【摘要】 随着自动驾驶感知技术的深入研究，基于单车的激光雷达3D目标检测算法已经达到较高的精度。然而，单车感知存在感知范围有限和视野盲区的固有局限，难以满足高级别自动驾驶对感知系统的更高要求。因此，协同感知技术近年来受到广泛关注。在真实场景中，定位设备和通讯延迟引起的时空异步会导致车路协同检测性能下降。本文提出了EFlow，一种基于自车特征流的异步协同检测方法。该流程由两个部分组成：首先，自车鸟瞰流图是从自车的连续历史帧中获取空间运动向量，进而移动特征到合适的位置；其次，本文设计了一种多尺度融合骨干，提升了模型对于异步特征的鲁棒性。本文在真实世界数据集DAIR-V2X和仿真数据集V2Xset上进行了大量的实验，实验结果表明本文所提方法可以有效减轻时空异步导致的检测性能下降，且性能明显优于基线方法。更多还原

【Abstract】 With in-depth research into perception technology for autonomous driving, single-vehicle LiDAR-based 3D object detection algorithms have reached a high level of precision. Nevertheless, perception from a single vehicle has intrinsic drawbacks, including a limited field of view and occlusion, making it difficult to satisfy the advanced requirements of highly automated driving systems. As a result, collaborative perception has become a subject of extensive research in recent years. In practical application, spatio-temporal misalignment arising from localization devices and communication delays can degrade the performance of vehicle-road collaborative detection. In this paper EFlow is proposed, which is an asynchronous collaborative detection method based on ego-vehicle feature flow. The framework is composed of two parts. Firstly, an ego-vehicle Bird’s-Eye View(BEV) flow acquires spatial motion vectors from its consecutive historical frames to move features to their correct locations. Furthermore, a multiscale fusion backbone is designed to improve the model’s robustness against asynchronous features. Extensive experiments on the real-world dataset DAIR-V2X and the simulation dataset V2 Xset show that the proposed method can effectively alleviate the decline in detection performance caused by spatio-temporal asynchrony and achieves superior performance over baseline methods.更多还原