Research On Object Tracking And Transfer Learning Algorithm Based Siamese Network

As an important topic and research hot spot in the field of computer vision,object tracking is widely used in intelligent video surveillance,video-based human-computer interaction,intelligent visual navigation and other fields.The basic task of target tracking is to label only the first frame of any target in a video sequence,while the location of the target can still be accurately predicted in subsequent frames.Although object tracking algorithm currently has been able to adapt to a variety of practical scenes,but in the face of occlusion,illumination changes,deformation,motion blur,similar interference and other complex scenes,accuracy,stability and real-time of tracking are still facing serious challenges.In addition,facing tracking tasks that lacking datasets,it is difficult to achieve satisfactory results when training neural networks with traditional methods.In view of the above problems,based on Siam RPN++ algorithm,this paper does two aspects of research work as follows:(1)A object tracking algorithm of siamese network combined with feature fusion and deep connected channel attention is proposed.Firstly,the features from third,fourth and fifth convolutional layer of the template and detection frame are extracted by the feature extraction network of Res Net50,respectively;Secondly,in the deep connected channel attention module,the feature channels of each convolutional layer are selected and the adjacent channel attention module is connected by the deep connected attention network,so that the feature information can flow among the attention modules and gaining more discriminative channel feature;Then,in the feature fusion module,in order to improve the feature representation ability of feature map in each convolutional layer,integrating the shallow feature containing rich position information with the deep feature containing advanced semantic information;Finally,the feature of each layer is fed into the Siam RPN module,respectively,and utilizing a depth-wise cross correlation layer to obtain the classification of the target-background and the regression of the anchor box.The results of tested on VOT2018、VOT2019 and OTB100 datasets show that the proposed algorithm has great tracking performance.(2)A multi-rotor drone tracking algorithm of siamese network based on transfer learning is proposed.Firstly,collecting image data of the multi-rotor drone,and making training datasets and testing datasets.Secondly,making a transferability analysis of deep neural networks,then taking the domain of tracking any target as the source domain,and taking the domain of tracking the multi-rotor drone as the target domain,the backbone network and the region proposal network of the pre-training model Siam RPN++ are fine-tuned layer by layer with a pre-training-fine-tuning method,so that completing transfer learning from the source domain to the target domain.Then,designing a pre-training-fine-tuning experience to find the tipping point for parameters of model fine-tuning,and validating it with the deep attribution graph.After that,designing several comparative experiences,the superiority of the transfer learning algorithm is proved,and the training parameter values are determined,training and obtaining the transferring model with the relatively best performance,then transplanting it to the DJI manifold 2embedding platform and testing performance;Finally,several universal regularities of the model-based transfer learning method are summarized,which can be extended to other related tasks about the special target.