Vectorized HD semantic map construction based on end-to-end object detection with transformers

To address the significant consumption of computing resources by traditional visual SLAM solutions when annotating and maintaining semantic information during the mapping process, which limits their scalability, this paper proposes a structured end-to-end vectorized high-definition map construction scheme. By representing map element features through point set prediction, this approach substantially reduces the computational overhead associated with numerous downstream tasks, mitigates the high costs of map updates, and achieves more robust and accurate high-definition map construction. Preliminary feature extraction is performed on input images capturing surrounding road conditions from a camera perspective, followed by secondary feature extraction using an efficient encoder-decoder Transformer attention mechanism to obtain BEV features based on prior geometry. A lookup table index structure is introduced to significantly enhance image inference speed. The Hungarian matching algorithm is employed to align the extracted instance features with ground truth data. This scheme enables mapping for a wide range of scenes using only visual images, demonstrating strong generalization capabilities. Additionally, leveraging the online learning capabilities of neural networks, this method maintains stable and efficient map construction quality even in complex and diverse environments.