METHOD FOR PREDICTING MULTIPLE FUTURES

A computer-implemented method comprising an operating phase comprising the steps of receiving one or several video frames from a plurality of modalities, so-called multi-modality video frames, of a vehicle's environment, corresponding to present and past timestamps; encoding into a latent representation, said multi-modality video frames by a spatial-temporal encoding convolutional neural network (E); combining into a composite representation (Z), said latent representation with encoded conditioning parameters corresponding to timestamps at the desired future time horizon; predicting multiple future multi-modality video frames corresponding to multiple future modes of a multi-modal future solution space associated with likelihood coefficients by a generative convolutional neural network (G) previously trained in a generative adversarial network training scheme.