Multiple Frame Point Cloud Object Detection Using Feature Fusion
Date
2023-09-21
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Object detection has important applications in the field of autonomous driving. While 2D object detection lacks depth information and suffers from the scale problems of objects, 3D point cloud object detection is a more suitable alternative for autonomous driving. However, due to the character of point cloud data, sparsity is an inevitable problem in point cloud object detection, and single-frame methods still have their limitations. This research investigates using multiple frame point cloud data to enhance detection precision. Facing the sparsity challenge in point cloud data, single-frame detection methods perform detection independently without all the data together to improve detection. However, with proper fusion scheme, using multiple frame data can be similar to using denser lidar. In this work, multiple association and fusion methods are discussed. And finally a probability-based proposal feature fusion (PPFF) module is proposed. Chapter 3 proposes a two-stage point cloud object detection method using proposal feature fusion. Proposal association methods and feature fusion methods are discussed. In Chapter 4, the MHT tracker is discussed, and in Chapter 5, combined with the knowledge of MHT, a probability-based proposal feature fusion (PPFF) module for multiple-frame point cloud object detection is proposed. We compared our methods to current multiple-frame methods to show the improvement and effectiveness of our methods.
Description
Keywords
autonomous driving, 3D object detection, multiple frame point clouds, feature and data fusion
Citation
Huang, M. (2023). Multiple frame point cloud object detection using feature fusion (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.