Neural Network Based Multiple Object Tracking for Automotive FMCW Radar

Tracking multiple objects from radar data poses several difficulties. In recent work, it has been shown that an algorithm consisting of a thresholding, clustering and multiple object tracking step using the Kalman filter can track multiple objects. Afterwards, features can be extracted from the range-Doppler map to classify the tracked objects. However, this method needs many heuristics on each stage and in the process, information which could be useful in subsequent steps is lost.To overcome these issues, in this paper we introduce a neural network based multiple object tracker. This removes the need for a separate thresholding, clustering, feature extraction and classification step because it combines those into one step which uses a neural network based on the You Only Look Once (YOLO) object detection system to classify and localize objects. The output of the neural network is fed into a Kalman filter based tracker to manage the tracks.We show that a convolutional neural network trained as an object detector can be successfully applied in the radar domain and we show the advantages of our neural network based multiple object tracker over the clustering based method for specific scenarios. These scenarios include tracking objects that cross each other and tracking objects while the radar is non-stationary.