Augmenting 3-D Object Detection Through Data Uncertainty-Driven Auxiliary Framework

Jianyu Wang

doi:10.1109/TIM.2024.3398077

Return

Augmenting 3-D Object Detection Through Data Uncertainty-Driven Auxiliary Framework

2024-01-01

PRE

DOI：10.1109/TIM.2024.3398077

Original

Original request for help

Save

Abstract

En 中文

Three-dimensional object detection algorithms are becoming increasingly crucial in autonomous driving, demanding high accuracy, and reasoning speed. Due to the uneven and extremely sparse distribution of point clouds in current point-cloud datasets, interference from background information is significant, potentially leading to fatal errors during driving. Therefore, existing 3-D object detection techniques in autonomous vehicles require a method to suppress false positive (FP) and false negative (FN) samples. Due to issues inherent in the dataset itself, data augmentation and dataset expansion are ineffective in resolving the problem of predicting object errors. To address these challenges, this study introduces a universal auxiliary framework, data uncertainty assimilator (DUA), for 3-D object detection. By applying the DUA framework based on data uncertainty for predicting classification and regression uncertainties, the number of projected FP samples can be minimized, and classification sensitivity enhanced. Bounding box regression can be predicted in a more reasonable manner. The methods for classification loss and regression loss are redesigned to improve detection accuracy. Additionally, a data uncertainty-based object filtering adjuster is designed. This added strategy can be flexibly employed with cutting-edge detectors to enhance model accuracy while largely preserving parameters in their original states. Experiments on the KITTI and Waymo Open datasets demonstrate that after inserting DUA, mainstream frameworks show a maximum improvement of 6.27% mAP in overall prediction performance, with an unavoidable average efficiency decrease of approximately 1.2 frames/s.

Keywords:

Three-dimensional displays

Point cloud compression

Uncertainty

Feature extraction

Object detection

Predictive models

Computational modeling

Autonomous driving

auxiliary framework

object detection

point cloud