RESEARCH ON CORRELATION FILTERS OF VISUAL TRACKING ALGORITHMS Information Technology 2017
Li, Chengxi (2018)
Li, Chengxi
Vaasan ammattikorkeakoulu
2018
All rights reserved
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-201802062089
https://urn.fi/URN:NBN:fi:amk-201802062089
Tiivistelmä
This thesis focuses mainly on the visual tracking algorithm, the Kernelized Correlation Filters (KCF) algorithm and Discriminative Scale Space Tracker (DSST) algorithm. They are widely applied in many fields. Moreover, the visual tracking framework with KCF and DSST outperforms in the perspective of tracking speed and accuracy, which has drawn increasing attention.
Although these target tracking algorithms achieve long-term and accurate tracking of the target, there are still many problems in the practical application en-vironment such as stability, adaptability and real-time performance. In view of these problems, some improved methods are proposed. Aiming at the problem that the detection module in the algorithm needs to detect the lack of accuracy of the fast-moving object, a Kalman filter is used to estimate the approximate appearance area of the target in the current frame. This approximate area is taken as the target de-tection area of the algorithm. Although the speed of the algorithm has a certain impact, but the accuracy of the algorithm has a certain degree of improvement
In this thesis, the Kalman filter is proposed to be utilized in the visual tracking framework with KCF, which is more robust to movements of the target area. Fur-thermore, the simulation results with test beds based on Matlab and OpenCV 3.3 show that the proposed framework outperforms the conventional KCF and DSST-based visual tracking framework. Experiments show that the two algorithms have their own advantages in the matching rate, the matching speed and the number of frames successfully tracked. And the improved algorithms are more effective than the original ones.
Although these target tracking algorithms achieve long-term and accurate tracking of the target, there are still many problems in the practical application en-vironment such as stability, adaptability and real-time performance. In view of these problems, some improved methods are proposed. Aiming at the problem that the detection module in the algorithm needs to detect the lack of accuracy of the fast-moving object, a Kalman filter is used to estimate the approximate appearance area of the target in the current frame. This approximate area is taken as the target de-tection area of the algorithm. Although the speed of the algorithm has a certain impact, but the accuracy of the algorithm has a certain degree of improvement
In this thesis, the Kalman filter is proposed to be utilized in the visual tracking framework with KCF, which is more robust to movements of the target area. Fur-thermore, the simulation results with test beds based on Matlab and OpenCV 3.3 show that the proposed framework outperforms the conventional KCF and DSST-based visual tracking framework. Experiments show that the two algorithms have their own advantages in the matching rate, the matching speed and the number of frames successfully tracked. And the improved algorithms are more effective than the original ones.