Siamese Network Based Visual Object Tracking

With the development of the Internet of Things and image sensors,there are more and more applications and demands for intelligent video processing algorithms.Visual object tracking is an important part of intelligent video processing algorithms.Traditional visual object tracking algorithms are often devoted to improving the performance of algorithms by improving manual designed features or optimizing machine learning models.It is difficult to design the most appropriate feature engineering and tracking framework.Deep learning can automatically learn feature engineering and achieve breakthroughs in many areas of computer vision.However,deep learning algorithms usually require a large amount of computing resources,making deep learning based object tracking algorithms difficult to implement fast tracking speeds despite their pretty accuracy.In order to improve the speed of deep learning based visual object tracking algorithm,we propose an improved object tracking algorithm based on the Siamese-network.First,we studied three representative algorithms deeply which are deep learning based MDNet,Siam FC and SINT.MDNet algorithm is used as the example to analyze the algorithm based on foreground and background classification.The Siam FC and SINT algorithms are used as the example to analyze the algorithm based on the Siamese network which tracking by similarity comparison.The training methods,tracking strategies and the performance,advantages and disadvantages of each algorithm are discussed in detail.According to the comparison,the object tracking algorithm based on the Siamese network can train the feature extraction network offline while the algorithm based on the foreground and the background classification is difficult to achieve real-time performance theoretically due to the necessary to perform back-propagation online to update the network parameters.In order to achieve higher-speed visual object tracking,we propose an improved fully convolutional siamese network object tracking algorithm combined with the above three algorithms.The algorithm uses two convolutional neural networks with identical structure and parameters that are identical and updated synchronously.The target branch receives the target image in frame t and the search branch receives the search image in frame t+1.After forward conduction,the target branch output feature map and the search branch output feature map are connected together then input into the convolution layers that used for the regression operation.The object external rectangle in frame t+1 is output directly at the last layer.The entire algorithm can be viewed as a regression function that predict the object in the current frame.The proposed neural networks does not obtain initialization parameters from any trained model of object detection or classification.Starting from a random initialization state,the neural network trained end-to-end on the Image Net Video 2015 dataset offline,and does not update the network during the tracking.The algorithm can perform tracking by only one forward conduction,and the tracking speed reaches 147 fps.Finally,performance evaluation and comparative analysis of the algorithm are performed on the most influential tracking benchmark VOT.The tracking overlap rate ranking chart of different scenes shows that the algorithm proposed in this paper mainly extracts the object appearance feature and movement feature.The average precision and robustness ranking of all the test sequences indicate that the object tracking algorithm full based on convolutional siamese network we proposed achieves state-of-art.