eduzhai > Applied Sciences > Engineering >

An LSTM Approach to Temporal 3D Object Detection in LiDAR Point Clouds

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 18 pages

Abstract: Detecting objects in 3D LiDAR data is a core technology for autonomousdriving and other robotics applications. Although LiDAR data is acquired overtime, most of the 3D object detection algorithms propose object bounding boxesindependently for each frame and neglect the useful information available inthe temporal domain. To address this problem, in this paper we propose a sparseLSTM-based multi-frame 3d object detection algorithm. We use a U-Net style 3Dsparse convolution network to extract features for each frame s LiDARpoint-cloud. These features are fed to the LSTM module together with the hiddenand memory features from last frame to predict the 3d objects in the currentframe as well as hidden and memory features that are passed to the next frame.Experiments on the Waymo Open Dataset show that our algorithm outperforms thetraditional frame by frame approach by 7.5 mAP@0.7 and other multi-frameapproaches by 1.2 while using less memory and computation per frame. To thebest of our knowledge, this is the first work to use an LSTM for 3D objectdetection in sparse point clouds.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×