Vehicle counting and traffic flow analysis with UAV by low rank representation

Abstract

We propose to synergize the motion and appearance information within a block-sparse robust principal component analysis (RPCA) framework for the estimation and analysis of traffic flow. The images are captured by a UAV with a downward looking camera, and the processing of images are performed on the measure line, called virtual gantry, resulting in much improved efficiency. Enforcing the low rank constraint on the spatiotemporal image which is generated via stacking image pixels on virtual gantry over time, we introduce the block-sparse RPCA algorithm in which the motion cue is leveraged to facilitate vehicle detection, followed by flow normalization to classify vehicles into light, small, medium, and large categories. Benefiting from the low rank representation, our detection method is parameter insensitive, robust to illumination changes, and requires no training. Lots of experiments have been carried out in a campus scenario, the UAV is launched and controlled to hover above roadways to capture images. Our method achieves nearly 100% accuracy in vehicle counting and classification, and significantly outperforms the available methods. Meanwhile, insightful observations of the obtained traffic information are given, which could be very valuable to the users, especially to the traffic management sectors.