Research On Object Tracking Using Local Invariant Descriptors

Object tracking algorithm contains two major aspects, target representation method andtarget searching method. The target representation method has a great impact on therobustness of the tracking algorithm. The main objective of this research is to investigatesome efficient novel representation methods based on the traditional color features. Moreprecisely, we will explore more effective target representation methods by improving texturedescriptors, which will increase the representation ability of the model on some local invariantfeatures.The main contents of this research are listed as follows：(1)In order to investigate the contribution of different color channels to theperformance of target representation, we introduce a new texture feature called OpponentColor Local Binary Patterns (OCLBP). Through extracting the correlation among differentcolor channels, three interactive patterns are selected to conduct the representation process.After analyzing all the10texture patterns of OCLBP, the proposed method extracts only thekey points of target region by using major pattern of OCLBP. Then, we use the OCLBPfeatures related to the key points to represent the target.(2)In order to reduce the computational complexity of the target model and investigatethe representation ability by combining the local invariant features with color features in thisresearch, we present a novel method of target representation for robust mean shift tracking.We add a new texture feature called Center-Symmetric Local Binary Patterns (CS-LBP) toreduce the number of patterns and simplify the computational process. This research analyzesall the16textures of CS-LBP and proposed a method which extracts only the key points oftarget region by using major patterns of CS-LBP. Then, we use the joint color and CS-LBPtexture features of the key points to represent the target.Finally, the proposed two target representation methods were applied to the mean shiftframework separately to track object. Compared with the traditional color based model, theproposed method increases the performance of object tracking by using only the major texturepatterns of OCLBP. Moreover, the proposed joint major CS-LBP texture and color patterns method increases the discrimination of the target and background, comparing with thetraditional color histogram model and the model including all the patterns. Experimentalresults on several representative videos under complex environments validate the accuracyand efficiency of our algorithm.