Research On Multi-modality Object Fusion Tracking Method Based On Spatiogram And KCF Representation

Target tracking is to realize visual surveillance,human-machine interaction,and vehicle navigation and so on.Many target tracking methods have been studied,and these methods can be classified into two categories: using single-modality and using multi-modality.The so-called singlemodality tracking refers to the tracking of target objects from a video source.And its mainstream methods include kernel density estimation,pattern classification,sparse representation and subspace analysis and so on.Multi-modality tracking is accomplished by using data of different modality,including multi-sensor fusion method,multi-spectral fusion method and multi-feature fusion method.Above three methods of multimodal tracking methods have themself characteristic,but there are few methods to combine them together,this paper proposes a unified framework to solve this problem,using modal data more complete track of unified target object.Our framework allows the modalities to be original pixels or other extracted features from single image or different spectral images,and provides the flexibility to arbitrarily add or remove modalities.Research on multi-source target fusion tracking method based on spatiogram and Kernelized Correlation Filters,this paper puts forward the following two unified multi-modality fusion tracking framework:1)Multi-modality target tracking algorithm based on spatiogram representation.Firstly,in the method,the second-order histogram is employed to represent a target,and weighted fusion of them to build the target function,Secondly,the feature similarity between each candidate target and its target modality is calculated using the Bhattacharyya coefficient and the Mahalanobis distance respectively.And spatial similarity and weighted fusion of them to build the target function;then,according to Taylor expansion and gradient minimization method to derive the joint displacement formula of multi-modality target.Finally,the automatic fast search of multi-modality target is accomplished by using mean shift method.2)Multi-modality fusion tracking algorithm based on Kernelized Correlation Filters.Firstly,a large number of samples are generated for each modality target image by means of cyclic shift sampling,and the characteristics of each modality sample are extracted,and the apparent modality is established for each modality target.Then,using ridge regression function to train the appearance modality of each modality target and obtain the modality target filter template;Then,each candidate sample in the corresponding test frame is detected by the modality filter template,and the maximum response value of each modality target is obtained.Finally,according to the maximum response value of each modal target,the weight of the corresponding position is allocated,and the final target position is obtained by weighted average.The test results on multiple datasets show that the image sequence based on spatiogram said multi-modality fusion tracking has good robustness performance,and multi-modality fusion tracking method based on Kernelized Correlation Filters can better distinguish target and background.The two multi-modality fusion tracking methods have achieved well results in the dealing with target intersection,illumination change and target occlusion.