Weak label for fast online visual tracking

• We propose to assign labels for examples with a strategy we call weak label.
• We use kernel matrix as the similarity function to approximate the weak label.
• We derive an approximation of model parameter to improve speed significantly.

Tracking by detection is becoming more and more popular in recent years. By training a classifier based on the image patches generated around the target, visual tracking can be formed as a classification problem. Most tracking by detection methods assign a constant label to the sample according to the distance between the sample and the center of the target. However, we find that the criterion to assign the label of each sample can be relaxed as long as some properties are preserved. By using the relaxed criterion which we call weak label, a visual tracking algorithm based on least square support vector machine (LS-SVM) and circulant matrix is proposed in this paper. Unlike the previous algorithm using LS-SVM and circulant matrix, the exploiting of weak label can make the coefficients of the support vectors a constant matrix. The calculation of the coefficients in each frame can be reduced, thus the proposed tracking algorithm can run very fast. On the test video sequences, the proposed tracker implemented in MATLAB runs over 300 frames per second on average on a machine with a 3.0 GHz CPU. Experimental results demonstrated the efficiency and accuracy of the proposed method.