Multi-level Feature Selection And Feature Fusion In The Application Of Visual Tracking

Visual tracking is an important field in computer vision.It is widely used in video surveillance,motion analysis and traffic regulation.Although there are a large number of literatures to present the visual tracking solution,it is still a challenging research topic due to the existence of unfavorable factors such as deformation,motion blur,occlusion and illumination variation in the scene.It's a common way to solve visual tracking task as a classification of the object and background.It does not need to construct a complex model to describe the object,but rather a classifier different from the object and the background.Grabner et al.proposed an online object tracking algorithm based on the Boosting method,what is a typical visual tracking algorithm based on classification.The algorithm distinguishes the object and the background from the random location Haar-like features,training weak classifier to select effective features.This paper attempts to use multi-level feature selection and feature fusion to achieve object tracking task.The Boosting object tracking algorithm is only used to select the position feature in the object area.In this paper we propose an improved Boosting algorithm,adding the selection of the filter type.In the framework of Boosting algorithm,different hierarchical features and different dimension features suitable for tracking in the depth network are selected.GPU-based parallel mechanism is used to accelerate the improved Boosting algorithm.1.Based on the Boosting tracking algorithm,we propose a two-level Boosting tracking algorithm.The improved method extracts the object local feature with many filter templates,selecting the small image's position and corresponding filters in the object area respectively and combining the two type features effectively,which improves the accuracy of object tracking.2.In this paper,the output of the middle layers of the deep neural network is used as the feature maps input Boosting algorithm to achieve the object tracking,the purpose is to select the suitable features for tracking tasks.Boosting is used to select different levels of feature and different dimension features in depth neural network,finding the suitable feature combination method for target tracking in the experimental results.3.This paper presents an acceleration strategy for the two-level Boosting tracking method.We accelerate complicated matrix operation in the two-level Boosting tracking method,with GPU's parallel computing power.Make the algorithm more time-efficient.