High-definition (HD) maps are the essential blueprints for autonomous driving, providing the geometric and semantic context—such as lane boundaries and crosswalks—that vehicles need to navigate. Traditionally, these maps were built offline through expensive manual annotation. The field has recently shifted toward online vectorized HD mapping, where models like VectorMapNet and MapTR generate these maps in real-time directly from onboard sensor data.
However, a critical "stability gap" has emerged. While these models achieve high per-frame accuracy, they often suffer from temporal flickering and geometric jitter. A lane line might appear in one frame, vanish in the next due to a passing truck, and reappear slightly shifted in the third. For a self-driving planner, this instability is more than a visual glitch; it is a safety hazard that can lead to erratic braking or steering.
The Independence of Accuracy and Stability
The central realization in recent research is that per-frame precision does not guarantee temporal reliability. Han Jiang, Hao Shan, and colleagues demonstrate in Stability Under Scrutiny: Benchmarking Representation Paradigms for Online HD Mapping that a model’s mean Average Precision (mAP) is largely independent of its mean Average Stability (mAS).
Traditional benchmarks evaluate maps as static snapshots. In reality, an autonomous vehicle moves through a continuous stream of data. When models process frames in isolation, they are vulnerable to transient noise, lighting changes, and occlusions. To solve this, the research direction has moved from "detecting" maps to "tracking" and "remembering" them.
From Detection to Tracking
If you see a lane divider, you don't just forget it exists the moment a car drives over it. You maintain a mental model of where it should be. Researchers are now building this intuition into neural architectures by reframing mapping as a tracking task.
In MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping, Jiacheng Chen and others argue that road elements should maintain consistent identities across time. Instead of re-detecting every point in every frame, MapTracker uses a memory buffer to propagate information. This ensures that a lane divider in frame $t$ is the same entity as the one in frame $t+1$, preventing the "flickering" that occurs when a model loses and then regains a detection.
Explicit History and Geometric Priors
While some models pass implicit latent embeddings between frames, others have found that explicit history maps provide more robust geometric anchors.
Sen Yang and the team behind HisTrackMap maintain rasterized records of previous perception results. By projecting these "history maps" into the current frame's coordinate system, the model gains a strong spatial prior. If the current sensor data is noisy, the model can rely on its previous, clearer observations to maintain a stable output. Similarly, HRMapNet by Xiaoyu Zhang uses historical rasterized maps to boost vectorized perception, showing that even a lightweight memory of the past can significantly reduce errors in challenging conditions like bad weather or heavy occlusion.
Refining Inner-Instance Geometry
Stability isn't just about whether an object exists; it's about its shape. Vectorized elements like lanes are often represented as ordered sets of points. Without temporal constraints, these points can "jitter," causing the line to look jagged or "zigzagged."
Yi Zhou in HIMap and Zhenhua Xu in InsMapper address this by focusing on inner-instance relationships. By modeling the connections between points within a single lane line and enforcing these relationships across frames, these frameworks ensure that the geometry remains smooth and consistent even as the vehicle's perspective changes.
The Path Toward Global Coherence
The ultimate goal of this research is to move beyond short-term temporal windows toward global consistency. GlobalMapNet by Anqi Shi attempts to bridge the gap between real-time online mapping and large-scale global maps. By integrating persistent memory and global coordinate systems, these systems aim to create a "seamless" map that doesn't just work for the next 50 meters, but remains stable across an entire city-scale trajectory.
By shifting the focus from per-frame accuracy to temporal stability, researchers are making online HD mapping reliable enough for the life-critical decisions required in autonomous driving.