Adaptive Complementary Learners with Diversified Color Attributes for Object Tracking

Histogram of oriented gradient(HOG)features used in the sum of template and pixel-wise learners(Staple)tracker have a poor ability of feature representation for deformation and scale variation.In addition,the fusion and updating of model can not be implemented adaptively.To solve these problems,an adaptive complementary learners with diversified color attributes for object tracking algorithm is proposed.Firstly,on the basis of HOG features,color names features with great invariance of shape and scale are added.The multi-channel features are used to calculate the response map of translation filter.Then,the response map of histogram is calculated by using color histogram features.The maximum peak and the average peak-to-correlation energy(APCE)of two response maps are used to obtain adaptively their weights of fusion.Finally,the high-confidence updating of model is implemented according to the maximum peak and the APCE of fused response map.Results compared with 5 state-of-the-art trackers are obtained on two benchmark datasets:OTB-13 and OTB-15.The experimental results demonstrate that the proposed algorithm performs high robustness in the scenarios with the interference of deformation,scale variation,illumination variation and occlusion.Moreover,the proposed algorithm performs favorably against Staple and other comparison algorithms in terms of tracking accuracy and success rate.