A vision-based parking lot surveillance framework is proposed in this paper. This framework utilizes multi-camera localization and tracking technologies to detect parking spaces and moving cars. The detected results are then used to estimate robust car moving trajectories for surveillance related event analysis and data queries. There are two main contributions of this framework-(1) it transforms each image from different cameras onto a global-view image that maximizes the field-of-view of the multiple cameras; (2) it utilizes Gaussian-Mixture-Model (GMM) to build a parking lot background model which is resistant to illumination variations. Moreover, this article also explored the possibility to build the system automatically through the use of the Scale Invariant Feature Transform (SIFT) and RANdom SAmpling Consensus (RANSAC) algorithms.