Estimating speed profiles from aerial vision — A comparison of regression based sampling techniques
Sebastian Freis, Miguel A. Olivares-Méndez, Francesco Viti · 2016
Since traffic congestion became an increasingly complex problem, research in traffic surveillance tremendously increased in the last decades. The data obtained from the vehicles' paths is non-linear with a high complexity. This type of data is frequently treated using regression methods to obtain reasonable results and to keep track of unrealistic outliers and measured errors. Vision-based techniques are used to obtain this data from static cameras located close by to highways or high congestion traffic points. The use of Unmanned Aerial Vehicles (UAVs) for traffic surveillance is another potential and promising application of this type of robotic platform. The main advantages are that it can be equipped with different sensors based on the requirement of the situation, such as RGB, infrared or multispectral cameras an.d it can change its' position easily based on the traffic situation requirements. This paper describes a vision algorithm for the background subtraction based vehicle tracking approach for vehicle speed estimation using aerial images taken from an UAV. The raw data obtained from the tracker are smoothed with several common regression methods to evaluate their effectiveness to be applied to this type of non-linear data. Furthermore, an extended version of the locally weighted regression algorithm was implemented, which was proven to be the most accurate method for smoothing the tracking results. This data was compared with the ground truth data acquired by a speed sensor installed in the vehicle.