Single-Frame Vulnerable Road Users Classification with a 77 GHz FMCW Radar Sensor and a Convolutional Neural Network

Road traffic accidents accounted in 2013 for over a million deaths worldwide. Pedestrians and cyclists are especially vulnerable in road accidents and therefore it is essential to identify them in a timely manner to foresee dangerous situations. Radar sensors are excellent candidates for this task since they are able to simultaneously measure range, radial velocity and angle while remaining robust in adverse weather conditions. In this paper, a method to classify moving subjects as pedestrians, cyclists or cars using single radar measurement frames from a 77 GHz FMCW radar sensor is proposed. To perform the classification the range-Doppler-angle power spectrum is run through a convolutional neural network. A dataset of around 9.1k frames gathered in urban scenarios is used to train the convolutional neural network. A classification accuracy as high as 97.3% is achieved on a set consisting of tracks not seen during training but on known locations. The classification accuracy drops to 84.2% when tested on unseen tracks gathered in an unseen location.