Occupancy grid mapping (OGM) is a critical component in automated driving for obstacle detection and free-space navigation. Radar is a key sensing modality for this task because it provides environmentally robust detection data for map estimation. However, existing radar-based OGM methods often struggle to jointly exploit temporal continuity across frames, together with spatial sparsity and correlation in the maps. In contrast, sparsity-aware mapping methods effectively capture spatial features in static maps but do not account for temporal dependencies across frames. To address this limitation, we present sparsity-aware Bayesian dynamic OGM algorithms that integrate temporal knowledge extracted from past frames in the Bayesian framework. This temporal information is encoded in the Bayesian priors through two approaches: either by adapting the map prior mean or by adjusting the model hyperparameters. Evaluations on the View-of-Delft public dataset demonstrate that our methods achieve superior mapping performance by effectively capturing temporal information compared to static sparsity-aware mapping and baseline dynamic OGM algorithms.
Track ID:
10
Track Name:
Sensor Systems: Signals, Processing and Interfaces
This paper addresses automotive radar-based dynamic occupancy grid mapping.
The challenge is that existing methods exploit either spatial sparsity or temporal correlation, but not both: sparsity-aware mappers ignore inter-frame dependencies, while Bayesian dynamic occupancy filters ignore sparsity.
We close this gap by extracting temporal knowledge from past radar returns and embedding it into the sparse Bayesian mapping framework.
Our specific novelties are:
1. We extract dynamic prior support information (PSI) from the previous frame's radar returns. We isolate moving targets by their Doppler velocities and project them into the current frame.
2. We embed the dynamic PSI in the Bayesian prior rather than the output. This lets the measurements correct an imperfect prior, and we do it two ways: shifting the layer-1 prior mean (DL1-PCSBL) or tuning the layer-2 hyperparameters (DL2-PCSBL).
3. On View-of-Delft, both methods improve detection and false-alarm suppression over baselines.
