People localization and tracking has always been an important and challenging task in vision-based surveillance system, and the growing number of developed techniques has created a lot of applications. In recent years, a lot of approaches of people localization and tracking have been proposed to deal with occlusion problem effectively using multiple cameras. In our work, algorithms based on vanishing points are adopted to generate 2D line samples of foreground images. The 2D line samples from all of views can be filtered via the proposed cross ratio analysis between each pair of cameras so that the computation of people localization can be reduced. In particular, the method developed in [11] is improved to achieve real-time 3D people localization. Experiment results show that our method can speed up the original method by two times without losing correctness and precision.