Computer Vision for Wide Area Aerial Imaging

Wide area aerial imagery presents unique challenges for automated analysis. It is captured by an array of cameras, at low frame rates, but in large format, resulting in a high data rate. Traditional approaches for well-known computer vision problems of mosaicking, video stabilization, object tracking, and activity recognition cannot be simply applied with an expectation of success. In this book, we take a fresh perspective at the fundamental computer vision problems as applied to this imagery, and offer practical solutions with an eye to computational efficiency.We begin with the camera array mosaicking problem, and present an algorithm that estimates high quality seamless mosaics in the presence of camera distortion, with only a small amount of overlap between cameras, and with varying camera exposure levels across the array. Once mosaicked, it is desirable to remove the motion of the camera array, stabilizing the video. We discuss approaches to this, which minimize the amount of drift and jitter. The mosaicked and stabilized video reveals the motion of hundreds, possibly thousands, of objects. The problem of tracking these objects at low frame rates, and quite often only in grayscale, is then addressed. We present scalable solutions, which achieve state of the art performance. Finally, the extracted trajectories open up possibilities for activity recognition. We consider recognizing interesting trajectory-based activities and motion patterns.Algorithms presented in the book are evaluated on publicly available datasets, and links to software implementations are provided where applicable, allowing the interested reader to get a deeper insight into this active research area.