Researches On Object Tracking Based On Convolutional Neural Networks

Object tracking is considered as an important issue in image processing and computer vision area.It has been widely used in both military fields such as precision guided weapons,detection of warning,unmanned aircraft and civilian fields such as intelligent traffic,robot navigation and medical imaging diagnosis.Object tracking is a process of estimating the state of a specific target in a continuous video sequence.Due to the complexity of the imaging conditions and the diversity of the scene,it is difficult to achieve the stable and effective tracking algorithm.According to the technical difficulties during the object tracking,this paper makes a deeply study on how to model the object appearance,based on which we propose two novel tracking algorithms based on convolutional neural networks.In traditional object tracking methods,the hand-crafted selection of feature representations would limit the performance of visual tracking and the predictability power of semantic information of the object.To address this problem,this paper proposes a new algorithm for object tracking based on multi-scale representation of convolutional neural network.We make a multi-scale convolutional network architecture generated from the Laplace pyramid.Each scale is trained from coarse to fine using video datasets and shares weights across all layers.To solve the saturation problem,we improve the multiple instance learning method and train an improved MIL classifier with the multi-scale feature representations to realize the online tracking.This makes the appearance model more robust to changes of the object,which can achieve more stable tracking effect and improve the tracking accuracy and success rate.To effectively alleviate the "drift" phenomenon caused by the appearance changes,an object tracking algorithm based on attention mechanism is proposed.The initial frame is regarded as the memory unit to make the network maintaining feature memory of the first frame.We design an attention layer with regard to the first frame according to the content correlation between video frames and construct a network based on attention to learn the importance of the image.From different layers,we extract features by spatial pyramid pooling and build multiexpert classifiers to achieve the target.Experimental results show the method can effectively alleviate the drift phenomenon,and achieve stable tracking results in a variety of scenes.