Correlation Filter Tracking Method For Complex Scenes

Visual tracking is an important research direction in the field of computer vision,including many cutting-edge technologies such as pattern recognition and graphic image processing.Today,our country is vigorously building smart cities,and visual tracking technology is mainly used in intelligent video surveillance,intelligent transportation,driverless cars,military and other fields.Due to the complexity of real-life scenes,the target’s own deformation and the real-time requirements of computer hardware and tracking algorithms,the existing single-target tracking technology based on video sequences has not been commercialized on a large scale.Therefore,finding a tracking algorithm that meets the requirements of real-time and has high robustness has important research significance and broad application prospects.In general,visual tracking techniques are divided into two categories: generation models and discriminant models.The generation models only considers how to accurately construct the target while ignoring the surrounding background information,and the tracking performance is not very good;while the discriminant model effectively utilizes the target information and makes full use of the background information,and it shows good performance.The visual tracking algorithm based on correlation filter is one of the discriminant models.In recent years,great progress has been made.The cyclic shift method is used to expand the number of samples,and the target region and its surrounding background region are used to train the classifier online.And the process of training the classifier and detecting the target is placed in the Fourier domain,which greatly speeds up the calculation.Such algorithms are very effective.The research finds that this kind of algorithm has excellent expansibility.This paper makes an in-depth study on the existing tracking algorithm based on correlation filter and deep learning.The main work is:A correlation filter tracking method for complex scenes is proposed,which is called multi-template tracking algorithm based on correlation filter.The algorithm extends and improves the correlation filter tracking method from three different aspects.Firstly,from the perspective of target expression ability,this paper deeply compares and studies the manual features and deep features extracted by convolutional neural networks.In order to ensure the accuracy and real-time at the same time,this paper extracts the deep feature and the color name feature respectively for the image or image region.The obtained two feature maps are respectively learned by the correlation filter to obtain different filter templates.Then,the correlation filter tracking algorithm is used to obtain the response set under the two features,and the obtained set is weighted and fused.The final target position is obtained by determining the maximum value of the maximum values.Finally,from the perspective of updating the target scale value,the existing method is based only on the target information of the current frame,so that the current frame target scale value estimation is not accurate.This paper uses Bayesian statistics to estimate the optimal scale value of the target by maximizing the posterior method and updating the relevant filter parameters to achieve adaptive scale visual tracking.The proposed algorithm is tested on two benchmark databases,OTB2013 and OTB2015,and compared with the current six algorithms.The results show that the accuracy and real-time considerations of this method are optimal.On two data sets,the success rate OP(AUE)is 10.7% and 12.4% higher than the classical tracking algorithm KCF,respectively.