Research On Robust Tracking And Super-resolution Reconstruction In Complex Scenes

Intelligence based on deep learning leads innovation and drives the rapid development of the industries.The modern warfare oriented to high intelligence has become an effective means to safeguard national security and ensure the safety of people's lives and properties.Accurate tracking and precise strike for target with autonomous perception and judgment are of great significance to the complex battlefield environment.The target trajectory predicted exactly and capturing key information of the target have great application value for the effective maintenance of public safety.All phases of target tracking and image super-resolution in deep learning are studied systematically based on the theories of deep learning,correlation filtering and sparse representation.This thesis focuses on inaccurate feature expression of the target and poor robustness of tracker in the complex scenes,and here by as a chance probe into the restoration for high frequency and details of the degraded image and carry out in-depth research.Effective target tracking and image super-resolution methods are proposed.The main contents and contributions in this thesis are summarized as follows.1.A Robust Tracker Integrating Particle Filter into Correlation Filter Framework.Aiming at the irrationality of hypothesis that the target does not rotate based on correlation filter method,which leads to poor adaptability for trackers to large deformation and rotation,we explore the principle of particle filter.We conduct sampling of particles based on the theory combined with affine transformation.The optimal rotation factor is obtained by similarity between the template and candidates,which breaks through the problem of low robustness caused by target rotation.In addition,the joint adaptive update strategy improves the accuracy of the tracking,and then proposes the method integrating particle filter into correlation filter framework.The experimental results have shown the effectiveness of tracking technologies.2.Convolutional Residual Learning with Sparse Robust Samples and Multi-feature Fusion for Object Tracking.Aiming at the problem of poor quality of training sample set and low accuracy of feature expression caused by fast motion,illumination and severe occlusion.The characteristic that candidates are sparsely represented by templates is fully utilized on the basis of sparse representation theory and we combine occlusion discrimination and peak detection to reduce effect of negative samples such as background.Furthermore,the multi-feature fusion strategy relied on residual learning enhances the adaptability of the target in complex scenes.A target tracking method is proposed with constructing sparse robust samples and multi-feature fusion.And the experimental results have shown the effectiveness about it.3.Stepwise image super-resolution using stacked GAN network.In view of the problem that it is difficult to restore high frequency detail and good visual perception from low-resolution images with lacking seriously detail.The predictive capability of single-network supervision learning is observed in the reconstruction process of high super-resolution image,and the essence of texture smoothness aboutL₂ loss is exploited.Therefore,we make full use of the confrontational idea of generative and discriminative network,and propose a stepwise super-resolution algorithm of stacked GAN network.GAN1 performs initial restoration based on the edge and visible details,and then GAN2 corrects the distortion of the reconstructed image,and supplements the details further.In addition,the Log-Cosh loss improves deficiency of ₁L and L₂loss function.The reconstructed high-resolution image without distortion has good visual perception.The experimental results have verified the effectiveness of the method.