A Tracking Algorithm Based On Learning Pixel-based Feature

As one of the core components of computer vision,visual object tracking aims to continuously locate the target in a video.It has been applied in many practical applications,such as surveillance systems,human-computer interaction,sports analytic,medical imaging etc.The primary factors,which limit the performance of trackers,are occlusion and scale change.On the one hand,occlusion and scale change cause the various target appearance in the video,which will make it difficult to detect and match the target.On the other hand,the known information of the target is generally only the single-frame image containing the target,which is insufficient to help construct an accurate and robust appearance model.Therefore,it is significant to construct a higher precision and more robust target apparent model with limited target informationIn this paper,we propose a tracking algorithm based on learning pixel-based feature.The apparent model of the target consists of two sub-models,which are the local apparent model based on pixel-based feature and the overall apparent model based on gray statistical histogram features.The former is a strong classifier model composed of multiple perceptron models,which are trained based on the Adaboost algorithm to label the pixels in the next frame as belonging to either target or background.The latter is a grayscale statistical histogram model based on the classifier output for target location and scale estimation.In the tracking process,the target histogram model and the mean shift algorithm are used to locate and scale the target,and the final tracking result is used to update the apparent model of the target.The tracking experiments based on the VOT2018 database show that the proposed algorithm can effectively deal with occlusion and scale changes.