視覺追蹤在目前電腦視覺的領域中為一重要的角色，其能應用於機器人上，幫助機器人達成許多任務。本論文以現有的壓縮追蹤演算法為基礎，提出一個多尺度分類之視覺追蹤方法，其依靠多個特徵分類器，在不同尺度空間條件下進行分類器訓練，使得追蹤效果更加強建。再者，在分類器初始化時，可透過隨機生成不同種類特徵的方式，產生較為強健之特徵分類器，獲得較精準的追蹤結果。為了提升追蹤系統的即時效能，在本論文中也針對所提出的視覺追蹤演算法進行各部份動作的執行時間分析，並針對執行時間較高的部分，進行CPU的平行加速。另外，也針對分類器公式，進行數值討論與公式化簡，以解決增加多尺度分類器後，演算法運算複雜度增加的影響。經由上述平行加速及運算簡化後，所實現的演算法在處理642x352影像時，可以達到每秒約45張影像(45 fps)的追蹤處理速度。而在實驗結果上，與四種現有的視覺追蹤演算法相比，不但在追蹤成功率有明顯提升外，在追蹤精確度上也較為優異。 Visual tracking acts an important role in fields of computer vision as it can apply on robots to accomplish many tasks. In this thesis, a robust visual tracking algorithm is proposed based on the existing compressive tracking method. The proposed algorithm adopts multiple Bayes feature classifiers, each of them was trained under a different scale condition, to realize online multi-scale classification, which can greatly improve the robustness of tracking system. Furthermore, each feature classifier was initialized by randomly generating different types of Haar-like features. By doing so, the robustness of feature classification can be improved to obtain more accurate tracking results. In order to enhance real-time performance of visual tracking system, the processing time of each stage of the proposed algorithm was recorded. Then, the processing speed of the proposed algorithm is accelerated by adopting CPU-based parallel processing on the computationally expensive stages. Also, the formula of the Bayes feature classifier is simplified by numerical analysis to speedup the processing speed of multi-scale feature classification. After accelerating by parallel computation and formula simplification, the proposed visual tracking algorithm reaches about 45 images per second (45 fps) tracking performance in dealing with images of 642x352 pixels. Experimental results show that the proposed algorithm outperforms four state-of-the-art visual tracking methods in terms of success rate, tracking accuracy, and visual comparison.
