Research And Application Of Object Tracking Based On Siamese Network

Object tracking is a basic but still challenging task in the field of computer vision,which has been concerned by experts and scholars at home and abroad.Object tracking can be applied to various visual fields,such as intelligent video surveillance,unmanned driving,human-computer interaction and so on.The task of the object tracking algorithm is to predict the state change of the object in the subsequent video frames and mark its position and size after the initial state of the object is given in the initial frame of the video sequence.Although great progress has been made in the research of object tracking algorithm in recent years,there are still some factors affecting the performance of tracking algorithm,such as occlusion,appearance change,scale change,etc.In addition,the real-time performance of the algorithm should be considered.Therefore,it is necessary to study the tracking algorithm more deeply.Although the methods based on correlation filter perform well in object tracking benchmark test,most of these methods only use hand-made appearance features to represent the tracking object,which makes the representation ability of the object limited,and often can not achieve satisfactory performance under the interference of occlusion and background speckles.Aiming at the problem of feature extraction,this paper proposes an end-to-end feature fusion framework based on siamese networks.This framework can effectively integrate the features of CNN and manual design,solve the problem of parameter learning in feature fusion,and improve the versatility of object tracker.Since 2016,object tracking algorithms based on siamese networks have appeared in public view and achieved good results in various competitions.One of the most representative is the full convolution siamese network tracking algorithm(SiamFC).Although it meets the real-time requirement in terms of speed,it is easy to cause tracking failure when the object changes dramatically due to the adoption of a shallow AlexNet and the absence of online updates.Aiming at the problems of SiamFC,this paper adopts the siamese network model with ten-layer convolution under the framework of tracking algorithm of siamese network,and optimizes it by two methods:(1)Change the feature extraction network from AlexNet to VGG network which is suitable for object tracking task after modification,and use the stronger information representation ability of deeper network to improve tracking accuracy;(2)On this basis,in order to further enhance the discriminant ability of the network model,a new algorithm using two attention mechanisms to adjust the model is proposed,which can selectively emphasize useful information and suppress less useful information,so that the algorithm can be better applied to actual scenarios.Finally,in order to verify the effectiveness of the algorithm,a siamese tracking framework based on attention mechanism is constructed and tested on a common OTB data set.The results show that the improved algorithm has higher accuracy and better robustness than other comparative algorithms without reducing the tracking rate.