The Research On Single Object Tracking Based On Siamese Fully-Convolutional Neural Network

Visual object tracking has become the basic technology of vision system,such as video surveillance,automatic driving and human-computer interaction.Because of its simple architecture and good performance,Siamese fully-convolutional network tracker becomes one of the important methods of object tracking.In this kind of trackers,convolutional neural network is used for feature extraction,and then feature is used for object location.Early Siamese fully-convolutional network tracker used shallow convolution neural network for feature extraction.The features extracted by this kind of network have limited discriminative power,so it is impossible to distinguish the object in complex situation.In addition,the Siamese fully-convolutional network tracker needs a lot of labeled video frame for training.If there is only one object in the image,there will be a large number of negative samples in the labeled data.And the common loss function is not easy to learn the difficult samples.For these problems,this thesis has carried out the following two aspects of research:1.In order to solve the problem that the feature discrimination ability of shallow convolution neural network is insufficient,a new Siamese fully-convolutional network architecture is designed based on the densely connected network with reusable features of each layer.In order to adapt to the tracking task,this thesis first analyzes the factors that affect the tracking results in the backbone network,and then improves the original densely connected network,adding the center clipping operation to remove the influence of convolution padding on the tracking.In addition,the output feature mapping size of the network is controlled by the influence of the feature receptive field on the tracker.The results of experiments show that the redesigned tracker based on Siamese densely connected network has better accuracy and robustness.2.To solve the problems of imbalance between positive and negative samples and hard to learn difficult samples in training data,two balance loss functions are introduced to improve the tracker of object location in two different ways.One of the loss functions is to improve the ability of the tracker backbone network to fit the object by introducing negative sample penalty parameters.Another loss function improves the ability of the tracker to learn difficult samples by using the confidence penalty parameter.Finally,the tracker improved by the balanced loss function are compared with tracker without improvement on several datasets.The experimental results show that the balanced loss function can enhance the discriminative power of the extracted features of the backbone network.The above works make the Siamese fully-convolutional network tracker benefit from the deep convolutional network and a large number of annotation data.Experiments on the object tracking evaluation dataset show that these improvements are effective.