Double-Prong ConvLSTM for Spatiotemporal Occupancy Prediction in Dynamic Environments

Predicting the future occupancy state of an environment is important to\nenable informed decisions for autonomous vehicles. Common challenges in\noccupancy prediction include vanishing dynamic objects and blurred predictions,\nespecially for long prediction horizons. In this work, we propose a\ndouble-prong neural network architecture to predict the spatiotemporal\nevolution of the occupancy state. One prong is dedicated to predicting how the\nstatic environment will be observed by the moving ego vehicle. The other prong\npredicts how the dynamic objects in the environment will move. Experiments\nconducted on the real-world Waymo Open Dataset indicate that the fused output\nof the two prongs is capable of retaining dynamic objects and reducing\nblurriness in the predictions for longer time horizons than baseline models.\n