Categorized Grid and Unknown Space Causes for LiDAR-based Dynamic Occupancy Grids

TL;DR

The paper addresses how LiDAR-based dynamic occupancy grids can benefit from explicit unknown-space analysis. It introduces a Categorized Grid (CG), a multi-label per-cell representation that first splits space by occupancy and then annotates occupied cells with dynamic behavior and reliability while attributing unknown space to causes such as FoV limitations, sensing gaps, or occlusions. By combining occupancy, dynamics, FoV analysis, sensing, and occlusion labeling, CG provides a richer, more interpretable view of scene understanding with real-world validations. This approach enhances perception transparency for autonomous driving by clarifying why certain regions remain unknown and how perception constraints influence occupancy estimations. The work has practical implications for safer planning and more robust perception pipelines, and points to temporally-filtered label updates as a direction for future improvement.

Abstract

Occupancy Grids have been widely used for perception of the environment as they allow to model the obstacles in the scene, as well as free and unknown space. Recently, there has been a growing interest in the unknown space due to the necessity of better understanding the situation. Although Occupancy Grids have received numerous extensions over the years to address emerging needs, currently, few works go beyond the delimitation of the unknown space area and seek to incorporate additional information. This work builds upon the already well-established LiDAR-based Dynamic Occupancy Grid to introduce a complementary Categorized Grid that conveys its estimation using semantic labels while adding new insights into the possible causes of unknown space. The proposed categorization first divides the space by occupancy and then further categorizes the occupied and unknown space. Occupied space is labeled based on its dynamic state and reliability, while the unknown space is labeled according to its possible causes, whether they stem from the perception system's inherent constraints, limitations induced by the environment, or other causes. The proposed Categorized Grid is showcased in real-world scenarios demonstrating its usefulness for better situation understanding.

Figures (8)

• Figure 1: Scheme of the proposed space categorization. M-FoV/O-FoV/F-FoV: Maximum, Occupied, and Free Field of View.
• Figure 2: Illustrative example of occupied space categorization. For better visualization, reliable cells are only drawn regarding their dynamic behavior.
• Figure 3: Illustrative example of the O-FoV, F-FoV and Sensed area calculation. The perception system used has a four-layer LiDAR sensor and two valid height ranges for occupied and free space measurements. O-FoV and F-FoV are calculated for $n_{iter} = 1$, so they only take into account the occupancy observation, which is primarily determined by the height ranges, i.e. they cover the space until the laser beams surpass them. The Sensed area is computed as the set of cells traversed by beams with valid information, i.e. it is limited by the obstacles, the ground and the maximum height value.
• Figure 4: Illustrative example of the process for computing the occlusion produced by a cluster with a complex form. a) Estimated occupancy. b) Example of some cluster's cells projected by their line of sight. c) Occluded space boundary calculation using the traversed cells and grid borders. d) Cell categorization considering occlusions' boundaries and obstacles' dynamic state.
• Figure 5: CG calculation example with special emphasis on the unknown space categorization. Unknown cells are only visualized regarding their possible causes, i.e. labels visible, sensed and non-occluded are not represented. In the case of unknown cell with these three labels at the same time, it is considered that the causes are unidentified and an auxiliary label other is set.