基于事件的视觉SLAM综述

Review of Event-based Visual SLAM

事件相机是一种生物学启发的新型视觉传感器。不同于传统相机输出强度图像帧,事件相机检测每个像素点的亮度变化,当亮度变化超过某一阈值,产生包含像素位置、时间戳及变化极性的像素级事件输出。同步定位与地图构建(SLAM)是一种让机器人独立估计自身位置并同步增量式地构建周围环境地图的技术,使用传统相机进行SLAM已经是一个成熟的话题,但在高速、高动态范围等挑战性场景下,传统图像帧易产生运动模糊。事件相机具有低延时、高时间分辨率、高动态范围等优秀特性,可以在SLAM领域弥补传统相机的不足,但“事件”这种非常规的异步输出形式导致了处理范式的转变。首先介绍了SLAM技术的基本流程,其次介绍了DVS、ATIS和DAVIS三种典型的事件相机,列举了基于事件的SLAM性能评估方法,梳理了事件寿命估计、时间平面表示、事件帧表示等事件数据的不同处理方式,然后回顾了事件相机在国内外视觉SLAM领域的应用研究并进行了总结和对比,最后对基于事件的视觉SLAM技术作了总结并展望了其未来发展趋势。综述表明,事件相机可以进一步拓宽视觉SLAM的应用场景,但新的数据形式和尚不成熟的硬件体系限制了基于事件的SLAM的发展。未来基于事件的SLAM可以围绕以下主题开展进一步研究:新的基于事件的传感器的研发、新的处理事件数据的方法、基于事件的SLAM与深度学习的结合以及基于双目事件的SLAM。

Event camera is a novel bio-inspired vision sensor.Unlike traditional cameras that output intensity image frames,the event camera detects brightness changes at each pixel point,when the brightness change exceeds a certain threshold,it generates pixel-level event output including pixel position,time stamp and polarity.Simultaneous Localization And Mapping(SLAM)is a technology allowing robots to independently estimate their location and simultaneously and incrementally construct a map of the surrounding environment SLAM with traditional cameras is a hot topic,in challenging scenes such as high speed and high dynamic range,motion blur,however,is easy to occur in traditional image frames.Event cameras have excellent features like low delay,high time resolution and high dynamic range,which can make up for the shortcomings of traditional cameras in the field of SLAM,but a paradigm change in processing is caused by the unconventional asynchronous output of events.This paper firstly introduces the basic flow of SLAM technology,next 3 typical event cameras(DVS,ATIS,DAVIS)are introduced,methods for evaluating the performance of event-based SLAM are listed and different processing methods such as events'lifetime estimation,time-surface,event frame are sorted out,then the applications research of event camera in visual SLAM at home and aboard is reviewed and a summary and comparison is made,finally the event-based visual SLAM technology is summarized and its future development trend is looking forward.The review shows that event cameras can further broaden the application scenarios of visual SLAM,but new data forms and immature hardware systems limit the development of event-based SLAM.In the future,event-based slam can be further studied around the following topics:the development of new event-based sensors,new methods for processing event data,the combination of event-based SLAM and deep learning,and event-based stereo SLAM.