ADMap: Anti-disturbance framework for reconstructing online vectorized HD map

TL;DR

This work tackles the instability of online vectorized HD maps for autonomous driving, where prediction bias induces jitter in instance point sequences that distort map topology. It introduces ADMap, an anti-disturbance framework with three components: Multi-scale Perception Neck (MPN) for robust multi-scale BEV features, Instance Interactive Attention (IIA) for cross-level inter-instance and intra-instance interactions, and Vector Direction Difference Loss (VDDL) to supervise point sequences via vector direction differences. The approach achieves state-of-the-art performance on nuScenes and Argoverse2, improving accuracy while maintaining real-time inference and reducing point jitter in challenging scenes. By combining cross-scale feature fusion, topology-aware attention, and direction-aware supervision, ADMap provides more stable, reliable, and actionable vectorized HD maps for planning and control in autonomous driving.

Abstract

In the field of autonomous driving, online high-definition (HD) map reconstruction is crucial for planning tasks. Recent research has developed several high-performance HD map reconstruction models to meet this necessity. However, the point sequences within the instance vectors may be jittery or jagged due to prediction bias, which can impact subsequent tasks. Therefore, this paper proposes the Anti-disturbance Map reconstruction framework (ADMap). To mitigate point-order jitter, the framework consists of three modules: Multi-Scale Perception Neck, Instance Interactive Attention (IIA), and Vector Direction Difference Loss (VDDL). By exploring the point-order relationships between and within instances in a cascading manner, the model can monitor the point-order prediction process more effectively. ADMap achieves state-of-the-art performance on the nuScenes and Argoverse2 datasets. Extensive results demonstrate its ability to produce stable and reliable map elements in complex and changing driving scenarios. Code and more demos are available at https://github.com/hht1996ok/ADMap.

Figures (7)

• Figure 1: Performance and visualization comparison between ADMap and baseline. The two endpoints of the line segment in the left figure indicate the results of the multi-modal and camera-only frameworks, respectively. The right figure shows that ADMap effectively mitigates the point sequence jitter problem.
• Figure 2: Schematic diagram of the overall framework of ADMap. The figure displays the MPN and IIA proposed in this paper. The process is performed only during training when indicated by the dashed line, and during both training and inference when indicated by the solid line. In decoder, Instance-Points query are defined to represent the topology of the map, and self-attention and cross-attention are used to interact with the BEV map. Instance self-attention and Points self-attention further interact with inter- and intra-instance information.
• Figure 3: Schematic diagram of Instance self-attention and Points self-attention. The point and channel dimensions in the Instance-Points query are merged and put into embedded layers consisting of multiple MLPs to compress the dimensions. These query are then used in multi-head self-attention for instance-level interactions. Groups the output query of instance self-attention so that each instance is fed into multi-head self-attention separately for point-level interaction.
• Figure 4: Flowchart of VDDL. The point sequence $P$ is modeled as a vector line $L$ and the vector direction difference between the predicted and ground truth is calculated. The weights of each instance point are obtained from the geometric topological relations of ground truth.
• Figure 5: Visualization results of the nuScenes dataset. Areas of discrepancy are indicated by red boxes. ADMap effectively reduces jitter within instances.