Covariance Descriptors on Moving Regions for Human Detection in Very Complex Outdoor Scenes