Based on background modeling that employs a nonlinear model, a new approach for slow moving object detection in rotating complex background is proposed. First, the rotational motion pattern is determined and the complexity of the background is evaluated. Then background modeling of rotational motion is conducted through the following three steps: (1) retrieving a set of matching feature points and obtaining feature point motion vector field; (2) filtering the vector field and obtaining precise matching feature point set; (3) estimating nonlinear model parameters and reconstructing background motion vector field. Next, background compensated image is obtained by applying pixel interpolation. Last, moving objects are detected by applying threshold segmentation and morphological processing to difference images and searching for complete object contour according to region growing rule. The main test video sequence consists of 1200 frames with a resolution of 720 × 560 pixels. We successfully detected objects with an inter-frame velocity as low as 1 pixel per frame. The false alarm rate, misdetection rate, and the accuracy rate of detection are 2%, 2.6% and 95.4%, respectively. The detection of each frame takes 38 ms, meeting the real-time requirement of object detection in videos with a frame rate of 25 frames per second.