Study On Visual Tracking Algorithm And Analysis Based On Binary Classifier

Video and images are important sources of information for people to obtain the appearance,color and other characteristics of the target.An important research direction of computer vision is how to make the computing device acquire the target independently and analyze the content of the corresponding video or image,just like human beings.As an important research topic in the field of computer vision,visual tracking is closely related to intelligent transportation,automatic control and other application scenarios and it is used to ensure the system that can capture the corresponding targets with long-term stability.Usually,an excellent visual tracking algorithm needs an effective target appearance model,perfect theoretical basis and efficient solution method.Although in current literature,researchers have achieved many research results in the field of visual tracking and put forward many related theories and algorithms,they are still confronted with enormous challenges.Due to the rich information in the video sequence and the complex scenes,it is often difficult for each algorithm to track the target robustly.In video sequences,when confronted with the challenges of target occlusion,out-plane rotation and motion blur,the drifting or losing of the target often occurs.Therefore,in practical applications,those algorithms are often not competent for visual tracking tasks in complex scenarios.In order to solve the above problems and ensure that the algorithm can effectively cope with various unfavorable factors which pollute the information of the target area and interfere with the video quality,this paper conducts the following research based on particle filter framework:(1)We propose a visual tracking algorithm based on a cooperative model,which combines the multi-task reverse sparse representation model(MTRSR)and the AdaBoost classifier to cope with the disturbing of target gradient information caused by motion blur or serious target occlusion,and at the same time,a descriptive dictionary is used to estimate the weights of each candidate.The specific methods are as follows: First,we use the MTRSR model to get the blur kernel k,which is used to get the blur target template set.Meanwhile,the confidence of each candidate is also obtained by the reconstruction error.Then,we use the HOG features of the target templates to get the descriptive dictionary to calculate the weights of each candidate,and an AdaBoost classifier is used to calculate the confidence of each candidate.Finally,the best candidate is retrieved by the sum of production of weight mulplied by two confidences.Compared with the related mainstream algorithms on the open source dataset,the average overlap rate of our algorithm is 12.2% higher than that of the second algorithm L1 APG,and the center position error is reduced by 22.4 pixels.It shows that our algorithm can cope with the change of the appearance information of the target,which is caused by moving blur and occlusion in complex scenes and it has better accuracy and robustness.(2)The computational cost of visual tracking algorithm based on particle filter framework mainly comes from the extraction of target features,and training and updating of the model.In order to ensure that the algorithm can effectively deal with both the interference of the video quality and the dramatic changes in the appearance of the target caused by the target occlusion and illumination variations,we propose a visual tracking algorithm based on robust appearance model.Firstly,we obtain a discriminative appearance description of the target by calculating the global and local Fisher vectors of the target area,and subsequently,we utilize a semi-supervised linear kernel classifier to calculate the confidence of each candidate in order to distinguish the foreground target from the background region.Meanwhile,because the change in the appearance of the target has caused serious pollution and interfered with the original information of the target area,we obtain the similarity between the sub-patches of the target template set and all the candidates,based on the pollution degree of the patches in each candidate.Then,we calculate the weight of each candidate according to the reconstruction error of the candidate and the template set.Finally,we retrieve the best candidate according to the multiplication of the weight,confidence and the similarity.The experimental data shows that,compared with the other algorithms,our algorithm can effectively improve the tracking accuracy and ensure a comparatively stable visual-tracking task.To summarize,the results show that,compared with the current mainstream tracking algorithm on the OTB dataset,our algorithms perform better in tracking the target with high precision,which can more effectively and more robustly accomplish the visual tracking task in complex scenarios.Of course,the visual tracking algorithm based on robust appearance model still has some deficiency when confronted with severe deformation and drastic size variation of the target.Although the target is accurately captured,the average overlap rate cannot yet perform well,and this needs to be further explored and improved.