Siamese Network Visual Object Tracking Based On Adaptive Update And Feature Optimization

As an important branch of computer vision,visual object tracking plays an important role in many fields,including video surveillance,UAV navigation,autonomous driving,medical imaging,military guidance and so on.The main task of visual object tracking is to obtain a tracker.Given the position of an object to be tracked in the first frame of a video,the tracker can accurately predict the position of the object in subsequent frames.With the rise of the artificial intelligence wave,increasing number of scholars have focused on the field of visual object tracking.A large number of algorithms that can complete visual object tracking tasks have been proposed.However,it is still a huge challenge to accurately complete the visual object tracking task,due to background clutters,deformation,illumination variation,scale variation,motion blur,occlusion and other issues.In this paper,the visual object tracking methods based on siamese network are researched,the main results are as follows:(1)A visual object tracking method based on adaptive update of siamese feature network is proposed.In view of the problem that the selected backbone network is mainly suitable for object classification tasks,the target-aware model is constructed,the target loss function and scale loss function are used to optimize the model parameters,and the importance of each channel in the feature map is determined according to the gradient during backpropagation.In this way,typical features suitable for visual object tracking tasks are screened out.For issues such as scale variations of the object,an adaptive update model is introduced,a template pool is established,and the siamese network is guided by the similarity between the depth features of the current frame image of the video sequence and the template features.The update of the template improves the algorithm's ability to deal with problems such as the scale variations of the object,and makes the tracking result more accurate.In the experiment,the method was ported to the embedded system,stable and accurate tracking results were also obtained.(2)A siamese network visual object tracking method based on the cross feature of the DownResidual is proposed.Aiming at the problem of the shallow depth of the backbone network,a backbone network composed of the Down-Residual model is constructed,methods such as cropping are used to alleviate the impact of the padding operation in the convolutional layer on the translation invariance of the network,increase the depth of the backbone network to improve the model's performance,while limiting the stride of the backbone network and reduce the impact of too large receptive field on tracking accuracy.For the matching problem of the object to be tracked,the cross feature model is integrated,and three different scale convolution kernels are used to obtain the cross feature of the object,so that the model can obtain the depth feature of different scales,so as to improve the matching ability and tracking effect of the model.(3)A siamese network visual object tracking method based on feature fusion evaluation is proposed.Aiming at the problem that different levels of the feature extraction network contain different information,a feature fusion model is introduced to fuse low-level features with high-level features,so that the fused features contain not only low-level information such as color and shape,but also rich high-level information.By this way,the model can adapt to a wider range of tracking scenarios.For the problem that the preset anchor will affect the tracking results,a feature evaluation model is constructed,the visual object tracking task is regarded as a combination of regression and classification tasks.The classification task regards the object area in the input image as a positive sample,and the background area as a negative sample.The regression task directly regresses to the bounding boxes at the position classified as positive,avoiding the influence of prior information on the tracking results,and filtering out the bounding boxes with lower scores through the score evaluation mechanism,thereby improving the tracking accuracy.