Abstract: With the rapid development of unmanned aerial vehicles, character verification and behavior recognition based on pedestrian dynamic shadow biometrics in the top-view state has become an important development direction, and obtaining pedestrian dynamic shadow biometrics from low-altitude UAV platforms has become a more challenging research hotspot. Most of the existing shadow extraction methods are based on fixed cameras, which are not suitable for motion platforms. In this paper, it combines machine learning and graph cutting theory to propose a new dynamic human shadow detection method for low-altitude UAV platforms. First, two independent views for co-training are constructed according to pixel features and regional features. It uses SVM as the classifier to semi-automatically extract shadow biometric features by machine learning methods. Then, the data items of the minimum energy equation are constructed based on the above motion results and the constraint terms of the energy equation are constructed according to the gradient features of the image. And it uses the graph cutting theory to optimize the above extraction results. The experimental results show that the proposed method has a better effect than the pure collaborative training method, and can further optimize the quality of shadow biometrics obtained under the low-altitude UAV platform.

Key words: unmanned aerial vehicle, collaborative training, graph cutting, shadow detection

摘要： 随着无人飞行器的迅猛发展，根据顶视状态下的行人动态阴影生物特征进行人物验证和行为识别成为一个重要的发展方向，而从低空无人飞行器平台获取行人动态阴影生物特征成为一个更具有挑战性的研究热点。现有的阴影提取方法大多基于固定摄像头，并不适用于运动平台。联合机器学习和图切割理论，提出了一种新的针对低空无人飞行器平台的动态人影检测方法。根据像素特征和区域特征构建协同训练的两个独立视图，以SVM为分类器，采用机器学习的方法对阴影生物特征进行半自动提取，根据上述运动结果构建最小能量方程的数据项，根据图像的梯度特征构建能量方程的约束项，运用图切割理论对上述提取结果进行优化。实验结果表明，所提出的方法比单纯的协同训练方法具有更好的效果，可进一步优化低空无人飞行器平台下所获取的阴影生物特征质量。

关键词: 无人飞行器, 协同训练, 图切割, 阴影检测