Research On Deep Siamese Network Based Object Tracking Algorithms

Object tracking is an important problem in computer vision and has applications in visual surveillance,robot navigation,virtual reality,autonomous driving,etc.The fully-convolutional Siamese network uses two deep convolutional network branches which are trained offline to solve the general similarity problem,and objects can be tracked by simply evaluating the network online.The fully-convolutional Siamese network achieves real-time frame rate in tracking and achieves competitive results in current tracking benchmarks.Although the current method based on the fully-convolutional Siamese network architecture has achieved remarkable results in tracking field,it still fails to track in the environment with complex tracking background,drastic changes in illumination and long-term occlusion.In order to design a robust object tracking system,we have done the following work:(1)We propose to introduce the channel-wise attention mechanism to help the network learn to select the most informative and discriminative channels in feature map.We observe that in the feature map of the Siamese network,not all channels are helpful for tracking.To improve the feature's discriminability in the network,we introduce channel-wise mechanism into Siamese network,which assigns higher weights to the more discriminative feature channels.(2)We propose a novel multi-scale feature fusion method which uses a top-down structure with horizontal connections to construct advanced semantic feature maps at multiple scales.In the original Siamese network architecture,in order to increase the representation capability feature map,the last output feature maps of two branches serviced as the input of the last connection operator.Although deeper feature maps have strong abstract representation ability,spatial information is lost due to down-sampling.On the other hand,feature maps in shallow layers may preserve spatial information well despite of their insufficient representation power.We propose in this paper a feature fusion strategy in wich deeper and more semantics features with lower resolution are upsampled and fused with high-resolution features from precedent layers.(3)We propose to introduce RAFT optical flow module to help Siamese network predict the movement trend of object.We fuse the image optical flow information obtained through the RAFT module and the image feature information obtained through the Siamese network,and finally obtain the image feature fused with the optical flow information.Based on this image feature,the tracking model can better predict the movement trend of object.(4)We make comparative experiments with other object tracking algorithms based on the accuracy and success rate of the object tracking benchmark dataset.The experimental results show that our proposed object tracking algorithm can effectively improve the robustness and success rate of tracking.