IEEE Consumer Electronics Magazine - September/October 2022 - 48

A Single Stage Monocular 3D Object Detection
asa single internal network.Weexploit thecorrelationbetween 2-Dand3-Ddetection
spaces, enabling 3-Dboxes to leverage featuremapsgenerated in imagespace. The 2-Dand
3-Dproposals areextracted througha proposal generationnetwork that is enhancedand
utilized for estimatingaccurate 3-Ddetectionand localization. Experimental results on the
KITTI dataset demonstrate that in comparison to othermonocularobjectdetectionmethods
the proposedmethodconsiderably improvedtheaccuracy of 3-Dobject detection.Themean
average precision of 3-Dobject detection in front viewis improved to25%and the bird's eye
viewto32%for thecar classonamoderate difficulty level.
& OBJECT DETECTION AND recognition play a crucial
role in designing effective, computationally
less expensive, and real-time 3-D solutions for
autonomous vehicles and robotic vision.1,2 The
existing 3-D object detection methods depend
on 3-D data sensors, such as point clouds from
sparse depth maps, which utilize expensive laser
sensors (lidar), depth cameras, and stereo cameras.3
The drawbacks of these sensors are the
limited range of depth, synchronization, calibration
of stereo cameras, and the expensive cost of
laser sensors, especially the high-resolution
lidar for detecting small objects. The previous
monocular-based methods used geometric constraints,
disentanglement transformation, and
depth-aware convolution for the prediction of 3D
objects.4-6 Most 3-D object detection methods
based on monocular or stereoscopic images are
too slow for a real-time requirement and the
overall accuracy is far lower than other methods
using 3-D data sensors. This motivates us to
improve the low accuracy and reduce the complex
processing steps in existing state-of-the-art
methods.
In this article, we propose a single end-to-end
deep learning model to simultaneously predict
7-tuple 3-D bounding box parameters (location,
height, width, length, and orientation angle of
the object) and 4-tuple 2-D (location, height, and
width) for objects. 2-D object detection can provide
enough information to improve and guide
the performance of 3-D object detection utilizing
a single internal deep learning model. Therefore,
3-D object detection utilizes both 3-D and 2-D
spaces for classification tasks. The contribution
of our work is as follows:
We formulated a 3-D regional proposal network
(RPN) that utilizes the geometric
constraints of both 3-D and 2-D detection
spaces to efficiently predict 3-D objects.
Twelve 3-D anchor boxes are precomputed
with different scales and aspect ratios.
We used a discriminative learning technique
to train the backbone network.
Compared to the state-of-the-art methods,
the proposed method shows significant performance
improvement on the KITTI dataset.
RELATED PRIOR WORKS
The work presented in this article is related
to 3-D object detection and monocular 3-D object
detection. Therefore, we summarize the related
work into the following two categories.
3-D Object Detection
3-D object detection can be defined as finding
or identifying an object in a 3-D view from
images, video sequences, or 3-D data. State-ofthe-art
3-D object detection methods are based
on 3-D data such as lidar point clouds and depth
maps. These methods are either utilizing lidar
point clouds only or both lidar and camera
images. Some methods proposed two subnetworks,
one for 3-D object proposal generation
and the other for multiview feature fusion utilizing
both lidar and camera images. Many methods
utilize stereo vision for estimating 3-D depth
information. The other methods use stereo
images to generates object proposals through a
stereo RPN to predict objects in the corresponding
left and right images simultaneously, and
pixel-level stereo matching is applied to refine 3D
bounding boxes. Stereo images provide essential
information than monocular images, but the
installation and computation of stereo cameras
have a higher requirement than monocular cameras,
which are more feasible for real-time cases.
48
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - September/October 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - September/October 2022
