Research On Object Tracking Algorithms Steered By Recurrent And Siamese Neural Network

Visual objects tracking is a fundamental and important research task in the fields of computer vision.And it is widely application and real value in many ways such as video surveillance,automatic drive,intelligent transportation,military targets location,etc.Because of the high complexity of tracking tast,traditional object tracking algorithms are challenged by a variety of difficulties.Recently,deep learning has the advantage of the mining data information and the power of fitting problems,so that make great progress in the fields of face recognition,images classification,object location and so on.Simultaneously,it has to provides a new perspective for solve the challenges and problems in objects tracking field.At present,many algorithms based on deep neural networks has improved the tradition algorithms,or put forward some new theories.But build a real-time and robust tracking system effective to process variety of complex scenes is still a huge challenge.Our paper utilize deep neural network model(especially the Recurrent neural network and Siamese network)as the theoretical basis,focus on feauter learning,background and boundary supression,real time and robust requirement in the tracking process,combining with related to the traditional tracking algorithms carried out the research work,the main works are as follows:1)We propose a spatial multi-directional recurrent neural network steered hierarchical regularized correlation filter tracker.The traditional tracking algorithms of correlation filters affected by the impure cyclic samples and the limitation representation capability of the handcraft features,its will decay the discriminative of the correlation filter.We use spatial multi-direction RNNs encoding the context information outside the target region,and to adaptively regularize the learning process of kernelized correlation filters using multi-scale convolutional features.The proposed tracker can more focus on target region so that significantly alleviate the boundary effects and background disturb,simultaneously extract the hierarchical convolutional feature to adapt various appearance changes,in order to reduce the risk of tracking failure.2)We porpose a multi-stage co-inference tracking algorithm via multi-task siamese neural network.Visual tracking is essentially a matching problem of video frames.The end-to-end Siamese network for sample similarity matching that without any prior information,no model update strategy and only simplified network structure.Result in that tracking model is difficult to effectively cope with target appearance changes for a long time.Although it can guarantee the real-time,but has a low accuracy.We pre-train an end-to-end multi-task Siamese network,and devide tracking problem into two sub-problems: regression and classfication.According similarity to determine whether directly predict the target location and scale offset,and through the classification branch bring prior information to adjust regression results.In order to further imporve the robustness,combined with the adaptive discriminative correlation filter to re-detect the lost target.The ensemble adaptive tracker model can actively trigger a multi-stage co-inference approach to enhance robustness and discriminative,and alleviating the complexity of the model and the real time can not achieve the consistency balance dilemma.