固定翼无人机地面车辆目标自动检测与跟踪

发布时间：2018-01-02 07:15

本文关键词：固定翼无人机地面车辆目标自动检测与跟踪　出处：《中国科学院国家空间科学中心》2017年博士论文　论文类型：学位论文

【摘要】：目标检测与跟踪是计算机视觉方向的重要研究领域,而无人机地面目标自动检测与跟踪在侦查、预警等领域均具有重要作用及战略需求。在这些应用场景中,地面目标在视场中一般尺寸较小,现有的检测算法对于小目标而言仍然比较乏力。无人机在进行目标跟踪的初始阶段,目标由于距离较远,在视场中尺寸过小(目标像素尺寸小于10×10),仅有边缘信息和少量甚至缺失的纹理信息,给目标检测与跟踪带来极大挑战。此外,小型固定翼无人机地面目标检测与跟踪还面临着背景环境复杂、视场变化剧烈、干扰物多等诸多难题。本论文主要研究小型固定翼无人机对车辆小目标的自动检测与跟踪方法。本文首先介绍了目前国内外的无人机目标检测与跟踪系统的研究和发展现状,跟紧趋势找出不足;同时算法层面上,分析了国内外目标检测、跟踪算法的发展趋势,找寻算法的难点和待突破的地方,从而为本篇论文的研究内容奠定方向。论文提出了一种基于目标特性的微小目标检测方法,通过一种融合了梯度响应的显著性检测来提取候选目标,同时引入局部敏感直方图特征,有效解决室外环境光照条件复杂变化的难题。将视场中相应像素位置的候选目标尺寸与固定翼无人机的位姿信息耦合起来,可以有效排除伪目标。对HOG特征进行傅里叶分析,生成具有旋转不变特性的FHOG特征,联合FHOG与LBP这两种旋转不变特征,充分利用小目标宝贵的边缘特征和极少量的纹理特征,将目标的检测问题转换为对候选目标的分类确认问题,同时大大减小了逐像素遍历搜索带来的巨大运算量,实现算法的实时性。训练时加入难样本挖掘重训练,将目标检测的精度提升10.23%。本文研究比较了最新的卷积神经网络在小目标检测中的应用,并分析其不足和原因,在负样本选取和难样本重训练上进行了优化和改进,有效解决了复杂背景下各种干扰物对小目标检测的影响,优化后的Faster R-CNN较原有默认参数在VEDAI数据库小目标的检测精度上高出40%多。另外提出一种基于目标特性的神经网络检测方法,对VEDAI航拍车辆数据库精度达85.34%。在检测到目标后,采用基于KCF的时空上下文目标跟踪方法实现对小目标的实时跟踪,固定翼无人机采用嵌入式ARM平台,基于嵌入式linux进行算法移植,实现对地面车辆目标的实时自动检测和跟踪。通过固定翼无人机在复杂背景环境下航拍模飞进行目标检测和跟踪试验,并且在公开数据集上比较验证算法。论文在多种场景下进行固定翼无人机模飞,尽可能多地模拟各种环境,验证算法的鲁棒性,同时在公开数据集上与目前state-of-the-art算法横向比较。通过试验分析验证了基于目标特性的微小目标检测算法、基于FHOG+LBP联合旋转不变特征的目标检测算法以及适用于小目标检测的神经网络方法,以及基于KCF的时空上下文目标跟踪方法,最后实现了相关算法的嵌入式平台移植。
[Abstract]:Target detection and tracking is an important research direction of computer vision, and UAV automatic target detection and tracking in the investigation, has an important role and strategic demand early warning and other fields. In these scenarios, the ground target in the generally smaller field of view, the existing detection algorithm for small targets is still relatively weak. In the initial stage of the UAV target tracking target, because of distance, the size is too small in the field of view (the target pixel size is less than 10 * 10), only little or lack of edge information and texture information, bring great challenge to the target detection and tracking. In addition, small fixed wing UAV Ground Target Detection and tracking is facing the complex background environment, multiple field change, and many other problems. This paper mainly studies the small fixed wing UAV vehicle target automatic detection and with Trace method. This paper introduces the research and development status of UAV target detection and tracking system at home and abroad at present, and find out the trend of tight; at the same time the algorithm level, analyzes the domestic and international development trend of target detection, tracking algorithm, algorithm and difficult to place to find a breakthrough, so this thesis the research content lay direction. This paper presents a small target detection method based on target characteristic, was detected by a combination of the gradient response to extract candidate targets, while the introduction of local sensitive histogram, effectively solve the problem of complex outdoor environment illumination condition changes. The candidate target size and pixel position corresponding fixed wing the field of UAV pose information coupling, can effectively eliminate false targets. Fourier analysis of HOG features, generate rotation invariant features of FHOG. Sign, combined with FHOG and LBP these two kinds of rotation invariant features, make full use of edge features of small targets valuable and very little texture features, converting the detection target to confirm the classification of candidate targets, and greatly reduces the pixel traversal search brings huge computational complexity, real-time implementation of the algorithm training. When added to sample mining heavy training application will enhance the precision of detection of 10.23%. in this study compared the convolutional neural network is the latest in small target detection, and analyze the problems and causes, in the negative sample selection and sample weight training to optimized and improved, effectively solves the influence of various complex interference under the background of small target detection, the optimized Faster R-CNN detection accuracy than the original default parameters in the VEDAI database of small targets on more than 40% higher. In addition we propose a target characteristics based on God The network detection method of VEDAI aerial vehicle database to the target in the detection accuracy of 85.34%., the temporal spatial context target KCF tracking method to achieve real-time tracking of small target based on fixed wing UAV Based on embedded ARM platform based on Embedded Linux algorithm transplantation, real-time automatic detection and tracking of ground vehicles. Through the fixed wing UAV in complex background aerial flight simulation for target detection and tracking test and verification algorithm. In the open data set in a variety of scenarios for fixed wing UAV flight simulation, as much as possible to simulate all kinds of environment, the robustness of the algorithm is verified, and in the public data set with the current state-of-the-art algorithm horizontal comparison. Through experimental analysis verified the small target detection algorithm based on the characteristics of target, FHOG+LBP combined with rotation invariant features based on The target detection algorithm and the neural network method suitable for small target detection, as well as the spatio-temporal context tracking method based on KCF, are finally implemented. Finally, the embedded platform transplantation of related algorithm is realized.

【学位授予单位】：中国科学院国家空间科学中心
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41;V279

【相似文献】