DESIRE: Distant Future Prediction in Dynamic Scenes with Interacting Agents

摘要

We introduce a Deep Stochastic IOC1 RNN Encoderdecoderframework, DESIRE, for the task of future predictionsof multiple interacting agents in dynamic scenes.DESIRE effectively predicts future locations of objects inmultiple scenes by 1) accounting for the multi-modal natureof the future prediction (i.e., given the same context, futuremay vary), 2) foreseeing the potential future outcomes andmake a strategic prediction based on that, and 3) reasoningnot only from the past motion history, but also from thescene context as well as the interactions among the agents.DESIRE achieves these in a single end-to-end trainable neuralnetwork model, while being computationally efficient.The model first obtains a diverse set of hypothetical futureprediction samples employing a conditional variational autoencoder,which are ranked and refined by the following RNNscoring-regression module. Samples are scored by accountingfor accumulated future rewards, which enables betterlong-term strategic decisions similar to IOC frameworks.An RNN scene context fusion module jointly captures pastmotion histories, the semantic scene context and interactionsamong multiple agents. A feedback mechanism iterates overthe ranking and refinement to further boost the predictionaccuracy. We evaluate our model on two publicly availabledatasets: KITTI and Stanford Drone Dataset. Our experimentsshow that the proposed model significantly improvesthe prediction accuracy compared to other baseline methods