With the development of people's security awareness, video surveillance systems have been widely used, resulting in an increasing amount of surveillance video data. Surveillance videos are usually fixed and continuous for a long time, which leads to a single surveillance video background, motion segments and static segments cross-exist. However, people usually only focus on the motion segments. There is an urgent need to quickly extract motion segments from massive surveillance videos, which has received a lot of attention from researchers. Motion segment extraction is the basis and prerequisite of behavior recognition, surveillance video synopsis and other subsequent processing, and it is also a research hotspot in the field of computer vision. The existing motion segments extraction algorithms are mainly divided into traditional methods and deep learning-based methods. The former needs to process the full amount of data in the spatial domain of the video and detect motion targets frame by frame, which is computationally intensive and time-consuming and can't meet the real-time needs. The latter requires a massive amount of sample data to pre-train the model, which has high algorithm complexity and high requirement for hardware devices. Addressing this problem, this paper proposed a fast motion segment extraction method via nested ellipse spatio-temporal tubes in surveillance video, which can extremely save the amount of calculation. Firstly, surveillance video is elliptically spatio-temporal sampled. The elliptical sampling lines are adaptively generated according to the video sequences with different aspect ratios and pixel on the sampling lines of each frame in the video sequence are extracted to form an elliptical spatio-temporal tube. Secondly, multiple elliptical spatio-temporal tubes sampled progressively according to surveillance scene are integrated to nested elliptical spatio-temporal tubes. Then, nested elliptical spatio-temporal tubes are expanded to generate spatio-temporal plane maps. Finally, the background of spatio-temporal plane maps is removed and the spatio-temporal flow model is constructed to extract motion segments. In this model, the instantaneous spatio-temporal flow curve reflects whether moving targets enter or exit the sub-surveillance area in the corresponding frame, and the accumulative spatio-temporal flow curve reflects the number of moving targets in the sub-surveillance area in the corresponding frame. Flowchart of proposed algorithm as shown in the figure. The proposed algorithm utilizes elliptical spatio-temporal sampling instead of traditional full spatial data processing and requires no pre-training, which can greatly reduce the amount of surveillance video data to be processed. The algorithm form nested elliptical spatio-temporal tubes by progressive sampling, which reduces the computational effectively and can also take into account the targets moving inside the surveillance areas. The experimental on SISOR, KTH and CAVIAR public data sets comparison with mainstream motion segments extraction algorithms. Experimental results show that the proposed algorithm has obvious advantage in calculating time, greatly reduces the amount of calculation under the premise of ensuring detection accuracy, and it can realize fast motion segments extraction in surveillance videos.