Focusing on low counting accuracy caused by serious occlusions and abrupt illumination variations, a new real-time statistical method based on Gaussian Mixture Model (GMM) and Scale-Invariant Feature Transform (SIFT) features for video crowd counting was proposed. Firstly, the moving crowd were detected by using GMM-based motion segment method, and then the Gray Level Co Occurrence Matrix (GLCM) and morphological operations were applied to remove small moving objects of background and the dense noise in non-crowd foreground. Considering the high time-complexity of GMM algorithm, a novel parallel model with higher efficiency was proposed. Secondly, the SIFT feature points were acted as the basis of crowd statistics, and the execution time was reduced by using feature exaction based on binary image. Finally, a novel statistical analysis method based on crowd features and crowd number was proposed. The data sets with different level of crowd number were chosen to train and get the average feature number of a single person, and the pedestrians with different densities were counted in the experiment. The algorithm was accelerated by using multi-stream processors on Graphics Processing Unit (GPU) and the analysis about efficiently scheduling the tasks on Compute Unified Device Architecture (CUDA) streams in practical applications was conducted. The experimental results indicate that the speed is increased by 31.5% compared with single stream, by 71.8% compared with CPU.