AutoSAR Dataset

In current development of autonomous driving technology, understanding image-like representations, such as optical images, is essential for enhancing the reliable environmental perception. However, under harsh weather conditions, typical sensors in artificial vision, often fail to meet the demands of accurate target recognition. Synthetic aperture radar (SAR) technology, which can deliver high-resolution images under all weather conditions, is being increasingly integrated into automotive systems. However, the challenges like dynamic trajectory estimation and real-time SAR imaging remain significant obstacles. In this paper, aimed at providing timely and high-quality SAR images for driving environment, a novel vehicular SAR imaging approach is proposed. By employing sub-aperture scheme for SAR system, the approach eliminates the need for complicated and timeconsuming range cell migration and motion correction. Due to the short coherent accumulation, the instant range doppler algorithm enables high efficient SAR image generation while maintaining two-dimensional (2D) high-resolution performance, which allows precise target detection. The theoretical analysis and experimental results confirm the effectiveness of the proposed scheme. Then, a novel benchmark is established based on the proposed millimeter-wave (mmW) SAR approach, to thoroughly cover the SAR images in real-time environment. To validate the radar data, diverse targets in SAR images are annotated, providing a robust database support for the automotive SAR target recognition. Furthermore, a variety of mainstream deep learning based-methods are performed to complete the SAR vision understanding task, facilitating the practical application in autonomous driving. 

在自动驾驶技术的当前研发进程中，理解光学图像这类类图像表征，对于提升可靠的环境感知能力至关重要。然而，在恶劣天气条件下，人工视觉常用的典型传感器往往难以满足精准目标识别的需求。合成孔径雷达（Synthetic Aperture Radar, SAR）技术可在全天气条件下输出高分辨率图像，正日益被集成至车载系统中。但诸如动态轨迹估计与实时SAR成像等挑战，仍是亟待突破的重大阻碍。针对自动驾驶场景下对实时高质量SAR图像的需求，本文提出了一种新型车载SAR成像方法。该方法采用SAR系统的子孔径架构，无需进行复杂耗时的距离徙动与运动校正操作。得益于短时间相干积累，瞬时距离多普勒算法可在维持二维（2D）高分辨率性能的同时，实现高效的SAR图像生成，从而支持精准的目标检测。理论分析与实验结果均验证了所提方案的有效性。随后，基于本文提出的毫米波（Millimeter Wave, mmW）SAR方法，构建了一个全新的基准数据集，可全面覆盖实时驾驶场景下的SAR图像。为验证雷达数据的可靠性，本文对SAR图像中的多种目标进行了标注，为车载SAR目标识别任务提供了鲁棒的数据库支撑。此外，本文还采用了多种主流深度学习方法完成SAR视觉理解任务，以推动该技术在自动驾驶场景中的实际落地应用。