DyFusion:基于动态融合的交叉注意三维目标检测 论文解析

106 IEEE LATIN AMERICA TRANSACTIONS, VOL. 22, NO. 2, FEBRUARY 2024

DyFusion: Cross-Attention 3D Object Detection

with Dynamic Fusion

Jiangfeng Bi , Haiyue Wei , Guoxin Zhang , Kuihe Yang , and Ziying Song

Abstract—In the realm of autonomous driving, LiDAR and

camera sensors play an indispensable role, furnishing pivotal

observational data for the critical task of precise 3D object

detection. Existing fusion algorithms effectively utilize the

complementary data from both sensors. However, these methods

sources. In addition, we acknowledge the issue of insufficient

data augmentation for these two diverse modalities. To combat

this, we propose a Synchronous Data Augmentation (SDA)

strategy designed to enhance training efficiency. We have tested

our method using the KITTI and nuScenes datasets, and the

results have been promising. Remarkably, our top-performing

model attained an 82.52% mAP on the KITTI test benchmark,

outperforming other state-of-the-art methods.

Link to graphical and video abstracts, and to code:

https://latamt.ieeer9.org/index.php/transactions/article/view/8434

task to equip unmanned vehicles with precise environmental

cognition [1]. Pioneers in 3D object detection have carried out

significant research, demonstrating excellent performance on

public datasets such as KITTI [2] and nuScenes [3]. As an

exemplar pioneering work, Qi et al. [4] devised PointNet, an

innovative deep neural architecture that directly learns global

features from point cloud data. Zhou et al. [5] proposed

VoxelNet with a Voxel Feature Encoder, which transforms

raw point clouds into voxel-wise features containing spatial

and physical information. These pioneering methods have laid

Jiangfeng Bi, Haiyue Wei, Guoxin Zhang, and Kuihe Yang are with

the School of Information Science and Engineering, Hebei Univer-

sity of Science and Technology, Shijiazhuang 050018, China (e-mail:

[email protected]; [email protected]; [email protected];

[email protected]).

Ziying Song is with the School of Computer and Information Technology,

Beijing Key Lab of Traffic Data Analysis and Mining, Beijing Jiaotong

University, Beijing, China. (e-mail: [email protected]).

Fig. 1. Quantitative Analysis of Camera, LiDAR, Fusion and Pro-

posed Methods.

Sensitive to lighting and

reflective surfaces.

Lower resolution, diffi-

culty in recognizing color

and texture.

detection [1]. These two sensors have highly complementary

output data, with their respective advantages and disadvantages

integrates 2D object detection outputs with point cloud data.

Building on the foundation of MV3D [7], Ku et al. [8]

introduced AVOD, a fusion-based approach utilizing both

image and point cloud features to achieve more accurate 3D

object detection.

Point cloud data contains the coordinate information of

3D spatial points in a scene, providing high precision and

reliability. Image data provides high-resolution information