In this paper, we propose a novel vision-based on-street parked vehicle detection method via view-normalized classifiers. Our method consists of two phases: (1) an offline process to train site-independent object classifiers and set various site-specific parameters and (2) a runtime process where streaming video frames are analyzed to determine the occupancy of the parking site. We incorporate temporal filtering, view-normalization, and temporal correlation into a core computer vision-based parked vehicle detection method to achieve real-time determination of on-street parking occupancy. Our method combines image processing techniques with machine learning to yield efficient and accurate results. It does not require site specific re-training of the classifiers and thus is most suitable for large deployment or for quick parking occupancy surveys covering a wide-area of interest. Two experiments are conducted. The first experiment consists of six cameras monitoring a block of a street. The results show that our method is robust against site variations as well as camera variations. The other experiment is a small-scale deployment, where 11 cameras are used to monitor four blockfaces of a city. The results confirm that our method can achieve adequate accuracy without re-training of vehicle classifiers or refinement of parameters.