Quantification of Uncertainties in Deep Learning-based Environment Perception

In this work, we introduce a novel Deep Learning-based approach to perceive the environment of a vehicle based on radar scans while accounting for uncertainties in the model predictions. We segment the environment of the vehicle into equally sized grid cells, which we classify individually. Our algorithm is capable of differentiating uncertainties in the system output as being related to an inadequate model (epistemic uncertainty) or noisy data (aleatoric uncertainty). To this end, we utilize probability distributions on the weights that describe uncertainties in the model parameters and that can be learned in a supervised fashion using gradient descent. We show that the uncertainties in our model output correlate with the precision of the model predictions. Compared to previous concepts that estimate uncertainties by utilizing Monte Carlo Dropout layers, we show the superior performance of our approach to reliably perceive the environment of a vehicle.