【作者】 徐斌；

【导师】 黎宁；

【作者基本信息】 南京航空航天大学 ， 信号与信息处理， 2018， 硕士

【摘要】 目标检测与跟踪是无人机视频巡视系统的重要研究课题。传统的目标检测算法因描述能力弱不能满足无人机视频巡视系统复杂环境下的目标检测需求。近年来,人工智能的发展为无人机视频巡视系统目标检测与跟踪的性能提升提供了新的突破口。本文以行人目标作为被测目标,引入深度学习对无人机视频巡视系统下的目标检测方法进行改进。同时,针对传统目标跟踪算法跟不稳定、实时性差的问题,在核相关滤波算法基础上,对无人机视频巡视系统下的目标跟踪改进方法进行研究。主要研究工作反映如下:1)无人机视频巡视系统下的目标检测改进方法研究。针对无人机移动速度快、拍摄场景复杂的情况,提出了显著性粗定位结合深度学习精定位的两步检测法。首先,根据行人目标多为直立状态的特点,对GBVS模型的Gabor方向、颜色特征进行针对性选择。通过加权合并,对GBVS通道各特征融合方式进行改进。通过显著性测量得到行人候选区域以实现目标粗定位。然后,采用迁移学习的设计思路,构建适合无人机视频巡视系统的轻量级卷积神经网络(CNN),用于目标精定位。所构建的轻量级CNN结构以深度可分离的卷积单元作为构建卷积结构的基础单位,将标准卷积操作分解为深度卷积和点卷积两部分,从而以更少的计算量实现与标准卷积操作近似的卷积效果。将基于显著性检测的目标粗定位候选区域,作为轻量级CNN的输入,通过对目标位置进行回归,实现目标的最终定位。实验结果表明,提出的目标检测改进方法因加入显著性目标粗定位,在无人机视频巡视系统下对小目标检测的精度优于常规CNN检测方法。2)无人机视频巡视系统下的目标跟踪改进方法研究。针对无人机视频巡视系统下行人目标存在较多的形变和尺度变化等情况,在核相关滤波算法(KCF)基础上,采用了融合HOG特征和颜色特征的跟踪改进方法。首先,引入对目标空间结构依赖性较低的颜色直方图特征,克服KCF目标形变检测鲁棒性较差缺陷。其次,训练独立的尺度滤波器对目标尺度变化进行估计,用于提高KCF对尺度变化目标检测的自适应性能。实验结果表明,改进后的核相关滤波跟踪算法对轻微形变和遮挡具有鲁棒性,能实现自适应尺度变化,达到实时跟踪的速度要求。更多还原

【Abstract】 Target detection and tracking are important research topic in UAV video navigation system.The traditional target detection algorithm can not meet the target detection requirements under the complex environment of UAV video navigation system limited by the poor descriptive ability.In recent years,the development of artificial intelligence provides a new breakthrough for the performance improvement of target detection and tracking in UAV video navigation system.Deep learning is introduced to improve the target detection performance under the UAV video navigation system.Meanwhile,to cope with the poor tracking speed of traditional target tracking algorithm under the UAV video navigation system,this paper studies the improvement method based on the kernelized correlation filtering algorithm.The main work of this paper includes the following aspects:1)Research on Improvement of Target Detection in UAV Video Inspection System.Considering the high moving speed and changeable characteristic of UAV video,a two-stage pedestrian detection method based on Graph-based Visual Saliency(GBVS)and Deep Learning is proposed.First of all,according to the characteristic that the pedestrian target is mostly upright,orientation and color features of the graph-based visual saliency model are modified.Also,the final saliency map is obtained by weighted summation instead of a direct sum of feature channels.Therefore pedestrian candidate region is obtained,realizing the rough detection of first stage.Then,based on tranfer learning method,we propose a lightweight Convolution Neural Network(CNN)for fine detection.The structrue is based on depthwise separable convolutions which is a form of factorized convolutions which factorize a standard convolution into a depthwise convolution and a 1×1 convolution called pointwise convolution to realize the similar effect with the standard convolution.The experiment results show that the improved detection method gain a better performance than other one-stage CNN based detection method in tiny target detection accuracy under the UAV system due to the coarse positioning introduced.2)Research on Improvement of Target Tracking method in UAV Video Inspection System.For the situation of pedestrian target existing more deformation and scale changes in UAV video,a modified tracking method combines HOG and color feature is adopted based on the KCF.First,the color histogram features that are less dependent on the target spatial structure are introduced in order to cope with the poor robustness of target deformation with KCF.Secondly,training an independent scale filter for robust scale estimation.The results show that the improved kernelized correlation filter tracking is robust to minor deformation and occlusion,and can achieve adaptive scale changing and meet the real-time requirements.更多还原