Mixture distribution estimation for future prediction

A computer-implemented method for mixture distribution estimation of multi-modal future predictions comprising a training phase of a convolutional neural network comprising the steps of: (1) inputting a set of images of a driving environment, each containing at least one object of interest, and a set of future ground truths corresponding to the objects of interest; (2) sampling the solution space of the multi-modal future of the object of interest with an evolving winner-takes-all loss strategy by generating a predetermined number of hypotheses, penalizing all hypotheses equally, gradually releasing one part of the hypotheses by penalizing only the other part of the hypotheses being closer to the corresponding ground truth, so-called winning hypotheses, until only the best hypothesis being the closest one is penalized, and outputting final hypotheses; (3) sequentially fitting a multi-modal mixture distribution of future predictions to the final hypotheses.