Research On Online Tracking Algorithm Of Single Target In Video

Visual object tracking plays an important role in the computer vision community,and it is crucial to application in numerous fields such as autonomous driving,unmanned aerial vehicle,pedestrian tracking,medical and other aspects.The target tracking algorithm is to determine the location and size of the target in the subsequent video sequence under the premise that only the first frame tracking target is given.However,due to the complex application scene in the process of target tracking,such as interference,background clutter,fast motion,scale transformation,occlusion and so on,which makes the algorithm research full of difficulties and challenges.In this paper,we analyzed the classic correlation filters tracking algorithm and the deep learning matching framework,we conducted the following research is carried out for single target online tracking:1.The traditional Siamese networks could not update the reference model in real time,so it is easy to lose the target when it moves or deforms rapidly.To overcome this shortcoming,we proposed the multi-classification dynamic Siamese network,called MCDSiam.The innovation is embodied in three aspects:1)We use the least mean square error method to get the correlation between the reference template and the current tracking target,and save the reference template after correlation filtering.Then,we update the dynamic classification according to the Euclidean distance between the target features in different frames.;2)We design a fast and independent scale filters to find the optimal scale estimation of the target separately;3)We propose a dynamic gaussian window to drive out the background clutter in the search area.The performance of MCDSiam is validated on OTB-2015 benchmarks,and the experimental results demonstrate that MCDSiam has the state-of-the-art tracking accuracy.2.At present,the most traditional discriminant correlation filters(DCF)were employed a rectangular bounding box to track visual objects.This way is simple and convenient.But the disadvantages are particularly obvious that will regard some background information as the target.In order to solve the above problems,we propose a novel correlation tracking via mask and multi peaks re prediction.Firstly,we extract the edge information in the target area by segmentation algorithm of MCG(multiscale combinatorial grouping),and use padding function which designed by ourselves to make the target mask.In this way,we can obtain a more precise target area.Then,we divide the target region into super-pixel according to its feature similarity by the SLIC method.And calculate the centroid deviation of each super-pixel between two adjacent frames to change aspect ratio which represent the target state.Finally,we use multiple peak points in the filtered response map as the origin to generates candidate areas when the target is disturbed.And we match these areas with the first frame target to select the most suitable target at present.We conduct extensive experiments on the public datasets to evaluate the proposed tracker.Experimental results show that our tracker performs favorably against other state-of-the-art trackers.