得益于无人机平台及其探测载荷的技术进步，无人机获取数据更加便捷、丰富，基于无人机航拍图像的目标检测在军警海事、监控安防等领域均有广泛应用。与固定监控平台相比，无人机载荷拍摄目标的多尺度效应更加剧烈、地物场景更加复杂；同时受光照、拍摄角度、成像参数影响，同一目标的外观呈现多样性，为基于无人机航拍图像的目标检测带来挑战。综合考虑无人机观测视野广、飞行自由度大的特点，本文基于卷积神经网络，开展无人机航拍图像目标检测研究，主要成果及创新点如下： (1)针对无人机观测视野广导致目标定位不准确的问题，提出基于有效感受野的精准候选区域提取方法。通过引入卷积神经网络有效感受野，设计基于有效感受野的候选区域提取技术，增加目标特征尺度与候选区域的匹配度。同时，依据有效感受野构建卷积特征金字塔，并为卷积特征金字塔中每一层分配对应尺度的候选窗口，进而适应不同大小的目标。在公开数据集进行的测试结果表明，相较于目前主流方法，平均准确率得以改善。 (2)针对无人机平台运动导致航拍图像同一目标出现外观多样性的问题，提出基于相邻帧图像信息的特征一致性表征方法。通过孪生网络引入相邻帧图像信息，结合相邻帧图像特征金字塔，构建互相关损失函数，利用特征匹配损失与中心位置损失强化航拍图像中同一目标特征提取相似性，显著提升无人机航拍图像目标特征一致性。通过公开数据集测试，相较于目前主流方法，平均准确率均有提升。 (3)针对无人机自由度大导致多尺度目标特征鲁棒性差的问题，提出基于尺度上下文的空间特征表征方法。通过引入尺度空间上下文关系，设计由顶至底与由底至顶相结合的卷积特征金字塔生成方式，利用上采样与池化过程建立卷积特征金字塔各层间的联系，平衡各层特征信息量，由此提高多尺度目标检测准确率。通过公开数据集测试，相较于目前主流方法，平均准确率均有提升。 (4)针对无人机航拍视频目标快速检测需求，提出基于光流的快速特征金字塔生成技术，提高了目标检测效率。通过引入时间维度信息，利用基于卷积神经网络的快速光流场提取技术，结合运动目标尺度信息，设计光流指导下的快速特征金字塔生成框架，克服了传统卷积特征金字塔计算过程复杂的问题。同时，结合前三点创新方法，构建了一种鲁棒、快速的无人机航拍视频目标检测框架，提高了无人机航拍视频目标检测的效率。通过公开数据集测试，相较于目前主流方法，在保证准确率的情况下，检测速度提升了近3倍。 关键词：无人机航拍图像；特征一致性；尺度上下文；光流

 With the rapid developments of Unmanned Aerial Vehicle (UAV) platforms and loads, the object detection from UAV aerial images are widely used in surveillance and security applications. Compared with fixed camera, objects from UAV aerial images have more serious multi-scale effects and backgrounds are more complex. Meanwhile, due to the illumination variations, different shooting angles and various imaging parameters, a certain object shows diverse appearances. These differences make the object detection task from the drone view more difficult. Considering the large observation field and high freedom of UAVs, this thesis uses convolutional neural network (CNN) to detect various objects from the UAV aerial images. The main contributions and innovative points are as following: (1)Due to the large observation field of UAVs, the accuracy of the object location is low. In order to overcome this problem, the candidate window extraction methods based on effective receptive field is proposed. The scale matching between candidate regions and object features is the key requirement for locating objects accurately in UAV aerial images. By introducing the effective receptive field, the candidate region extraction method based on effective receptive field is designed to increase the scale matching ratio between object features and the area of candidate regions. At the same time, based on the effective receptive field, the convolutional feature pyramid is constructed and candidate windows of corresponding scales are allocated to each convolutional feature layer. In this way, the object detection accuracy in UAV aerial images is improved. The proposed method has been tested on the public dataset and compared with advanced object detection methods, the mean average precision increases. (2)The UAV platforms are greatly affected by the external environment, which leads to various appearances of objects in UAV aerial images. The feature consistency representation method based on adjacent frame information is proposed. The feature consistency of moving objects is important for stable detection, for enhancing the feature consistency of moving object in UAV aerial image, the Siamese network, which involves the information of adjacent frames, and feature pyramid of neighboring frames are related. Then, by designing correlation loss, which consist of the center value loss and coordinate loss, the feature consistency could be guaranteed. In this way, the detection accuracy of objects in UAV aerial images could be improved. The proposed method has been tested on public datasets. Compared with advanced object detection methods, the mean average precision increases. (3)UAVs have the high freedom, which leads less robustness of multi-scale feature representation. For solving this issue, the multi-scale space feature representation based on scale context is proposed. The information imbalance between objects with different scales is the prime factor. By introducing the scale context information, a top down and bottom up feature pyramid network is established, which uses convolutional and pooling procedures to build the relationship between each feature layer. In this way, the information of multi-scale feature representation could be balanced and the detection accuracy for objects with different scales is improved. The proposed method has been evaluated on public datasets. Compared with advanced object detection methods, the mean average precision increases. (4)Aiming at the requirement of the fast object detection in UAV aerial videos, the fast feature pyramid generation method based on optical flow is proposed. The generation of convolutional feature pyramid is complex and needs much more computational resources, which affects the efficiency of the object detection. By introducing the fast optical flow generation network, the fast feature pyramid based on the optical flow is designed. Then, considering the scale information, the optical flow features sand convolutional features are combined to detect objects in UAV aerial videos with fast speed. Last but not the least, incorporate with the above three parts, a robust and fast object detection architecture for UAV aerial videos is established, which provides the speed guarantee for various UAV applications. Key Words: UAV aerial images; feature consistency; scale context; optical flow