Font Size: a A A

Research On Infrared Target Tracking Technology Based On Multi-feature Kernel Correlated Filtering

Posted on:2020-01-24Degree:MasterType:Thesis
Country:ChinaCandidate:C C PanFull Text:PDF
GTID:2428330599462094Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of the field of computer vision,the research of target tracking technology,as one of important research directions,has also made breakthrough progress with the tide of science and technology.As one of the target tracking fields,infrared target tracking is widely used in many aspects such as video surveillance,precise guidance,individual combat and unmanned driving.Compared with visible target tracking,it can not be disturbed by dim light and haze,and has the advantages of anti-interference,concealment and all-time.However,the research of infrared target tracking should not only deal with many challenges,such as target occlusion,scale change and target similarities,but also deal with the problem of insufficient feature model and scene information in infrared image due to the lack of information.In this paper,based on the knowledge of kernel correlation filtering,an infrared target tracking algorithm based on multi-feature kernel correlation filtering is proposed for the difficult problems in infrared target tracking.The main work and innovative research results of this paper are as follows.1)From the point of view of feature representation and model updating,a kernel correlation filter infrared target tracking algorithm based on multi-feature adaptive fusion is proposed.Multi-peak detection and high confidence model updating strategy ensure accurate target location and prevent model drift and model contamination.At the same time,it reduces the times of model updates and improves the speed of the algorithm.Calculate the feature fusion coefficient with relative confidence,and adaptively select which feature's confidence response map is more suitable for different scenarios.Based on the scale pyramid,the refined scale updating strategy can estimate the scale change more accurately by using Newton iteration method.Experiments show that the accuracy of the algorithm is significantly better than other popular tracking algorithms,and the speed can reach 80 frames/sec.2)From the point of view of feature extraction and feature fusion in convolution network,a kernel correlation filter infrared target tracking algorithm with convolution feature adaptive fusion is proposed.A large number of labeled visible light images are transformed into infrared-like images by using antagonistic neural network,the new residual network(ResNeXt)is trained by generating infrared-like images and real labeled infrared images,the pre-training model is used in tracking algorithm to extract convolution network features of infrared targets.Based on the fusion of deep feature extracted by convolution network and shallow feature(directional gradient histogram feature and gray feature),a quality evaluation method of response map,which can reflect both accuracy and robustness is proposed,and the adaptive fusion coefficients are calculated to complete the adaptive fusion of deep feature and shallow feature,which enrich the expressive ability of target.The appearance changes of deep feature for rotation and deformation are constant,and the robustness of tracking is effective,however the shallow feature mainly contains texture and gray information of the target,which has high spatial resolution and is suitable for high-precision target location.After fusion,combine the advantages of the deep feature and the shallow feature to obtain high-precision and robust tracking algorithm.
Keywords/Search Tags:infrared target tracking, correlation filtering, feature fusion, residual network
PDF Full Text Request
Related items