Object Tracking Based on Compressive Features and Extreme Learning Machine

Abstract

Effective object tracking is a great challenging task in computer vision due to the appearance changes of an object. Extreme learning machine (ELM) can analytically determine the weights of feed-forward neural networks, while compressive features can be extracted via a random sparse matrix. Thus, their common advantages are running fast. Moreover, the strong learning ability of ELM is helpful to make a tracker robust target appearance change. In this paper, we propose an effective tracking method using ELM and compressive features. First, a bank of compressive Haar-like features is extracted with the sparse random matrix. To make the feature invariant to illumination, the mean illumination of each image patch is removed from each Haar-like feature. Second, a bank of ELM-based weak classifiers are combined together to form a strong classifier to complete tracking task. Considering the discrimination performance of features is uneven, the training accuracy of each ELM is employed as the confidence (weight) of the weak classifier. Thus, less discriminative features will be weakened. The proposed tracker has been tested on several representative sequences, and the results show that it achieves the highest tracking success rate and the tracking accuracy among the compared trackers including the FCT, CNT, MIL, KCF, and ESPT.