Research And Implementation Of Similar Target Tracking Algorithm Based On Siamese Network

With the development of 5G,big data,AI and several technologies,a large amount of video data is generated and stored in various fields,such as national defense,remote sensing,smart city and security.Thus,the requirements of analyzing video data are remarkable increasing.Visual single object tracking,which aims at tracking the specific target of videos continuously and accurately,is an important research of video data analyzation,widely applying in military application,intelligent monitoring,human-computer interaction and augmented reality.These single object tracking algorithms based on deep neural network can effectively handle light changes,motion blur and complex backgrounds problems.However,the interference problem among similar targets is still not robustly solved by current trackers.To solve this problem,this thesis puts emphases on single object tracking algorithm based on Siamese network,and studying the excitation ability of feature channels in convolutional neural network from different objects.We propose a tracking algorithm,which is based on squeeze and excitation mechanism.Considering that the degree of reducing similarity is not obvious,the tracking algorithm based on trajectory fitting is proposed.Innovations and main work of this thesis are as follows.1.On account of the tracking failures occurred when feature distribution of multiple areas in video image is similar to feature distribution of template image.A single object tracking algorithm is proposed based on squeeze and excitation mechanism(SE-Siam FC).High speed and excellent precision are advantages of visual tracking algorithms based on Siamese network,and these algorithms have got a wide application.However,the backbone network aims at learning general representation of target appearance on large scale dataset.Therefore,the tiny difference of similar targets cannot be represented better.In order to solve this problem,squeeze and excitation mechanism is fused to Siamese network model,inspired by that different weights of feature channels can be obtained.Firstly,the features of template image and video image are extracted by fully convolutional neural network.Then the extracted features are recalibrated by squeeze and excitation mechanism.However,because of squeeze operation,abundant spatial information is lost.The final tracking hot map is composed of weighted fusion feature before recalibration and after recalibration.The result of tracking is obtained by the location of maximum response mapping to the corresponding location in video image.The real target and interfering targets are successfully distinguished by reducing the similarity of feature distribution.2.When the similar targets and real target belong to a same category,the difference of feature distribution between them is not obvious.Inspired by the continuity of moving objects' trajectory,we propose a trajectory fitting strategy based on least square method and a method of template update.Firstly,predicting k candidate locations according to SESiam FC.Then,the locations of previous tracking results map to 2D Cartesian coordinate system,which uses least squares to fit target motion trajectory.Finally,the degree of deviation of candidate locations and trajectory is calculated to obtain the location of target in video image.Meanwhile,considering tracking failures caused by the dramatic changes of targets' appearance and the poor-quality template image,a deviation awareness mechanism of template image update is designed.The tracker achieves state-of-art performance,and operates at frame-rates beyond real-time.3.According to the visual tracking project requirement of a research institute,a software is designed and implemented based on the tracking algorithm proposed in this thesis,called aerial similar target video processing system.There are three modules named video image preprocessing,feature analyzation and tracker.The first one labels video test data and generates annotation files that can calculate the precision and success rate.The second one takes advantage of feature visualization to observe the feature distribution of similar target regions.The last one labels and stores the tracking results on the basis of similar object tracking algorithm.In order to verify the effectiveness of the proposed tracking algorithm,we evaluate the tracker on benchmark datasets: OTB50,OTB100 and similar object dataset obtained by filtering.The experimental results show that the precision and success rate of the proposed tracker are improved.On similar target test dataset,precision and success rate are separately 86.1% and 78.1%.On OTB50 and OTB100 datasets,precision and success rate are separately 82.6% and 62.3%.Meanwhile,supported by the target tracking project of a research institute,the tracking software implemented in this thesis is tested.The functions of video preprocessing and feature analyzation work normally.What's more,the precision and success rate satisfy the requirements of project indicators.Until now,the software has been delivered to the research institute and used in the integrated project of infrared aerial targets.