Algorithms Study On Multi-Clue Object Tracking Via Deep Learning And Transfer Learning

The target of visual object tracking algorithm is to allow the computer to simulate the human brain and visual,using the tracking algorithm to automatically identify the trajectory of the target object and do something. The study of visual object tracking algorithm can not only effectively promote the theory development of science of computer vision and machine learning, but also has important application value in intelligent human-computer interaction, unmanned, intelligent traffic monitoring,etc.In view of the important theoretical significance and application value of the target tracking algorithm, many experts and scholars at home and abroad for the construction of the target tracking algorithm of three key technologies(namely,object appearance model,search mechanism and model update), respectively, launched a broad and in-depth study.The performance of the target tracking algorithm has been greatly improved, and it has been used in certain occasions.However, because of the complex real scene(such as occlusion, light illumination changes, etc.),the accuracy,robustness and real-time performance of target tracking algorithms are still difficult to achieve large-scale application and industrial standards.In order to solve the problems mentioned in the above, we will introduce what we have done.In this paper,we use deep learning to extract features, transfer learning,selective search,multiple cue fusion etc.Object appearance model,search mechanism and model update are studied respectively.The main research contents are as follows:1) We propose a robust visual tracking algorithm based on convolutional deep belief network and transfer learning. The traditional manual selection method has the shortcomings of domain knowledge, non optimal, time consuming, and poor generalization ability.Compared with hand design features,the algorithm automatically learns more effective features from images and videos.In the framework of particle filter, the target apparent model is constructed by using the convolutional deep belief network.First, the model is trained off-line to learn the general characteristics.Then,the characteristics of offline learning are migrated toonline model.In order to effectively alleviate the "drift" phenomenon,the positive samples are obtained by the following three methods:positive samples of initial stage(obtained from the first frame by hand),positive samples obtained from online tracking(automatic adjustment of samples in recent video frames),positive samples obtained from the current frame.Experimental results show that the proposed algorithm can achieve more accurate tracking results.2) A target tracking algorithm based on online PCA network is proposed.The model parameters of target tracking algorithm based on deep learning model are large,requiring large data sets for training.Online training sample is short,and offline training data sets may not be consistent with the specific tracking target categories,as a result,this can't get a good discriminant model.The algorithm does not require a large number of training samples,alleviating the problem of over fitting due to the small number of samples in the target tracking problem.This algorithm combines principal component analysis(PCA) and deep learning model,effective fusion of global and local information of the target object.Finally, through the use of a variety of ways to obtain online training samples, further easing the "drift" problem.The experimental results verify the effectiveness of the proposed algorithm.3) A multi-clue tracking algorithm based on selective search and deep learning is proposed.Traditional single cue tracking algorithm can not deal with complex scenes.When the scene is complex and changeable, the single clue can't identify the object.For this reason, a multi model fusion method is proposed. To extract the robust feature,the algorithm used complementary multi-clue.In addition, because of the complexity and uncertainty of the motion of the object, a simple search strategy and a single motion model used have been unable to obtain high quality candidate particles.So, the algorithm uses particle filter and selective search method to obtain particles.The experimental results show that the algorithm can improve the accuracy of tracking.