Anti Occlusion Improvement of Kernel Correlation Filter Target Tracking Algorithm

This paper proposes a target tracking algorithm that combines kernelized correlation filter (KCF) and Kalman filter (KF) prediction to address the problem of poor tracking performance or even failure in target tracking when the target is occluded. This article uses peak sidelobe ratio to determine whether the target is occluded. When the target is unobstructed or partially occluded, the learning rate is optimized to update the target appearance model; When the target is severely occluded, stop updating the KCF model and use the Kalman filtering algorithm to predict the trajectory of the moving target to estimate its position at this time. This article uses the OTB100 dataset for experiments, and the results show that the improved KCF (KCF-A) target tracking algorithm has improved accuracy and success rate. Compared with other target tracking algorithms, its tracking accuracy and success rate are better, and it can achieve better tracking performance when the target is occluded, effectively improving the algorithm's anti occlusion ability.

I. Introduction

Target tracking is an important research direction in the field of computer vision. Motion target tracking algorithms can be divided into generative methods and discriminative methods according to their working principles, and correlation filtering algorithms are a discriminative tracking method [1]. It generates training samples through cyclic shifting and calculates them in the Fourier domain, greatly improving algorithm efficiency and receiving extensive research. Bolme et al. [2] proposed the minimum output sum of squared error (MOSSE) algorithm and introduced correlation filtering for image tracking for the first time. Henriques et al. [3] proposed a cyclic structure of tracking by detection with kernel (CSK) tracker, and subsequently improved the CSK tracker by proposing a multi-channel feature kernel correlation filter (KCF) [4]. Based on the problem of target scale changes, Danelljan et al. [5] proposed a fast discriminative scale space tracker (fDSST) algorithm, which introduces a scale pyramid based correlation filter to detect scale changes and combines correlation filtering with scale filtering.