Target Recognition and Tracking Based on Scale Adaptive Kernel Correlation Filter for Dynamic Perspective of Robots

Abstract

An autonomous target recognition and tracking method based on improved Kernel Correlation Filter (KCF) is proposed to deal with camera jitter and changing environments in dynamic perspective for mobile robots. Firstly, the Speeded Up Robust Features detection algorithm is combined with Random Sample Consensus to recognize targets and mark target areas. And then, KCF is used for rough tracking of the above target areas. Meanwhile, the target re-identification strategy is applied to update the tracking data in real time to solve the problem that the scale of KCF tracking box is too inflexible for accurate tracking. Finally, tests are carried out not only indoors but also outdoors. The experimental results show that the proposed method can obviously enhance the stability and accuracy of dynamic perspective. Its accuracy is higher than SIFT and KAZE when the camera moves rapidly, and it has better real-time performance and tracking accuracy than KCF and MIL.