本篇論文提出利用線性判別分析 (LDA, Linear discriminant analysis) 以及 part-based 的追蹤策略於複雜環境中進行單目標視覺物件追蹤。使用相同預先一次性訓練得到的負例樣本平均值與共同變異量，再加上目標物不同位置的正例樣本即可快速訓練出不同的部件 (part) 分類器。利用快速傅立葉轉換卷積計算追蹤畫格影像特徵的相似度，再使用投票方式結合不同影像特徵的部件與整體的目標物位置，來完成正確及快速的視覺物件追蹤。在100個影片中使用三種不同強健度評估方式來測試追蹤器之精確度與成功率，由量化與質化的實驗結果顯示我們所提出的方法在正確性與速度上都有很好的表現，並能與其他先進的追蹤演算法匹敵。 Tracking-by-detection methods treat the target location as a classification problem in which the approach SVM + HOG shows a good performance. However, training a good SVM classifier is cost expensive. In this paper, we replace SVM by linear discriminant analysis (LDA) for classification where the mean and covariance of negative examples are evaluated only once. Not only the training is much cheaper, but testing time is also very efficient. The proposed method uses HOG and color features for image representation. To defense partial occlusion issue, part-based tracking strategy is adopted and the model is updated according to Peak to Sidelobe Ratio (PSR) of parts. And conclude the classification results from parts and holistic detections by voting. To speed up classification of LDA classifiers with features from search window, the FFT convolution is employed to reduce the computational efforts of dot product on feature vectors. We evaluate our approach on 100 public benchmark challenging video sequences, both qualitative and quantitative experiments show that our approach is competitive to state-of-the-art methods.
