Adaptive Multi-Strategy Observation of Kernelized Correlation Filter for Visual Object Tracking

Abstract

Visual object tracking leads a vital role in multiple fields such as intelligent surveillance system, intelligent transportation system, human-computer interaction, behavior analysis, and intelligent driving assistance. In recent years, research of object tracking tends to focus on improving accuracy. Kernelized Correlation Filter (KCF) is considered as a baseline algorithm for real-time object tracking in term of high computation speed and accuracy by using correlation efficiently in the Frequency domain. However, correlation filter-based tracker is still prone to model drift due to incorrect predictions. This condition caused by varied appearance model especially in fast motion and motion blur. We proposed a new concept of KCF based tracker by adding confidence score scheme to detect tracker loss. Our tracker also introduces observation model with adaptive multi-strategy to find the lost target. We test the proposed method using OTB100 data that has strong characteristics in fast motion and motion blur. The result demonstrates that the proposed method was capable of recovering the lost target. The proposed tracker achieves better performance compared to the existing tracker in term of 0.887 in accuracy and 0.895 success rate.