近十年來，壓縮感知是信號處理領域中的重大發現之一，它能夠以極少量的信號重建出原本的信號，達到即時計算的目的。2012年 “Real-Time Compressive Tracking”成功的應用壓縮感知於視覺追蹤上，由於方法簡明計算迅速，受到許多學者的關注。他們的方法利用滿足壓縮感知理論RIP條件的隨機稀疏矩陣，對多尺度的影像特徵進行降維，接著以簡易貝氏分類器進行分類達到追蹤的目的。然而，若在追蹤的過程中持續更新模型，模型很容易於發生偏移(drift)的情況下更新錯誤的資訊導致追蹤失敗。因此，本文提出兩點改良，第一，採用動態更新模型的方式，以分類器得到的分數來決定是否模型更新而確保模型的精確性；第二，基於被追蹤的目標大多落於追蹤視窗的正中間，因此特徵與視窗中心的距離以及常被挑到特徵，都分別賦予較高的權重，再以兩者的相關性 (correlation) 選擇出最佳的一組特徵。經過大量實驗後，第一點改良確實降低drift問題的機率而增加模型的精確性，然而第二點改良沒有明顯的改進，可能是因為限制特徵挑選的方式與壓縮感知的方法產生矛盾所導致，或者是因為強調視窗中心而忽略背景所蘊含的資訊。另外，我們發現隨機選取特徵和隨機選取更新的訓練樣本導致較差的穩定性，而不同的影片所適合的特徵也有差異性，也就是特徵不具有推廣性。 The past ten years, compressive sensing is an important discover on the topic of signal processing, it can use a few signals to represent the source signals for the purpose of the real-time computation. In 2012, Zhang et al. in “Real-Time Compressive Tracking” [2] successfully applied it on visual tracking and attracted attention of researchers in computer vision field due to the simple algorithm and low computation. They use a sparse random matrix that satisfies the RIP condition of compressive sensing to reduce the dimension of the multi- scales Haar-like feature of images and build positive and negative Gaussian models for each selected feature. Then, target is localized by the naïve Bayes classifier. However, models are apt to update wrong information when the drift problem happens and the system remains updating models in the tracking. Thus, we propose an improvement to ensure the accuracy of models, which adaptively updates models depending on the score from the classifier. After a large number of experiments, we demonstrate that our improvement increases the accuracy of models by reducing the drift problem. In addition, we find out that the tracking results may be quite different due to randomness in selected features and best feature combination may not be applicable for different tracking scenarios.
