Incremental perception on real time 3D data
Incremental perception on real time 3D data
复制标题
DOI：
10.1145/3508396.3512875
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
A. Sarker;F. Lin
中科院分区：
文献类型：
--
作者：
A. Sarker;F. Lin
基于3D数据的自主车辆感知以较高的计算代价追求较高的精度。它延迟了反病毒管道中的下游模块-计划、预测和控制。这延长了无人机的决策时间，并最终损害了车辆的机动性和安全性。为了更有效地感知，我们的见解是，感知模块应该关注与下游模块更相关的3D像素。因此，我们提出了增量式感知：每个3D帧在渐进迭代中进行处理；早期迭代向下游模块发出不准确的感知结果；后续迭代合并来自下游模块的反馈，并相应地细化3D帧的最相关部分。我们的早期结果表明，AV的决策时间缩短了，因此安全性得到了改善。
Perception based on 3D data in autonomous vehicles (AV) pursues high accuracy at a high computational cost. It delays the downstream modules in an AV pipeline --- planning, prediction, and control. This prolongs the AV's time-to-decision and ultimately hurts the vehicle's maneuver and safety. Towards efficient perception, our insight is that the perception module should attend to 3D pixels that are more relevant to the downstream modules. Accordingly, we propose incremental perception: each 3D frame is processed in progressive iterations; early iterations emit inexact perception results to the downstream modules; later iterations incorporate feedback from the downstream modules and accordingly refine a 3D frame's most relevant portions. Our early results show reduction in AV's time-to-decision and therefore safety improvement.