Incremental Perception on Real Time 3D data



Real time 3D data processing in autonomous vehicles (AV), has a significant overhead. To get better accuracy, there is a latency in the whole pipeline of perception, planning and control, which causes the failure of the system. Existing solutions focused on improving the network models and algorithms of 3D vision. But the latency between data perception and planning incentive is crucial to a safe execution of control decisions. In this paper, we represent a comprehensive analysis on AVs and find different factors of causing delays which impact the critical motion planning. In order to reduce the latency and better planning, we propose a method of incremental perception with priority object vectors, which helps an AV system to act based on partial processing with low delays and on demand refined perception. Our preliminary observations show that incremental perception can reduce delays on motion planning and help to make the right control decision.


  • A key challenge an AV, is the tension between high computation cost and the need for low latency.

  • The delay should be less than 100 ms to ensure safety and passenger comfort [1].

  • While much prior work focuses on optimizing 3D vision for efficiency [ 6 ], our key observation is that the attention over an entire 3D point cloud, which is a waste of resources.

  • Though processing a full point cloud frame, increase the accuracy of object detection, but motion tracking to keep the system in the right path, has a significant overhead.

  • The accuracy of perception ( mAP , IoU ) fails to capture its relevance to downstream tasks when the motion control depends on a critical task that is highly dependent on latency.