Research On Single Object Tracking Algorithm In Video Sequence

In recent years,a lot of cities are striving for “National Civilized City”.The first task is to ensure the safety of the city.Therefore,building a “Safe City” has become a key demand for realizing “National Civilized City”.Nowadays,all kinds of cameras can be seen in the streets,government enterprises and some public places.The intelligent processing of video can “liberate” humans from complicated operation.It automatically processes the video information captured by the camera,and responds in a timely manner.Therefore,the demand of video intelligent processing system is more and more significant.However,object tracking technology,as an important part of the system,has become a hotspot and difficult point in designing a real-time and accurate tracker.This paper studies and evaluates the effect of tracking through three modules: feature representation,update model and integration strategy.Then,we propose two algorithms: an adaptive spatio-temporal context learning for visual tracking and an object tracking algorithm based on siamese network with fused score map.With introducing the correlation filtering,the tracking performance has been greatly improved.However,many traditional algorithms only considered the information of the tracking object,ignored the relevance between the background information around the tracking object and object information.There are researchers proposed a spatial-temporal context model,but it adopted a single feature to represent the target.The learning rate of the model was constant and cannot adapt well to the appearance change of the object.Aiming at the above lacking,multi-feature fusion strategy is proposed.The histogram of gradient and color naming features are used to describe the target with the gray value,and the learning rate is updated according to the difference of the object position in two consecutive frames so that the model can adapt to the changes of object.Analysis by experimental results,it can be seen that the tracking performance of the proposed algorithm,compared with the traditional algorithm,is greatly improved,especially for the video sequence of severely occluded or have very similar background information.Nowadays,most of the correlation filtering uses the first frame of the video sequence to train model online.The introduction of deep learning has became a key method to solve the problem.The target tracking algorithm based on the siamese network,which uses the idea of similarity to realize the task of tracking.It uses a convolutional network to extract object features,and uses only one score map to predict the object location so that discards information from other score maps.In view of the above deficiencies,this paper studies the influence of depth features on object tracking,makes full use of the information contained in each score map,and adopts weighted fusion processing on the score maps from different search areas,which can finally predict the object location more accurately.End-to-end offline training and online tracking are used throughout the tracking phase.From the experimental results and analysis,compared with the traditional algorithm,the proposed algorithm has been greatly improved in robustness of tracking performance,and the tracking speed has also been guaranteed.