面向自动驾驶场景的脉冲视觉研究

Spike-Based Vision for Autonomous Driving Scenarios

自动驾驶是计算机视觉重要的研究方向,具有广阔的应用前景。纯视觉感知方案在自动驾驶场景中具有重要的研究价值。不同于传统相机,脉冲视觉传感器能更加灵敏地感受光子,具备比传统视频快千倍以上的成像速度,具有高时间分辨率、高动态范围、低数据冗余和低功耗等优势。本文面向自动驾驶场景,总结了脉冲相机的成像原理、感知能力与优势;围绕自动驾驶相关视觉任务,详细介绍了脉冲视觉影像重建原理与方法,讨论了基于传感器融合的影像增强技术路线;归纳总结了基于脉冲相机的运动光流估计、目标识别检测分割与跟踪,以及三维场景深度估计算法进展及技术路线;梳理了脉冲相机数据及感知系统的发展现状,分析了脉冲视觉的研究挑战;研究提出了潜在解决方案及未来研究方向。脉冲相机及其算法和系统在自动驾驶领域具有巨大潜力,是未来计算机视觉的主要研究方向之一。

Autonomous driving is an important research direction in computer vision which has broad application prospects.Pure vision perception schemes have significant research value in autonomous driving scenarios.Different from traditional cameras,spike vision sensor offers imaging speeds over a thousand times faster than traditional cameras,possess advantages such as high temporal resolution,high dynamic range,low data redundancy,and low power consumption.This study focuses on autonomous driving scenarios,introducing the imaging principles,perception capabilities,and advantages of the spike camera.Besides,focusing on visual tasks related to autonomous driving,this study elaborates on the principles and methods of spike-based image/video reconstruction,discusses the approach to image enhancement based on sensor fusion with spike cameras,and provides a detailed description of the algorithms and technical routes for motion optical flow estimation,object recognition,detection,segmentation,and tracking,and deep estimation of three-dimensional scenes based on spike cameras.It also summarizes the development of the spike camera data and systems.At last,it analyzes the challenges,potential solutions,and future directions for spike vision research.Spike cameras and their algorithms and systems hold great potentials in the field of autonomous driving and represent one of the future research directions in computer vision.