A KCF-Based Incremental Target Tracking Method With Constant Update Speed

With a good balance between tracking accuracy and speed, kernelized correlation filter (KCF)-based trackers have now become dominant approaches in the field of online object tracking. However, a KCF-based tracker updates templates by linearly combining historical objects, and its performance inevitably degrades when the accumulated error causes the model to look more similar to the background than the foreground. To address this model drifting problem, we propose an incremental kernel principal component analysis (IKPCA)-based KCF (IKPCA-KCF) tracking scheme. In IKPCA-KCF, IKPCA technology is used to incrementally learn the target features to alleviate the model drifting problem. Moreover, reduced-set (RS) expansions are adopted to compress the historical target samples to maintain a constant template update rate. Additionally, KCF-based trackers use the response map generated by the correlation filter to determine the target location. However, when the target leaves the field of view, using the maximum response value to reflect the target’s position information will certainly induce tracking errors. To further solve the tracking stability problem, a simple and effective nearest neighbor classifier is used to confirm the tracking target during the tracking process. Extensive experimental results demonstrate that the proposed IKPCA-KCF tracker is competitive with state-of-the-art trackers on eight challenging video datasets.