Experimental Research Of Target Tracking Algorithm Based On DJ Manifold

Target tracking technology has a very important position in the field of computer vision,and it has a wide range of applications in many fields,such as video surveillance,autonomous navigation,human-computer interaction,military defense,and so on.In recent years,the target tracking technology has made great progress,but it is also faced with many problems,such as morphological changes,scale changes,occlusion and disappearance,blurred images,etc,these problems may lead to tracking failures.In addition,the experimental verification of target tracking research literature is mostly running data sets,which lacks practical verification.In response to the above problems,this paper improves the KCF target tracking algorithm,and transplants it to the DJI Manifold 1 microcomputer for tracking experiments;Considering the excellent performance of the deep learning target tracking algorithm,configure environment and transplant the Siamese RPN target tracking algorithm to the more outstanding DJI Manifold 2-G microcomputer,and conducts tracking experiments.First,considering that the traditional KCF target tracking algorithm uses a single feature,do not monitor the output,and the size of the target box is unchangeable.It can easily cause tracking failures when the target under abnormal conditions such as occluded or scaled.In response to the problems,this paper proposes an improved Kernel Correlation Filter target tracking algorithms.Firstly,the cascade of CN feature and FHOG features is applied to train and update filters.Secondly,the tracking output is monitored by the PSR value.When an abnormal tracking output is detected,the self-correction mechanism is activated to accurately re-track the target.Finally,a scale filter is added to adapt to changes in the target scale.Simulation experiments of this algorithm were performed on the OTB-2015 data set,and physical experiments were conducted on Manifold 1.The experimental results verified the effectiveness of the algorithm in dealing with abnormal conditions such as scale changes and occlusion.Second,in the VOT2015,VOT2016 and VOT2017 data sets,the Siamese RPN target tracking algorithm has achieved leading results in performance,and also reached a running speed of 160 FPS.In this context,based on the DJI Manifold 2-G equipment in the laboratory,this paper builds an operating environment for the Siamese RPN target tracking algorithm,and writes relevant code to transplant the Siamese RPN target tracking algorithm to DJI Manifold 2-G for tracking.The performance analysis is performed and the experiment is compared with the traditional KCF target tracking algorithm and the improved algorithm proposed in Chapter 3 of this paper,it shows the advantages of Siamese RPN target tracking algorithm when the target scale changes,the target is occluded,and the target is out of view.