LOPR: Latent Occupancy PRediction using Generative Models

Abstract

Environment prediction frameworks are essential for autonomous vehicles to facilitate safe maneuvers in a dynamic environment. Previous approaches have used occupancy grid maps as a bird's eye-view representation of the scene and optimized the prediction architectures directly in pixel space. Although these methods have had some success in spatiotemporal prediction, they are, at times, hindered by unrealistic and incorrect predictions. We postulate that the quality and realism of the forecasted occupancy grids can be improved with the use of generative models. We propose a framework that decomposes occupancy grid prediction into task-independent low-dimensional representation learning and task-dependent prediction in the latent space. We demonstrate that our approach achieves state-of-the-art performance on the real-world autonomous driving dataset, NuScenes. We provide our implementation at https:/github.com/sisl/LOPR.

Dynamics-Aware Spatiotemporal Occupancy Prediction in Urban Environments

“…Subsequent work by Toyungyernsub et al [12] integrates static-dynamic object segmentation module into their method to determine stationary and moving objects in the environment, instead of relying on object detection and tracking information as done in the prior work [11]. Lange et al [20] reduce blurriness in predictions by making predictions in the latent space of a generative model. Different from these works, we propose using environment semantics to retain valuable contextual information in the scene, rather than only giving the model access to static and moving object information.…”

Modeling of Coupling Effects in Neural Networks for Ship Motion Prediction

“…A widely adopted representation called Occupancy Grid Map (OGM) captures the spatial arrangement of obstacles and free space where each grid cell represents the estimated probability of an agent's presence within. Predicting future OGM allows the formation of occluded areas, thus offering a more comprehensive understanding of the environment [40], [41]. Nevertheless, these approaches based on OGM can be computationally expensive, particularly for high-resolution, large, and complex environments.…”