Enhancing Autonomous Driving with PointMapNet: A Scalable Approach for High-Definition Map Construction

Perception of the surrounding environment based on sensor data, such as the task of online high-definition (HD) map construction, is of great significance for the downstream planning and control modules of autonomous vehicles and robots. Previous approaches predominantly adopt a methodology grounded in dense bird's-eye view (BEV) features for fusing multi-view images and other sensor data. Nevertheless, this approach leads to a substantial amount of redundant computation within detection tasks. Consequently, it squanders the valuable computing resources on the vehicle side, also imposing limitations on its application within the robotics domain. In this paper, we introduce PointMapNet, a novel framework for the construction of high-definition (HD) maps that shifts from traditional dense Bird's Eye View (BEV) feature reliance to a point feature-based approach. By integrating sparse, target-oriented point features, PointMapNet drastically reduces computational demands—by approximately 95% in encoder layer feature quantity—thereby significantly increasing the inference speed. Specifically, it attains 30.3 FPS on NVIDIA 4090 compared to the previous SOTA's 24.2 FPS. Our method employs an innovative attention mechanism with three-dimensional spatial position embeddings, which optimizes the detection and vectorization of map elements with unprecedented accuracy. Extensive validation shows that PointMapNet outperforms existing state-of-the-art methods, facilitating faster and more efficient HD map construction. The broader implications of our work extend beyond autonomous driving, offering substantial benefits to robot planning, logistics, and real-time environmental monitoring, paving the way for advanced real-time geospatial analytics in various applications. The code is available at https://github.com/liamkuan/PointMapNet .