On-Line learning-based Fuzzy Clustering Modelling in Real-Time Object Tracking

Abstract

Real-time object tracking has been one of the challenging issues in autonomous navigated robotic design. Most of the solutions proposed the off-line calculation or powerful computation unit on-board, where collection of data samples is provided in prior. While a complete data collection is hard to achieve practically, this paper proposed an on-line learning-based fuzzy adaptive resonance theory (Fuzzy ART) with Takagi-Sugeno-Kang Fuzzy model aimed to cluster and predict object positions from live-stream video frames. In this study, an object is extracted from a video frame, and object' centroid position is fed for fast-learning. Fast-learning mode of Fuzzy ART is used for real-time clustering the positions. While the position obtained are considered fuzzy, the clusters formed unique patterns at each cluster and represent them as fuzzy sets. The fuzzy sets are used to predict the object position using the Takagi-Sugeno-Kang Fuzzy Inference System (TSK-FIS) model. The proposed algorithm contributes to an on-line and unsupervised fuzzy modelling for object tracking to predict the object's position in real-time. In order to validate the practicality of the proposed approach, two experiments are conducted, i.e., using simulated data samples, and a case study using real-time hand movement tracking. A bounding box for region of interest on a frame is moved according to the predicted positions, while the robustness of the tracking is measured through the box's movements. The proposed model is self-improving over time, which has great advantage for selfcorrection in autonomous applications.