A Long-term Single Target Tracking Algorithm Based On Improved Correlation Filtering

The existing single target long-term tracking algorithms cannot effectively deal with the problems of target occlusion and scale variation.To solve the abovementioned problems,a single target long-term tracking algorithm based on improved correlation filtering is proposed.The main works of this research are listed below.(1)A long-term tracking algorithm with scale self-adaptiveThe KCF algorithm does not change its tracking box,cannot find back the lost target,and updates the template in each frame.Aiming at these problems,a long-term tracking algorithm with scale self-adaptive is proposed.By adding a scale pool to KCF,it can effectively respond to scale change of the target during the tracking process.By exploring the setting rules of the scale pool,a more universal setting method can be found.A target re-detection mechanism is set up.When a frame with poor tracking condition occurs,the position of the target is found out again by expanding the search range.This makes the proposed algorithm can deal with the challenges of fast motion and occlusion easily.By updating the tracking template conditionally,the quality of the tracking template is improved,and the effects caused by negative factors such as fast motion,background clutter,and occlusion are alleviated.The experimental results show that the success rate and accuracy of the proposed algorithm on OTB2013 are improved by 18.1% and 15.3% respectively compared with the KCF algorithm.Compared with other multi-scale algorithms,the success rate and accuracy of the proposed algorithm are also improved to some extent.(2)A tracking algorithm based on channel selection and target re-detectionIn order to break through the limitation of traditional features in improving the accuracy of target tracking,the filter is trained using selected convolutional features.The Hierarchical Convolutional Features(HCF)algorithm is time-consuming in extraction of convolutional features,has no dimensionality reduction of the convolution channel,is difficult to recover missed targets,and updates each tracking template unconditionally.Aiming at the abovementioned problems,a channel-based selection and targets re-detection tracking algorithm is proposed.By using the lightweight neural network,MobileNet,to obtain convolution features,the amount of calculation is reduced,and the tracking speed is improved.The channel of the convolution layer is selected to obtain more effective convolution features to train the filter.This improves the tracking accuracy of the algorithm.According to the characteristics of fusion of multiple response graphs,a target re-detection mechanism for multiple filters is designed,and the optimal template is used to detect the target position again,which effectively reduces the occurrence of target missing.Updating the tracking template conditionally can also improve the tracking accuracy.It can be seen from the experiments that the proposed algorithm improves the success rate and accuracy by 9.2% and 1.5%,respectively,and performs more stably during tracking of longer video sequences.