Micro-Doppler-based target classification capabilities of the automotive radars can provide high reliability and short latency to the future active safety automotive features. A large number of pedestrians surrounding vehicle in practical urban scenarios mandate prioritization of their treat level. Classification between relevant pedestrians that cross the street or are within the vehicle path and those that are on the sidewalks and move along the vehicle rout can significantly minimize a number of vehicle-to-pedestrian accidents. This work proposes a novel technique for a pedestrian direction of motion estimation which treats pedestrians as complex distributed targets and utilizes their micro-Doppler (MD) radar signatures. The MD signatures are shown to be indicative of pedestrian direction of motion, and the supervised regression is used to estimate the mapping between the directions of motion and the corresponding MD signatures. In order to achieve higher regression performance, the state of the art sparse dictionary learning based feature extraction algorithm was adopted from the field of computer vision by drawing a parallel between the Doppler effect and the video temporal gradient. The performance of the proposed approach is evaluated in a practical automotive scenario simulations, where a walking pedestrian is observed by a multiple-input-multiple-output (MIMO) automotive radar with a 2D rectangular array. The simulated data was generated using the statistical Boulic-Thalman human locomotion model. Accurate direction of motion estimation was achieved by using a support vector regression (SVR) and a multilayer perceptron (MLP) based regression algorithms. The results show that the direction estimation error is less than $10^{\circ}$ in $95\%$ of the tested cases, for pedestrian at the range of $100$m from the radar.