TICMapNet: A Tightly Coupled Temporal Fusion Pipeline for Vectorized HD Map Learning

Abstract

High-Definition (HD) map construction is essential for autonomous driving to accurately understand the surrounding environment. Most existing methods rely on single-frame inputs to predict local map, which often fail to effectively capture the temporal correlations between frames. This limitation results in discontinuities and instability in the generated map.To tackle this limitation, we propose a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ti</i>ghtly <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</i>oupled temporal fusion <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Map</i> <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Net</i>work (TICMapNet). TICMapNet breaks down the fusion process into three sub-problems: PV feature alignment, BEV feature adjustment, and Query feature fusion. By doing so, we effectively integrate temporal information at different stages through three plug-and-play modules, using the proposed tightly coupled strategy. Unlike traditional methods, our approach does not rely on camera extrinsic parameters, offering a new perspective for addressing the visual fusion challenge in the field of object detection. Experimental results show that TICMapNet significantly improves upon its single-frame baseline model, achieving at least a 7.0% increase in mAP using just two consecutive frames on the nuScenes dataset, while also showing generalizability across other tasks.