Online sequential-extreme learning machine based detector on training-learning-detection framework

Tracking unknown objects using adaptive tracking-by-detection approaches are widely used in computer vision. In these approaches, tracking problem is treated as an online classification problem, where the object classifier model is updated in the current frame to be used for classification process in the next frame. One of the approaches is based on Tracking-Learning-Detection (TLD) framework, where a combination of generative tracker and discriminative detector models are used for tracking process. In TLD, a cascaded detector with three stages is used to localize the tracked object in the current frame and correct the tracker if necessary. The second stage of the cascaded classifier uses a randomized ensembles classifier. This classifier relies on simple pixel comparison that may not offer a good feature value to well generalize for the new appearances of the object in the detection process. In this paper, we introduce online sequential — extreme learning machine (OS-ELM) to obtain a more generalized detector in TLD framework. In our approach, the ensemble classifier in TLD framework is replaced with OS — ELM based on pixel feature values. OS-ELM provides good generalization performance with the ability to detect and track a target object in sequential frames besides the ability to learn online data arriving sequentially with different sizes. A series of experiments were conducted to evaluate the performance of the proposed approach to compare with the existing TLD with ensemble classifier. The results show that the proposed approach is able to offer better performance in six videos out of nine videos with better generalization.