航拍红外视频帧间图像配准是航拍红外视频运动目标检测的一个重要步骤，用来将前后两帧图像变换到同一坐标系下，消除相机运动带来的帧间图像坐标偏差。由于无人机拍摄范围广和红外摄像头通过热辐射成像等原因，航拍红外视频具有背景复杂多变、图像模糊的特点，这给航拍红外视频的相关研究带来了较大的挑战。另外，为满足实时、准确地从航拍红外视频中提取出运动目标，帧间图像配准算法应具有较高的配准精度和效率，进一步增加了配准算法的难度。配准评估是算法选择的依据，采用统一标准对算法性能进行评价，能够有效衡量算法的有效性和准确性。本文针对无人机平台下的航拍红外视频帧间图像配准方法和配准效果评估技术两部分关键内容，展开深入研究。论文主要工作如下： （1）针对无人机航拍红外场景较模糊的特点，提出了一种适用于航拍红外视频的帧间图像快速配准算法。算法首先提出改进的SUSAN（Smallest Univalue Segment Assimilating Nucleus）角点检测算法保证在较为模糊的红外图像中仍能准确定位特征点；接着设计了角点聚类和自适应双阈值相结合的运动目标区域角点去除算法，以解决在场景简单或运动目标数量较多时角点过于集中在运动目标区域给配准带来的干扰；然后对剩余的角点利用HOG（Histogram of Oriented Gradients）进行描述并根据向量间的KL（Kullback-Leibler）距离求取特征匹配对；最终采用RANSAC（RANdom Sample Consensus）算法估计帧间变换模型参数。实验证明，该算法能够在不同背景复杂度和不同运动目标个数的场景中对相邻帧进行准确的运动估计，平均耗时为62ms，达到了实时的处理速度。 （2）提出两个面向运动目标检测的航拍视频帧间图像配准评估准则。对运动目标检测来说，图像配准是为求残差服务的，其注重的是配准后像素间的位置对应关系的精度。为刻画图像配准的精度，我们设计了位置误差比例（Position Error Rate，PER）准则和正确匹配率（Correct Match Rate，CMR）准则。PER准则计算当前帧各点根据变换矩阵得到的在参考帧中的投影位置和其基准数据之间的距离小于1个像素点和0.5个像素点的点的比例，通过对每次配准得到的PER值进行平均可以统计算法在整个视频中的结果，以反映准确配准的像素的比例。CMR准则根据每次配准的PER值是否大于一个固定阈值判断配准是否正确，进而计算在一个视频中匹配正确的帧的比例，以反映正确匹配的图像帧的比例。为获取两个准则的基准数据，采用了稍微复杂的、环境适应性强的改进块匹配算法与人工校验相结合的方法产生基准数据，利用多个标准选择可靠的块匹配对计算变换参数，求取点的对应位置坐标，生成的基准数据具有亚像素级精度，对配准得到的残差图进行人工校验可以保证基准数据的准确性。基于PER准则和CMR准则，我们对SUSAN、FAST（Features from Accelerated Segment Test）、DoG（Difference of Gaussian）、Harris、SUSAN_M五种特征检测算法和灰度、HOG、SIFT（Scale Invariant Feature Transform）、BRIEF（Binary Robust Independent Elementary Features）四种特征描述算法进行了评估。实验证明，DoG类配准算法...
英文摘要
Image registration between adjacent frames in aerial infrared videos is an important step for moving object detection and used to transform adjacent frames to the same coordinates in order to eliminate the influence of camera motion. Since UAV (Unmanned Aerial Vehicle) has a wide shooting scope and infrared camera photographs the target scene based on objects’ thermal radiation, aerial infrared videos usually have complex and ever-changing backgrounds and are blurry with no color cue, which make the relative research become a tough challenge. Besides, to detect moving objects accurately in real time, image registration methods between adjacent frames are required to possess high registration accuracy and efficiency, which further increases the difficulty of the registration algorithm. Registration evaluation is the basis for the selection of different algorithms. Through using the same standard to evaluate the algorithms’ performance, we can measure their effectiveness and accuracy. In this paper, we mainly focus on two key aspects, image registration methods in aerial infrared videos and their performance evaluation technologies. The main works are summarized as follows: (1) According to the characteristics of UAV aerial infrared videos, a fast image registration algorithm between adjacent frames is proposed. Firstly, the algorithm improves the original SUSAN (Smallest Univalue Segment Assimilating Nucleus) detector so that corners are detected stably and accurately in blurry infrared images. Then, on-moving-object corner rejection step consisting of corner clustering and adaptive dual threshold is designed to solve the problem of corners centralizing on moving object region under the flat background or with several moving objects. After that, HOG (Histogram of Oriented Gradients) is utilized to form the descriptors of the remaining corners and KL (Kullback-Leibler) distance is employed to measure the similarity and obtain correspondances of corners. Finally, RANSAC (RANdom SAmple Consensus) is performed to estimate the parameters of the transformation model with the correspondences. Experimental results show that this algorithm can register adjacent frames accurately under the backgrounds with different complexities and several moving objects. The average time cost is 62 ms, which reaches real-time processing speed. (2) Two criteria are proposed to evaluate the performance of image registration algorithms for object detection in aerial videos. Imag...
修改评论