Though the deep learning methods have achieved effective moving vehicle detection in satellite videos, there is a non-negligible premise that these methods require lots of object-level annotations for hundreds of small and blurry vehicles in the vast observation scene. These annotations can be quite labor-intensive and time-consuming. To address this problem, this paper is committed to realizing the vehicle detection based on point-level annotations, and a motion-aligned and hardness-aware dynamic update network is proposed, which consists of the basic detector, motion-aligned initialization method and online pseudo label update scheme. Specifically, the high-quality pseudo bounding boxes are initialized by revising the Gaussian mixture model to fully exploit the motion information in the video sequence and the location information from the point annotations. Then, the pseudo bounding boxes are utilized as the supervision for the basic detector. During the training phase, an online label refinement scheme is designed to refine the pseudo bounding box continuously, and the confidence-aware loss function is defined to adjust the example weight dynamically according to its learning hardness. Extensive experiments on the Jilin-1 and SkySat satellite video datasets show that our method achieves the comparative performance compared with fully-supervised learning methods.