Scale Adaptive With Feature Fusion And Multi-classifiers Parallel For Object Tracking

Visual object tracking is one of the fundamental applications in computer vision,which is widely used in the fields of behavior analysis,intelligent video monitoring,human-computer interaction and so on.However,the challenges of illumination variation,occlusion and background clutter in the actual scenes will have adverse effects on the algorithm.Therefore,it is of great significance to study a robust,high real-time and high-precision tracking algorithm.Under the framework of multi-classifier parallel tracking,scale adaptive with feature fusion and multi-classifiers parallel for object tracking algorithm is proposed.The main works are as follows:(1)A tracking algorithm based on feature fusion and multi-classifier parallel is proposed.Firstly,multi-feature fusion and cascade iterative training model are combined to improve the expression ability of the classifier.Then the scale of the initialization window is adaptively adjusted according to the aspect ratio of the target in the initial image frame.Finally,the multi-classifier parallel strategy is adopted to further improve the robustness of the algorithm against dynamic background interferences.(2)In order to reduce the integration of pseudo-target information into the classifier,an update strategy with conservative and adaptive learning rate is proposed.On the basis of setting a threshold to achieve conservative update,the learning rate is adaptively adjusted according to the classifier response,which greatly reduces the integration of pseudo target information into the classifier.(3)A scale adaptive tracking strategy is proposed for the problem that the target scale cannot be accurately estimated.Firstly,the scale estimation method is integrated on the basis of position estimation.Then,the strategy of updating the position classifier with only the target scale in the initial frame is proposed,which eliminates the problem of position classifier performance degradation caused by inaccurate scale estimation in the traditional algorithm.(4)In order to improve the robustness of the algorithm to long-term occlusion or severe deformation,a long-term tracking mechanism is proposed.A redetection module,combining a long-term filter with an online SVM detector,is designed to achieve robust and long-term tracking.Finally,the position estimation,scale estimation and long-term modules are tested on the OTB2015 and MEEM respectively,to verify the robustness of the algorithm to dynamic background interferences,severe deformation and long-term occlusion.The experimental results show that the tracking results of the proposed algorithm are greatly improved in the central location error,precision and success rate indicators.