Research On Superpixel-based UAV Machine Vision Algorithm

UAVs have become an important filed of rapid development in today's world,and its applications cover many fields of military and civilian use.And the machine vision algorithm based on drone has become a research hotspot.Among them,the research of target saliency detection and video target tracking algorithm has important application value.Superpixel,as an important image feature description,can reduce the complexity of subsequent visual processing to a large extent due to its rich image redundancy information.Aiming at the problems of complex background,target scale change and position drift of UAV aerial image,this thesis combines the research on saliency detection and target tracking in UAV aerial videos with superpixel,and focusing on the improvement of superpixel segmentation algorithm.Aiming at the requirements of image saliency detection and target tracking algorithm on image information features,two improved algorithms are proposed to improve the accuracy,quality and speed of superpixel segmentation algorithm respectively.At the same time,combined with the improvement of traditional target saliency detection algorithm and target tracking algorithm,the accuracy of UAV aerial video target saliency detection and the robustness of target tracking algorithm are improved.In addition,this thesis also proposes a super-resolution reconstruction algorithm based on sparse representation as a restoration technology of UAV images.The specific work content is as follows:(1)A UAV image restoration technique based on sparse representation for super-resolution reconstruction is proposed.The low resolution of UAV image is caused by the limitation of acquisition,transmission and storage of image signal,weather and other environmental factors.In this thesis,the sparse representation theory is used to the reconstruction of process of low-resolution images to achieve the image restoration effect of UAV with high speed and quality.(2)A target saliency detection algorithm based on SLIC improvement and multi-scale superpixel is proposed.Firstly,the basic principle and common algorithms of superpixel generation have been studied and analyzed.Combined with the characteristics of aerial image of UAV,the measurement method and feature usage of pixel distance in SLIC algorithm are improved,and the initial segmentation constraint framework is added.Then,the improved SLIC algorithm is used to generate multi-scale superpixels,combined with Bayesian framework for target saliency detection.Finally,experiments show that the improved superpixel algorithm and the use of multi-scale superpixels can significantly improve the detection accuracy of target saliency detection in UAV aerial images.(3)A target tracking algorithm based on non-iterative clustering superpixel and keypoint structure is proposed.Aiming at the real-time and robustness requirements of aerial video target tracking,the process of clustering in SLIC algorithm is improved by discarding iteration,improving distance calculation efficiency and forcing connection at the beginning of the original algorithm.The experimental results show that the improved algorithm not only greatly improves the speed of the algorithm,but also improves the boundary compliance of the superpixel.By combining the superpixel generated by the improved algorithm with the image key points,a superpixel-keypoint structure is constructed for the feature description of the target tracker.At the same time,combined with the new appearance model,the update strategy is improved to avoid model drift.Finally,the experiment proves that the tracker combined with superpixel and keypoint information is more robust to the target tracking of UAV aerial video.