Object Tracking Based On Multi-scale Feature Fusion And Squeeze-excitation Model

The main task of object tracking is to detect the positions of the targets in each frame from a continuous video sequence.With the growing understanding of the field of computer vision,object tracking has been widely used and developed.At present,there are a lot of tracking algorithms has been achieved on object tracking.There are still a lot of challenges to be solved due to occlusion,background clutter,appearance deformation and illumination variation.In this paper,we focus on the challenges mentioned above,and the main contributions of our work are as follows:1.A method based on the features of distribution field is proposed.When using traditional features to smooth the target image,it's easy to cause the loss of pixel information.This work uses the descriptor of distribution field as features.The work introduces the scale filter to solve the problem of scale variation.The work uses the adaptive hedged algorithms to update the weights of template.The representative algorithms are selected to compare with the methods in this chapter.Experiments show that our method is robust when encounter with the problem of illumination changes,occlusion and fast-moving.2.A method based on fully–convolutional Siamese network with squeeze-and-excitation blocks is proposed.This work uses a fully–convolutional Siamese networks to learn similarity between targets for object tracking,and improves the network by squeeze-andexcitation blocks.By introducing the squeeze-and-excitation blocks,the network can learn the importance of each feature channels online and enhance the useful feature channels.The work uses two ways to enhance the representation ability of network.Compared with the traditional tracking algorithms,our methods can adapt to the change of target and background,and achieve real-time tracking with good performance.3.A method based on deep-residual network with squeeze-and-excitation blocks is proposed.This work expands the training dataset with combined samples to solve the problem that the dataset of object tracking is too small to train the deep convolution neural network.This work introduces squeeze-and-excitation blocks into the deep-residual network to constructa two-class network model.The network is trained by extended dataset.Compared with representative algorithm,the experimental results show that our method can deal with a variety of complex tracking scenarios and has a certain robustness.