成功辨識出多種不同於箱內隨機擺放之堆疊物體，並且在運算速度或辨識準確度上都能達到不錯的效能。當使用GPU平行運算技術加速時，本文所提出的演算法能達到了每秒25張640x480影像的運算速度。 Object recognition and detection play important roles in various computer vision applications. When the image contains different types of objects, detecting and tracking an object-of-interest (OOI) from the multiple objects become a difficult task and a topic worth exploring. In this thesis, a novel multiple object clustering and target tracking algorithm is proposed to address this issue efficiently. The proposed method first employs a keypoint extraction and matching algorithm to extract keypoint descriptor matches between an input image and a target reference image. Next, setting a control point on the target reference image, a candidate point computation method is used to compute the position of each candidate point around the target control point in the input image associated with each keypoint match obtained from the previous stage. Then, the center point of each detected OOI in the image can be seqarated by applying a mean shift clustering approach to classify the computed candidate points into different clusters, each of them indicating an OOI to be tracked. Finally, a template-based visual tracking method is adopted to locate and track the center position of the top OOI detected in the image based on a template matching model trained from the target reference image. Experimental results show that the proposed method not only successfully recognizes each OOI from multiple stacking objects randomly placed in a box, but also achieves high recognition accuracy with real-time performance. When using GPU parallel computing technology to accelerate the proposed method, the entire system reaches about 25 frames per second in processing images with size 640×480 pixels.
