Research On Siamese Network Object Tracking Algorithm Based On Higher Order Statistics

Object tracking is the course of estimating the scale and accurately positioning the object in the subsequent frames given the object information in the initial frame of the video.At present,object tracking technology has been extensively used in the domains of human-computer interaction,unmanned driving and military applications.Although a host of phased research progress has been made in the domain of object tracking,there are still many challenges in practical applications,such as: similar backgrounds,fast motion,and object occlusion et al.For the past few years,due to the increase of deep learning technology,methods based on deep learning have been expansively used in the domain of object tracking,among which the algorithms based on Siamese network are favored.As a representative of deep learning-based methods,the Siamese network-based algorithm uses a convolutional neural network with the same structure and shared weights to convert the object tracking task into a similarity matching task.Despite the algorithm based on Siamese network has achieved good tracking performance,there are still some problems: Firstly,this kind of algorithm tends to erroneously track the same kind of objects,resulting in tracking drift.The reason is that this kind of algorithm has insufficient ability to discriminate specific object in similar objects,and cannot distinguish specific object from similar objects well.Secondly,most of these algorithms use a relatively shallow backbone network,and the extracted features are shallow apparent features,lacking deep semantic features.In addition,this type of algorithm uses the initial frame as a template and does not consider the template update problem.Based on the Siamese network algorithm,this thesis proposes some optimization measures for the problems existing in this kind of algorithm,and has completed the following two aspects:(1)Aiming at the problem that has insufficient ability to discriminate similar objects of Siamese network algorithm,a lightweight Siamese network object tracking algorithm based on Second-order Pooling feature fusion is proposed.The main idea of the algorithm is as follows: firstly,use the convolutional neural network as the backbone of the Siamese network to obtain the deep features of the object;secondly,a Second-order Pooling network and lightweight channel attention are added in parallel at the end of the Siamese network structure to obtain the Second-order Pooling features and channel attention features of the object;finally,the depth features of the object,the Second-order Pooling features and the channel attention features are fused,and the fused features are used to achieve object tracking through cross-correlation operations.(2)Aiming at the problem of insufficient feature expression ability and discriminative ability of Siamese network algorithm,a lightweight double branch response Siamese network object tracking algorithm based on Second-order attention is proposed.The main idea of the algorithm is as follows: firstly,use a convolutional neural network with better performance and deeper depth as the backbone of the Siamese network to obtain the deep features of the object;secondly,the Residual Second-order Pooling network proposed in this thesis and the Second-order Spatial Attention network are used in parallel at the end of the Siamese network structure to obtain the Second-order attention features with channel correlation and the Second-order attention features with spatial correlation;finally,object tracking is achieved through a double branch response strategy using Second-order channel attention features and Second-order spatial attention features.