Inferring Driving Maps by Deep Learning-based Trail Map Extraction

TL;DR

Problem: Static HD maps lag behind dynamic driving scenes and sensor configurations; approach: TrailTR, an offline pipeline that converts crowdsourced trail data into a tile-based grid G ∈ $R^{w×h×(n+1)}$ with n directional channels and a speed channel, using a MapTRv2-inspired transformer decoder with a ResNet50 backbone to predict centerlines per tile; contributions: (1) a novel trail-based offline mapping framework, (2) a robust input representation with directional bins and speed, (3) demonstrated superiority over online HD mapping baselines and strong generalization across geographic splits and sensor setups, (4) extensive ablations validating design choices; significance: enables continuous, sensor-agnostic map updates and lays groundwork for extending to lane boundaries and full lane graphs.

Abstract

High-definition (HD) maps offer extensive and accurate environmental information about the driving scene, making them a crucial and essential element for planning within autonomous driving systems. To avoid extensive efforts from manual labeling, methods for automating the map creation have emerged. Recent trends have moved from offline mapping to online mapping, ensuring availability and actuality of the utilized maps. While the performance has increased in recent years, online mapping still faces challenges regarding temporal consistency, sensor occlusion, runtime, and generalization. We propose a novel offline mapping approach that integrates trails - informal routes used by drivers - into the map creation process. Our method aggregates trail data from the ego vehicle and other traffic participants to construct a comprehensive global map using transformer-based deep learning models. Unlike traditional offline mapping, our approach enables continuous updates while remaining sensor-agnostic, facilitating efficient data transfer. Our method demonstrates superior performance compared to state-of-the-art online mapping approaches, achieving improved generalization to previously unseen environments and sensor configurations. We validate our approach on two benchmark datasets, highlighting its robustness and applicability in autonomous driving systems.

Figures (4)

• Figure 1: Visualization of the ground truth lane centerlines and predictions of a popular online mapping model (MapTRv2) and our approach (TrailTR) for a nuScenes validation sample. Both models were trained on the geographically disjoint near extrapolation split lilja_localization_2024. Centerlines with no corresponding trail data were removed from the ground truth in our approach.
• Figure 2: The overall architecture of the trail-based mapping approach. Starting with the trail aggregation (a), object and ego trails are converted into a grid map format. The covered area is divided into tiles (b) and for training the ground truth is obtained from a centerline HD map (c). The resulting samples are then fed into a deep learning model (d) that, once trained, infers tile-wise centerline predictions for the input trail grid maps.
• Figure 3: Visualization of an example sample from the nuScenes dataset.
• Figure 4: Visualization of ground truth (red), model predictions (blue) and its input grid map on an aerial image. The samples are from the nuScenes geograhical validation split (a-b) and nuPlan validation split (c-f). The samples from the nuScenes dataset lack sufficient amounts of trails to depict the entirety of lanes in the map tile, so the ground truth was refined as described in \ref{['sec:input_and_gt']}.