【作者】 张恒；

【导师】 于起峰；

【作者基本信息】 国防科学技术大学 ， 航空宇航科学与技术， 2008， 博士

【摘要】 无人机因其特有优势而一直备受各国军事专家的青睐。特别是在最近几次局部战争中,无人机更因其赫赫战功,而成为各国军事发展的重要方向之一。论文以无人机平台对运动目标检测跟踪和无人机辅助着落导航系统为背景,研究了图像角点特征提取、像机自运动消除、复杂背景下运动目标跟踪、目标尺度方向自适应跟踪、小目标实时高精度跟踪和无人机辅助着陆系统等方面的相关算法,系统总结了作者在攻读博士学位期间所做的研究和取得的成果。论文创新点如下:(1)传统梯度算子抗噪性不好,而且在梯度方向计算准确性上存在一定误差,论文在分析Gabor小波的基础上,提出了一种新梯度算子。新算子在抗噪性和梯度方向计算的准确性上优于传统梯度算子。针对图像角点特征提取问题,本文提出一种基于Gabor梯度的角点检测算法。新算法在定位精度、噪声抑制等方面比经典算法有一定的提高。(2)对于像机自身存在运动情况下对运动目标进行检测跟踪问题,常用的一类算法是先消除像机自身运动,然后再对目标进行检测跟踪。论文针对像机运动消除问题,提出了一种基于Gabor特征描述的像机自运动消除方法。实验结果表明,新算法可以有效地消除像机自身运动,检测出和像机存在相对运动的目标。(3)目标在复杂背景中运动时,变化的背景会在一定程度上影响跟踪算法的稳定性和可靠性,所以应尽量抑制背景。对于目标跟踪而言,目标中和背景差异大的区域对于跟踪的有效性、稳定性贡献较大。根据以上原则,论文提出了一种目标区域差异性权值计算方法。实验结果表明,权值计算方法有效地抑制了背景,权值计算结果符合人类直观感受。在获取目标差异性权值的基础上,本文提出了一种基于目标差异性加权的最小二乘影像匹配运动目标跟踪算法。仿真实验分析和自然图像跟踪结果表明,对于复杂背景下运动目标进行跟踪,新方法比传统方法更为有效准确。论文将差异性权值与Mean Shift跟踪算法结合起来,提出了一种基于目标差异性加权的Mean Shift运动目标跟踪算法,并对算法的收敛性及收敛条件进行了讨论。实验结果表明,改进的Mean Shift算法比传统跟踪算法在目标跟踪稳定性、可靠性等方面有一定提高。(4)对于目标跟踪问题,目标的初始特征建立或目标的描述区域是否准确是整个跟踪过程中至关重要的一环。针对该问题,本文在分析尺度空间理论的基础上,提出了一种目标最佳椭圆描述区域计算方法,该方法为准确描述目标区域特性奠定了基础。针对图像上同时存在旋转缩放变化的运动目标进行自适应跟踪问题,在分析Mean Shift跟踪算法和尺度空间理论的基础上,本文提出了一种尺度方向自适应Mean Shift跟踪算法。实验表明,新算法比现有的Mean Shift改进算法更为有效,即可以准确获得目标的位置、大小、方向信息。(5)针对小目标实时高精度检测跟踪问题,提出一种基于正负正则化LOG算子的尺度自适应小目标实时高精度检测跟踪方法。实验证明,新方法在实时性、尺度自适应性、检测准确性以及抗噪声性上均有较好的表现。(6)无人机安全进场着陆过程中,其引导控制尤为关键。目前GPS导航因其精度高、使用简单而普遍使用,但信号易受干扰,特别是战争时期不能获得准确的GPS信息。论文提出了一种基于摄影测量的无人机辅助着陆导航系统。实验表明,该系统可以实时高精度地获得无人机相对跑道的相对位置,同时可以实时保存无人机着陆过程图像,为事后分析提供有效数据。更多还原

【Abstract】 Unmanned Aerial Vehicle (UAV) is extra favored by the military experts for its own dominance. Specially,UAV has already been one of the very important military development fields for its success in the last brushfires. Moving target tracking on the platform of UAV and the assistant landing system of UAV are taken as the backgrounds of this dissertation. The image corner detection, camera ego motion compensation, moving target tracking in the cluttering background, the scale and rotation adaptive tracking of the target, puny target real-time and high-accurate detection, the assistant landing system and other relative technologies are investigated in this dissertation. The main contents and achievements of this dissertation are as follows:(1) There is definite error to calculate the gradient orientation by traditional gradient operators. Based on the analysis of the gabor wavelet, a new gradient calculating operator is proposed. The new operator can obtain the orientation of the gradient more accurately, and restrains the noise more effectively than the traditional gradient operators. And a new corner detecting algorithm based on gabor gradient operator is proposed. Experimental results with some synthetic and real images show that this new algorithm detects the corner more efficiently, locates the corner more accurately, and restrains the noise more effectively than the classical algorithms.(2) For the tracking of the moving target on the moving platform, one commonly used algorithm is to eliminat the camera ego motion firstly, and then track the moving target. An camera ego motion compensation algorithm based on the gabor feature descriptions is proposed. Experimental results show that the new method can effectively compensate the camera ego motion and detect the moving targets.(3) When a target moves in the cluttering backgrounds, the variational backgrounds can reduce the stability and reliability of the tracking algorithms. So the backgrouds must be restrained. For the heterogeneous target, the target areas in the image are different for the target’s ununiformity. If some areas in the target are more different from the backgrounds, it shows that the diversity of these areas is stronger and these areas are more helpful for target tracking. Based on the above-mentioned principle, a diversity weight calculation algorithm is proposed. Experimental results show that the new algorithm can restrain the backgrounds effectively, and that the weights accord with the human intuitionistic recept. Based on the diversity weights, a diversity weighted least square image matching method is proposed. Simulation experimental analysis and real image tracking results show that the new method can track moving target in the cluttering backgrounds more effectively and accurately than the traditional tracking methods. A new diversity weighted Mean Shift tracking method (DWMS) is proposed. Its convergence condition is discussed. And its convergence is proved. Experimental results show that DWMS can effectively track moving targets in the cluttering background too.(4) It is one of the most important problem for target tracking whether the initial features and the describing areas of the target are accurate. Based on the scale-space theory, a new method of the optimal ellipse describing area of the target is proposed. The new method provide a foundation for the feature analysis of the target. It is one of the most difficult research fields of computer vision to track target which takes the change of rotation and scales at the same time. Based on the analysis of the scale-space theory , the current Mean Shift algorithms and the optimal ellipse describing area of the target, a scale and rotation adaptive Mean Shift tracking algorithm is proposed. Experimental results show that the new Mean Shift tracking algorithm can more effectively track the target which takes the change of both rotation and scale than the current mean shift algorithms. At the same time, the new algorithm can supply the size and angle information of the target.(5) To detect and track puny target high-accurately and in real-time, a new scale adaptive little target detecting and tracking method based on the positive and negative LOG operator is proposed. Experimental results demonstrate that the new method has great capability in real time detection, precise detection and noise restraint.(6) In the landing process of UAV, navigation control is very important. At present, the GPS navigation is usually used because it is very accurate and easy. But its signal can be easily disturbed during the war. The vision-based navigation system becomes one of the most important research fields for its superiority. A assistant landing system based on the photogrammetry is proposed. Experimental results show that the system can obtain the accurate location of UAV according to the runway, and save the images of the whole landing process of UAV synchronously. which can provide effective information for post analysis.更多还原