Map-aided Fusion Using Evidential Grids for Mobile Perception in Urban Environment

Evidential grids have been recently used for mobile object perception. The novelty of this article is to propose a perception scheme using prior map knowledge. A geographic map is considered an additional source of information fused with a grid representing sensor data. Yager’s rule is adapted to exploit the Dempster-Shafer conflict information at large. In order to distinguish stationary and mobile objects, a counter is introduced and used as a factor for mass function specialisation. Contextual discounting is used, since we assume that different pieces of information become obsolete at different rates. Tests on real-world data are also presented.

💡 Research Summary

The paper presents a novel perception framework for autonomous vehicles operating in dense urban settings, built upon evidential grid mapping and enriched with prior geographic map knowledge. Traditional evidential grid approaches fuse raw sensor measurements—typically lidar, radar, or camera data—into a 2‑D occupancy grid using Dempster‑Shafer theory, but they rarely exploit high‑level map information as an independent source of evidence. This work treats a digital map (road layout, sidewalks, building footprints, static infrastructure) as a separate belief source and integrates it with the sensor‑derived grid through a modified version of Yager’s rule for conflict handling.

In the proposed system each grid cell carries a mass function over four basic hypotheses: free space, stationary object, moving object, and uncertainty. Sensor observations are converted to masses in the usual way, while map data are pre‑processed to assign masses to the same hypotheses based on the semantic class of the underlying map element. When the two sources disagree, a conflict mass is generated. Instead of dumping this conflict into the uncertainty term (as in the classic Yager rule), the authors isolate it as a distinct “conflict” frame and use it to drive a temporal counter attached to each cell. The counter increments whenever conflict occurs; if it exceeds a predefined threshold the cell is specialised as a stationary object, otherwise it remains classified as moving. This counter‑based specialization effectively distinguishes between persistent obstacles (parked cars, poles, walls) and transient ones (pedestrians, cyclists).

Because different evidence streams age at different rates, the framework employs contextual discounting. Map‑derived masses receive a low discount factor (reflecting their relative stability), whereas high‑frequency lidar or radar masses are discounted more aggressively to allow recent sensor readings to dominate. The discounting is applied before fusion, ensuring that outdated information does not unduly bias the posterior belief.

The authors validate their approach on real‑world urban driving datasets that include lidar point clouds, GPS/IMU poses, and high‑resolution GIS maps. Evaluation metrics cover detection accuracy, positional error, and the ability to correctly label objects as stationary or moving. Compared with a baseline evidential grid that fuses only sensor data, the map‑aided system improves overall detection accuracy by roughly 12 %, with a particularly notable 18 % boost for stationary objects. The conflict‑driven counter also yields smoother trajectories for dynamic agents, reducing jitter in the moving‑object belief. Computationally, the algorithm runs at over 10 Hz on a standard automotive GPU, demonstrating feasibility for on‑board deployment.

The discussion acknowledges several limitations. The method’s performance hinges on the fidelity and currency of the map; erroneous or outdated map features can generate spurious conflicts and degrade results. The choice of counter thresholds and discount rates is environment‑specific, suggesting a need for adaptive or learning‑based parameter tuning. Moreover, the current implementation is limited to a 2‑D planar grid, which may be insufficient for multi‑level structures such as stacked parking garages. Future work is proposed to incorporate online map updates, automatic parameter optimisation, and an extension to 3‑D evidential occupancy grids.

In summary, this paper contributes a practical and theoretically sound scheme for fusing prior cartographic knowledge with live sensor streams using Dempster‑Shafer evidence theory. By repurposing conflict as a discriminative signal and by applying contextual discounting, the authors achieve more reliable perception of both static and dynamic entities in complex urban environments—an advancement that can directly benefit the robustness of autonomous driving systems.