| Visual object tracking is a basic and important research topic in the field of computer vision,which enjoys a wide range of applications such as intelligent video surveillance,video-based human-computer interaction,robot vision,and unmanned aerial vehicle.General visual single object tracking is the task of estimating the state information(position information,scale information)of any object in the video sequences.A specific object is selected by using a bounding box in the first frame of a video or image sequence,a given object tracking algorithm is used to determine the position of the object by using the bounding box in subsequent frames.Although many studies have been conducted by national and international researchers on visual object tracking,with great progress having been made recently,visual object tracking is still considered to be a challenging problem due to many factors such as object size and scale variation,illumination,background clutter,fast movement,and motion blur.In addition,the real-time requirements have also become a major bottleneck for the actual application of tracking algorithms.Based on the analysis of existing tracking algorithms,this paper takes the Siamese network tracking algorithm as a basic framework to carry out in-depth research.The specific work of this paper is as follows:(1)A tracking algorithm of siamese network based on attention mechanism has been proposed.Firstly,the cropped template frame and detection frame are sent to the Res Net50 feature extraction network,which adds feature fusion mechanism for shallow,middle,and deep feature extraction.Secondly,in the feature adjustment module,the features after deep cross-correlation are adjusted by series channel attention and space attention,so as to improve the sensitivity of the network to the tracking object.Finally,the adjusted features are sent to the classification and regression module for the binary discrimination of tracking object and the regression of boundary box,so as to obtain an accurate object boundary box.The results of experiments on OTB100 and UAV123 datasets show that the proposed algorithm can achieve accurate object tracking.(2)A tracking algorithm of the siamese network based on Mobile Nets has been proposed.Firstly,a regression branch loss function based on IOU loss is designed in the offline training phase,by adding a scale sensing factor to the loss function,the sensitivity of the tracking algorithm to the tracking object scale is improved;Secondly,in the online tracking stage,the improved Mobile Nets is used as the feature extraction network to complete the object feature extraction,which reduces the parameters of the tracking algorithm and improves the efficiency of the algorithm.Finally,a similarity calculation module is designed to complete the deep cross-correlation between template frame and detection frame features by means of dilation convolution,which improves the perception of different scale objects.The results of experiments on OTB100,UAV123,and La SOT validate the effectiveness of the proposed algorithm. |
