Research On Object Tracking Based On Siamese Network

As an interdisciplinary research hotspot of image processing,pattern recognition,computer vision and many other disciplines,object tracking has been continuously concerned by experts and scholars at home and abroad.Object tracking is widely used in video surveillance,military reconnaissance,unmanned driving,intelligent transportation,robot and other fields,which has important theoretical research significance and practical application value.Object tracking aims at the region or object of interest in image sequence or video stream,according to the given position and scale information in the initial frame,predict and mark the target in the subsequent frame,and finally get the motion track,specific shape and position information of the target.Although people continue to expand the application field of object tracking technology,however,in the actual application scenarios,object tracking faces great challenges due to the uncertainty of the object and the complexity of the background environment.Therefore,it is a great challenge to design a fast,accurate and robust object tracking algorithm.With the development of artificial intelligence,deep learning is widely used in the field of object tracking.Compared with the traditional method,the object tracking algorithm based on deep learning has a great improvement in speed and accuracy.In recent years,the siamese network object tracking method has achieved the balance of robustness and real-time performance,and achieved breakthrough results.In this paper,object tracking method based on siamese network is proposed.The input is video stream information,and the output is the location information of the target in the video stream.In this paper,the siamese network structure is used to map the input to a new space and form a new space representation through network branches.Then,the similarity evaluation of the input is realized through the calculation of the loss function,so as to calculate the relationship between frames in the video stream.In view of the existing problems of Siam RPN,this paper optimizes it by two methods:(1)In this paper,attention mechanism is introduced.Through spatial attention mechanism and channel attention mechanism,significant appearance features are selected to improve the network's ability to distinguish semantic background and target foreground,so that the algorithm has better robustness to background similar interference.(2)In this paper,through the way of multi-layer feature fusion,we extract the low-level and middle-level features that contain the spatial information of the target object and the high-level features that reflect the semantic information of the target,and design a number of anchor free region proposal networks,so that the network can not only predict the specific location of the target more accurately,but also better adapt to the multi-scale changes of its appearance.Finally,the effectiveness of the algorithm is verified by experiments on OTB2015 data set.The experimental results show that the improved algorithm is more accurate and robust.There are 23 figures,6 tables and 56 references in this paper.