Due to the critical role of traffic flow information in urban activities, we propose a UAV-based system for traffic flow estimation and analysis. Deploying a quadrotor unmanned aerial vehicle (UAV) equipped with a downward-looking camera and range finder, we capture traffic videos for region of interest from a 'bird's eye view'. To effectively process such videos where significant camera jitter may exist, we firstly design a low rank representation based method to simultaneously align the frames and detect the location of road. Then, we represent the well-aligned frames using spatiotemporal image, followed by a discriminative robust principal component analysis (RPCA) algorithm which fuses motion information with appearance for vehicle detection. Subsequently, leveraging on the camera height, the optical flow is exploited to classify vehicles into various size categories. Extensive experiments performed on the traffic data collected on university campus, achieving better performance than available methods.