关键词: 鸟瞰图, 三维目标检测, 预训练, 时序特征, 高度特征, 大模型

Abstract: Visual perception, as one of the core technologies of environmental understanding, provides accurate environmental information for intelligent mobile systems (such as autonomous vehicles) and is an important prerequisite for ensuring safety decisions.  3D object detection technology based on Bird's Eye View (BEV) has become the mainstream paradigm in the field of environmental perception because of its efficiency and accuracy. To further promote the research of three-dimensional object detection algorithms based on BEV, the covered algorithms were first systematically classified, and according to the modes of the input data, they were divided into three categories: pure camera algorithm, pure lidar algorithm and camera-lidar fusion algorithm. Secondly, the role of pre-training algorithms in improving detection performance was explored. Then, the advantages of fusion timing characteristics were analyzed in dynamic scenarios and the performance of fusion high-level characteristics in complex environments. Besides, the breakthrough progress made in target detection accuracy and scenario understanding of large language model collaborative BEV target detection was sorted out. Finally, the core conclusions were summarized and future research directions were looked forward to provide new ideas for research work in this field.

Key words: Bird's Eye View（BEV）, 3D object detection, pre-training, temporal features, height features, large language model