Visual Object Tracking Based On Deep Learning

Visual target tracking is one of the hot issues in computer vision,during visual target tracking working,there are some problems like target deformation and complicated background,which bring about researchers a series of difficult challenges.With the development of deep learning and enhancement of computer hardware,visual target tracking abandons traditional methods and employs deep network to extract more abundant target characteristics and build more robust appearance model.Although the tracking difficulty is reduced,there are still some problems to be solved.The paper started with the difficulties easy to be ignored during tracking and improved the existing algorithm applied into target tracking by means of deep learning,obtaining some research achievement in building appearance model and settle target drift.Our major contributions of the paper are summarized as:To build an appearance model which could adapt to complicated scenes varying,we proposed deep learning algorithm based on meta learning and improved existing algorithm by meta learning framework,realizing fast learning in a data set with a small amount of data.We used the obtained general target model to conduct adaptive learning in the future frames of video sequence to build a well-timed target appearance model.To obtain a prediction bounding box with higher expected score than target bounding box predicted by selective search,we replaced selective search with deep search network based on regional recommendation,conducted regional recommendation search in local region rather than overall region and transformed localization problem of regional search into regression problem of Convolutional Neural Networks without increasing the number of regional recommendations,realizing end-to-end deep learning and alleviating target drift.