Research On Moving Object Tracking Algorithm Based On Correlation Filtering

As an important research content of computer vision,object tracking has a wide range of applications in surveillance security,intelligent driving,etc.The moving object tracking algorithm based on correlation filtering still has low tracking accuracy and poor real-time performance due to challenges such as limited features,insufficient samples,and scale variation.In view of the above problems,this thesis uses the feature fusion to obtain richer feature representation,and uses the sample storage and sample library compression to improve the expression ability of model.The tracking quality evaluation and the prior distribution of target position are utilized to reduce the influence of background interference on model.Finally,correlation filtering and instance segmentation are combined together to solve scale variation problem.The proposed algorithm effectively improve the performance of the tracker.Feature is the foundation of the correlation filtering object tracking algorithm.Manual and deep learning features are the two most common features in tracking field.The current object tracking algorithms are mainly based on manual features or a single layer deep feature,which have limited application scenes due to their low tracking accuracy and speed.This thesis adopts the frequency domain feature fusion method,which combines the tracking response maps of different resolutions according to frequency and takes advantages of semantic feature of deep layer and texture feature of shallow layer of convolutional neural networks to improve the accuracy and stability of tracking algorithm.Training samples play an important role in the correlation filter solving.The traditional correlation filtering algorithm uses the samples generated by the tracker to train the correlation filter online,which easily results in the loss of target information.In this thesis,the sample library is used to store historical samples,and the correlation filter is solved with conjugate gradient method.Then,the sample library is compressed according to the sample similarity,which is beneficial to reduce the information redundancy of sample library.The tracking quality evaluation method is introduced to adjust the prior weight of the training samples to avoid the model drift and improve the stability of the tracking algorithm.The priori distribution is utilized to eliminate the background interference and solve the short-term target occlusion problems under fixed cameras.Scale variance adaption is an important part of the object tracking algorithm.The traditional scale pyramid method has a large amount of calculation and low real-time performance.In this thesis,a two-stage tracking algorithm combining correlation filtering and instance segmentation is proposed to solve scale variance problem of object tracking.In the first stage,the target is searched on a single-scale image,and the candidate bounding box is generated from the target position.The second stage,the RoI feature of the target area is extracted by the candidate bounding box,and we use the target template to weight each layers'feature through convolution and elementary summation.The proposed method combines the features from deep layers and shallow layers to achieve pixel-level segmentation of the target,and then it determines the target position using the minimum envelope box.The tracking accuracy and success rate of the proposed algorithm on the OTB100 dataset are 0.858 and 0.641,respectively.And the proposed tracker can run faster than traditional correlation filtering based trackers.Finally,the research work carried out in this thesis is summarized,and the future research direction is prospected.