Depth Correction++ for pseudo-LiDAR

Qiu, Yujia

Permalink

https://hdl.handle.net/2142/109452 Copy

Description

Title

Depth Correction++ for pseudo-LiDAR

Author(s)

Qiu, Yujia

Issue Date

2020-12-09

Director of Research (if dissertation) or Advisor (if thesis)

Chen, Deming

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Object detection
• pseudo-Lidar
• depth correction

Abstract

In recent years, objection detection is one of the most important tasks in the autonomous driving area. There are many research studies done for 2D object detection that have achieved very high detection accuracy. However, the recent trend is 3D object detection. 3D object detection can provide more information from the 3D bounding boxes, compared to 2D bounding boxes, which can better describe surrounding obstacles with details, providing further help with route planning of autonomous vehicles. Researchers commonly use LiDAR data in 3D object detection, because it can provide accurate depth information. However, LiDAR sensors are expensive, which means it is not economical to use LiDAR sensors and data in real life. How- ever, using only cameras for 3D object detection has low accuracy. This resulted in the introduction of the pseudo-LiDAR, which uses stereo images as input and generates a fake point cloud. Then this pseudo-LiDAR data is used instead of a real point cloud as an input for 3D object detection algorithms. After one year, pseudo-LiDAR++, which was based on the original pseudo-LiDAR algorithm, was published. Furthermore, it uses sparse real LiDAR data to correct the depth of generated LiDAR data. The pseudo-LiDAR algorithm only uses 2D image data, which lacks depth information, hence the accuracy is only a little higher than image-based systems. At the same time, LiDAR-based 3D object detection algorithms, which use full LiDAR data, do not fully utilize all LiDAR points, because many points are redundant. Sparse LiDAR data can provide accurate depth for a few points. However, only using sparse real LiDAR points as the input for 3D object detection cannot generate high detection accuracy, although it can be useful to adjust a pseudo-LiDAR point cloud. Pseudo-LiDAR++ uses K-nearest-neighbor (KNN) to classify the 3D points, and then uses a linear system to adjust the depth estimation. Based on this algorithm, we proposed a new algorithm “Depth Correction++” (DC++) that is able to further correct the depth information. DC++ focuses on pedestrians and cyclists. It uses the closest real LiDAR point to adjust the predicted depth of a target point, which can effectively decrease the deviation of a small object’s predicted depth. Therefore, the algorithm enhances the 3D object detection accuracy.

Graduation Semester

2020-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/109452

Copyright and License Information

Copyright 2020 Yujia Qiu

Owning Collections