Occupancy Flow Fields for Motion Forecasting in Autonomous Driving

Mahjourian, Reza; Kim, Jinkyu; Chai, Yuning; Tan, Mingxing; Sapp, Ben; Anguelov, Dragomir · 2022 · Crossref

DOI: 10.1109/lra.2022.3151613

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper introduces Occupancy Flow Fields, a novel spatio-temporal representation for motion forecasting in autonomous driving that addresses limitations in existing trajectory-based and occupancy-grid methods. While trajectory sets struggle with non-parametric distributions and occupancy grids lose agent identity and motion information, Occupancy Flow Fields combine both by predicting a grid where each cell contains an occupancy probability and a two-dimensional flow vector indicating motion direction and magnitude. This approach enables the model to capture joint probabilities of agent existence, recover agent identities through flow tracing, and predict "speculative agents"—objects currently occluded or outside the field of view that may appear in the future. The proposed deep learning architecture utilizes a PointPillars-inspired encoder to process sparse inputs, including past agent states, road structure points, and traffic light states. These inputs are converted into per-pillar embeddings and processed by an EfficientDet-based decoder to generate occupancy and flow predictions for vehicles and pedestrians over multiple future timesteps. The model employs a backward flow formulation, where flow vectors point to an agent’s previous location, allowing a single flow field to represent multiple potential futures without requiring complex multi-vector predictions. Training involves supervised losses for occupancy and flow, alongside a novel self-supervised "flow trace loss." This loss warps current occupancy predictions through the predicted flow fields over time and enforces consistency with ground truth, ensuring that flow and occupancy predictions remain physically coherent. Experiments were conducted on a large in-house "Crowds" dataset derived from Waymo data and the public INTERACTION dataset. The model predicts occupancy and flow for up to six seconds into the future using one second of past observations. Results demonstrate that incorporating the flow trace loss significantly improves performance across all metrics, including occupancy accuracy (AUC and Soft-IoU), flow error (EPE), and agent ID recall. Specifically, the flow-traced occupancy metrics, which evaluate the consistency of the entire motion trajectory, showed substantial gains when the trace loss was applied. The method outperformed state-of-the-art trajectory prediction models like MultiPath and contemporaneous occupancy-based methods like MP3. Additionally, the model successfully predicted speculative agents, handling scenarios where agents enter the scene from occluded areas, a task not addressed by prior trajectory-based models. The significance of this work lies in providing a unified, non-parametric representation that captures both the probabilistic location and motion of multiple agents while preserving identity information. By enabling the recovery of agent IDs and predicting speculative agents, Occupancy Flow Fields offer a more robust input for downstream planning algorithms in autonomous driving. The introduction of the flow trace loss establishes a new standard for evaluating consistency in motion forecasting, demonstrating that enforcing temporal coherence between occupancy and flow predictions leads to more accurate and reliable forecasts.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation