Citation: Xu, W.; Hu, J.; Chen, R.; An,
Y.; Xiong, Z.; Liu, H. Keypoint-Aware
Single-Stage 3D Object Detector for
Autonomous Driving. Sensors 2022,
22, 1451. https://doi.org/10.3390/
s22041451
Academic Editors: Yangquan Chen,
Subhas Mukhopadhyay,
Nunzio Cennamo, M. Jamal Deen,
Junseop Lee and Simone Morais
Received: 9 January 2022
Accepted: 11 February 2022
Published: 14 February 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Keypoint-Aware Single-Stage 3D Object Detector for
Autonomous Driving
Wencai Xu, Jie Hu *, Ruinan Chen, Yongpeng An, Zongquan Xiong and Han Liu
Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology,
Wuhan 430070, China; vehfish@whut.edu.cn (W.X.); chenruinan@whut.edu.cn (R.C.);
yongpeng20_@whut.edu.cn (Y.A.); zq.xiong@whut.edu.cn (Z.X.); liuhan77@whut.edu.cn (H.L.)
* Correspondence: auto_hujie@whut.edu.cn
Abstract:
Current single-stage 3D object detectors often use predefined single points of feature maps
to generate confidence scores. However, the point feature not only lacks the boundaries and inner
features but also does not establish an explicit association between regression box and confidence
scores. In this paper, we present a novel single-stage object detector called keypoint-aware single-
stage 3D object detector (KASSD). First, we design a lightweight location attention module (LLM),
including feature reuse strategy (FRS) and location attention module (LAM). The FRS can facilitate
the flow of spatial information. By considering the location, the LAM adopts weighted feature fusion
to obtain efficient multi-level feature representation. To alleviate the inconsistencies mentioned
above, we introduce a keypoint-aware module (KAM). The KAM can model spatial relationships
and learn rich semantic information by representing the predicted object as a set of keypoints. We
conduct experiments on the KITTI dataset. The experimental results show that our method has a
competitive performance with 79.74% AP on a moderate difficulty level while maintaining 21.8 FPS
inference speed.
Keywords:
3D single stage object detector; feature reuse strategy; location attention module;
keypoint-aware module
1. Introduction
Nowadays, object detection has become a fundamental task of scene understanding,
attracting much attention in various fields, such as autonomous vehicles and robotics. The
tasks include traffic sign detection [
1
–
3
], traffic light detection [
4
,
5
], 2D object detection [
6
],
and 3D objection detection [
7
,
8
], which rely on sensors installed on the autonomous vehicles.
Since LiDAR (light detection and ranging) can provide accurate distance information about
the surrounding environment and is not impacted under low-light conditions, it has become
one of the main sources of perception. The purpose of 3D object detection of LiDAR point
cloud is to predict the bounding box, classification, and direction, an essential job for
downstream perception and planning tasks.
Recently, 3D object detection methods based on deep learning have been widely
adopted, and achieved dramatic developments in industry and academia [
7
]. Despite huge
advantages, it is important to note that point clouds suffer some drawbacks: (1) The original
point cloud is sparse, while the image is dense; (2) Point cloud data have an unstructured
and unordered nature [
8
]; (3) Point cloud data are sensitive to occlusion and distance; (4) 3D
features introduce a heavy computational burden. Instead of learning feature for each point,
volumetric-based methods encode point clouds into regular 3D grids, called voxels, so as
to achieve robust representation and then apply a Convolution Neural Network (CNN)
for feature extraction and prediction instance object. Furthermore, a regular data format
can naturally transfer previous mature knowledge from the image domain. Although
the point cloud can reflect the real geometric structure and object size, the image may
suffer from these information losses. Thus, applying image methods directly may deliver
Sensors 2022, 22, 1451. https://doi.org/10.3390/s22041451 https://www.mdpi.com/journal/sensors
