An algorithm combining Center Surround Extremas(CenSurE) and spatial-temporal information was proposed to improve the detection speed and detection integrity for a moving object in dynamic scenes. Firstly, in terms of the rapidity and accuracy of CenSurE feature point extraction, an inter-frame image registration method for the motion sequence base on CenSurE and a homography transformation model was proposed to compensate the translation factors, rotation angles and scaling coefficients caused by a moving camera. Then, the foreground mask was generated according to the temporal inter-frame difference of the registrated background in a time domain. Moreover, a dynamical updating background was established with spatial information of the foreground mask and a complete moving object was extracted through background subtraction and a self-adaptive threshold segmentation method based on probability and statics. Finally, the experiments on several standard video sequences were tested to evaluate the performance of this algorithm. Experiment results show that the operating speed of this proposed algorithm can be 15 frame/s, meanwhile, complete moving objects can be obtained with high detecton speeds. The proposed algorithm basically meets the demands of moving object detection in dynamic scenes for speeds, noise resistances, light adaptability, object integrity and so on.