Multi-Feature Fusion Correlation Filtering Object Tracking Algorithm Based On Scale Adaptation

Object tracking has always been a key issue in the field of computer vision and plays an important role in video retrieval,driverless,intelligent monitoring systems,industrial robots and other fields.In recent years,with the continuous improvement of computer hardware technology and the rapid development of artificial intelligence,the problem of moving object tracking has received more and more attention.Although there are many algorithms for object tracking,the problem of achieving accurate tracking under complex conditions has not been well solved.After the Kernel Correlation Filtering(KCF)is proposed in 2014,it attracts wide attention of researchers.It transforms the calculation from the time domain into the frequency domain by fourier transform,which greatly simplifies the calculation,not only improves the tracking speed,but also greatly improves the tracking accuracy.This article addresses the problem of object tracking under complex conditions.On the premise of ensuring the real-time performance of the algorithm,the features,scales and model updating mechanism are improved on the basis of KCF.A scale-adaptive multi-feature fusion object tracking algorithm is proposed,which significantly improves the tracking accuracy of the object tracking algorithm under complex conditions(scale variation,object occlusion and image blur).The main innovations of the algorithm in this paper are as follows:(1)For the problem of low image quality in object tracking process,this paper proposes an multi-feature fusion algorithm based on kernel correlation filtering algorithm.Combine each feature for its advantages in different environments.Due to the various interferences that occur in the object tracking,we combine the histogram of gradient(HOG),the color feature(CN),gray feature(Gray)and the convolutional neural network feature(CNN).The weighted result is used as the feature used in the final tracking.At the same time,the model space clipping strategy is added in the process of feature extraction,and the number of samples is improved under the premise of ensuring the sample quality.Through experimental analysis,we get:Compared with the classical KCF algorithm,the multi-feature fusion algorithm we use has higher tracking accuracy,but it also has defects.Next,we have improved the object scale and model update for the existing defects.(2)Aiming at the problem of object scale change in object tracking process,this paper proposes a classification tree-scale adaptive algorithm based on kernel correlation filtering.The tree size search method is used to quickly locate the size of the object scale to find the best response position.Compared to traditional algorithms,a more accurate object scale can be obtained with the same number of calculations.(3)Aiming at the occlusion and disappearance problems in the object tracking process,this paper proposes an adaptive model updating strategy algorithm based on kernel correlation filtering.The accuracy of the model is affected by the error frame updating during the tracking process.This paper firstly determines whether the current frame has reached the updated standard through multi-peak detection,and does not update the pictures that do not meet the requirements;The peak comparison strategy method are then used to determine the size of the learning rate in the model update.Through experimental analysis,the algorithm of this paper can still achieve accurate tracking under the conditions of scale change,object occlusion and image blur.In the data set OTB-2013,the tracking accuracy of the algorithm reached 87.4%,and the success rate reached 67.1%.Compared with the MOSSE algorithm,the tracking accuracy is improved by 26.4%,and the success rate is increased by 21.4%.Compared with the KCF algorithm,the accuracy is improved by 9.0%,and the success rate is increased by 12.0%,which has a very large increase.