| dc.description.abstract | As an emerging technology, LiDAR point cloud has been applied in a wide range of fields.
With the ability to recognize and localize the objects in a scene, point cloud object detection
has numerous applications. However, low-density LiDAR point clouds would degrade the
object detection results. Complete, dense, clean, and uniform LiDAR point clouds can
only be captured by high-precision sensors which need high budgets. Therefore, point
cloud upsampling is necessary to derive a dense, complete, and uniform point cloud from
a noisy, sparse, and non-uniform one.
To address this challenge, we proposed a methodology of utilizing point cloud upsam pling methods to enhance the object detection results of low-density point clouds in this
thesis. Specifically, we conduct three point cloud upsampling methods, including PU-Net,
3PU, and PU-GCN, on two datasets, which are a dataset we collected on our own in an
underground parking lot located at Highland Square, Kitchener, Canada, and SUN-RGBD.
We adopt VoteNet as the object detection network. We subsampled the datasets to get
a low-density dataset to stimulate the point cloud captured by the low-budget sensors.
We evaluated the proposed methodology on two datasets, which are SUN RGB-D and
the collected underground parking lot dataset. PU-Net, 3PU, and PU-GCN increase the
mean Average Precision (under the threshold of 0.25) by 18.8%,18.0%, and 18.7% on the
underground parking lot dataset and 9.8%, 7.2%, and 9.7% on SUN RGB-D. | en |
