| Target tracking is an important branch of artificial intelligence.It is a process of processing sequential images,that is,locating the accurate position of a designated target in successive frames and forming a motion trajectory.At present,the application of target tracking is mainly concentrated in the fields of precise guidance,intelligent monitoring,human-computer interaction and so on.In the field of target tracking,from the traditional algorithm to the deep learning related algorithms,the target's feature expression directly affects the tracking accuracy and success rate.How to obtain a more comprehensive feature representation method is one of the key points of this article.In recent years,target tracking algorithms can be roughly divided into traditional target tracking algorithms and deep learning related target tracking algorithms.Traditional target tracking algorithms generally have real-time speed,but they are only used in specific scenarios and have poor generalization performance.Deep learning related target tracking algorithms use the human visual system's hierarchical processing of external information and combine lower-level features to form more abstract highlevel features,but the real-time performance is somewhat poor.Aiming at these shortcomings,a lightweight end-to-end tracking network based on deep learningrelated target tracking algorithms is presented in this paper,which not only optimizes the feature representation method,but also adds a correlation filter layer to improve the tracking speed,having certain advantages in accuracy,success rate and speed.The work of this article is summarized as follows:Firstly,in order to obtain the temporal and spatial information of the image sequence,a spatiotemporal sampling network is designed to extract the target features,and a deformable convolution layer is added to solve the problem of target deformation and size change.In the feature aggrated stage,an Attention mechanism is added to optimize the features extracted by the spatiotemporal sampling network.When the target feature changes during the tracking process,the features of the former K historical frames are used to enrich and enhance the features of the current frame.The spatiotemporal sampling network can extract more robust target features without complicated network design and a large amount of optical flow data.Secondly,in order to test that the spatio-temporal sampling network can obtain robust target features,an experimental simulation was conducted.The aggrated feature obtained from the spatiotemporal sampling network is used as the feature of the template frame,searching in the feature map of the current frame.The feature matching is carried out by the normalized cross-correlation method.The target is located at the highest score of the similar score map,and the tracking result is obtained.On the OTB2013 data set,experimental test analysis shows that the data of accuracy,success rate and speed are 0.833,0.636 and 34.8 FPS.In the face of deformation and target size changes,it achieves better tracking effect,comparing with some other classic algorithms.Lastly,in order to improve the tracking effect,this paper continues to combine the features extracted by deep learning with correlation filtering.The features extracted by the spatiotemporal sampling network are sent to the correlation filtering layer,achieving the target tracking results.The cyclic matrix in the CF layer can perform fast calculations in the frequency domain of the Fourier space,which improve the tracking speed of the target,makeing the DDCF algorithm have obvious speed advantages.DDCF algorithm obtains 42.3FPS in tracking speed,accuracy and success rate are 0.866,0.666.Finally,the DDCF algorithm is tested on the laboratory data set,and the tracking speed reaches 47.8FPS,indicating that the algorithm has research prospects in engineering applications. |
