不确定性建模在2D和3D目标检测中的应用

doi:10.12305/j.issn.1001-506X.2023.08.10

摘要/Abstract

摘要：

目标检测算法在自动驾驶领域有着不可或缺的地位, 其检测精度和速度往往可以作为评判一个自动驾驶系统好坏的标准。如何提升目标检测精度和速度已成为当前目标检测算法的主要研究方向。对此, 提出了一种基于不确定性建模的目标检测改进算法, 在原有二维单次多柜检测器上通过对物体的边界框进行高斯建模并引入新的位置损失函数, 实现对原有检测结果的微调。同时, 在原有单阶单目三维检测器的基础上引入深度以及航向角的不确定性来调整目标在热力图中的高斯半径, 并用马氏距离替代原有的L1距离以提升对较远目标和斜向目标的识别率。对比实验证明, 改进后的2D和3D检测算法在KITTI自动驾驶数据集上对比原始2D和3D算法, 检测精度分别提升了近5%和2%。

关键词: 深度学习, 目标检测, 自动驾驶, 不确定性建模

Abstract:

Object detection algorithm plays an indispensable role in the field of autopilot. Its detection accuracy and speed can often be used as the standard to judge the quality of an auto drive system. How to improve the accuracy and speed of object detection has become the main research direction of the current object detection algorithm. Therefore, an improved object detection algorithm based on uncertainty modeling is proposed. Based on the original two-dimensional (2D) single shot multibox detector (SSD) object detection algorithm, the original detection results are finely tuned by Gaussian modeling the boundary box of the object and introducing a new position loss function. At the same time, based on the original three-dimensional (3D) single-stage monocular (SMOKE) object detection algorithm, the uncertainty of depth and heading angle is introduced to adjust the Gaussian radius of the target in the heatmap, and the Mahalanobis distance is used to replace the original L1 distance to improve the recognition rate of distant targets and oblique targets. Comparative experiments proves that the detection accuracy of the improved 2D and 3D detection algorithms is improved by nearly 5% and 2% respectively compared with the original 2D and 3D algorithms on the KITTI autonomous driving dataset.

Key words: deep learning, object detection, autonomous driving, uncertainty modeling

中图分类号:

TP391.41

汪萌, 诸兵. 不确定性建模在2D和3D目标检测中的应用[J]. 系统工程与电子技术, 2023, 45(8): 2370-2376.

Meng WANG, Bing ZHU. Application of uncertainty modeling in 2D and 3D object detection[J]. Systems Engineering and Electronics, 2023, 45(8): 2370-2376.

图/表 11

图1

图2

图3

图4

表1

图5

表2

表3

表4

表5

图6

参考文献 21

1	KIRAN B R , SOBH I , TALPAERT V , et al. Deep reinforcement learning for autonomous driving: a survey[J]. IEEE Trans.on Intelligent Transportation Systems, 2021, 99 (2): 1- 18.
2	VAFAEIPOUR M , RAHBARI O , ROSEN M A , et al. Application of sliding window technique for prediction of wind velocity time series[J]. International Journal of Energy and Environmental Engineering, 2014, 5 (2): 1- 7.
3	UIJLINGS J R R , VAN DE SANDE K E A , GEVERS T , et al. Selective search for object recognition[J]. International Journal of Computer Vision, 2013, 104 (2): 154- 171. doi: 10.1007/s11263-013-0620-5
4	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2014: 580-587.
5	GIRSHICK R. Fast R-CNN[C]//Proc. of the IEEE International Conference on Computer Vision, 2015: 1440-1448.
6	REN S Q , HE K M , GIRSHICK R , et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Trans.on Pattern Analysis and Machine Intelligence, 2017, 39 (6): 1137- 1149. doi: 10.1109/TPAMI.2016.2577031
7	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]//Proc. of the European Conference on Computer Vision, 2016: 21-37.
8	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 779-788.
9	IANDOLA F N, HAN S, MOSKEWICZ M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size[EB/OL]. [2022-06-17]. https://arxiv.org/abs/1602.07360.
10	HOWARD A G, ZHU M L, CHEN B, et al. Mobilenets: efficient convolutional neural networks for mobile vision applications[EB/OL]. [2022-06-17]. https://arxiv.org/abs/1704.04861.
11	MOUSAVIAN A, ANGUELOV D, FLYNN J, et al. 3D bounding box estimation using deep learning and geometry[C]// Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 7074-7082.
12	LI B Y, OUYANG W L, SHENG L, et al. Gs3D: an efficient 3D object detection framework for autonomous driving[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 1019-1028.
13	QIN Z Y, WANG J L, LU Y. Monogrnet: a geometric reasoning network for monocular 3d object localization[C]//Proc. of the AAAI Conference on Artificial Intelligence, 2019, 33(1): 8851-8858.
14	ZHOU X Y, WANG D Q, KRÄHENBVHL P. Objects as points[EB/OL]. [2022-06-17]. https://arxiv.org/abs/1904.07850.
15	CHOI J, CHUN D, KIM H, et al. Gaussian YOLOv3: an accurate and fast object detector using localization uncertainty for autonomous driving[C]//Proc. of the IEEE/CVF International Conference on Computer Vision, 2019: 502-511.
16	HE Y H, ZHANG X Y, SAVVIDES M, et al. Softer-NMS: rethinking bounding box regression for accurate object detection[EB/OL]. [2022-06-17]. https://arxiv.org/abs/1809.08545v1.
17	LU Y, MA X Z, YANG L, et al. Geometry uncertainty projection network for monocular 3d object detection[C]//Proc. of the IEEE/CVF International Conference on Computer Vision, 2021: 3111-3121.
18	LIU Z C, WU Z Z, TÓTH R. SMOKE: single-stage monocular 3D object detection via keypoint estimation[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020: 996-997.
19	GEIGER A , LENZ P , STILLER C , et al. Vision meets robo-tics: the kitti dataset[J]. The International Journal of Robotics Research, 2013, 32 (11): 1231- 1237. doi: 10.1177/0278364913491297
20	XU B, CHEN Z Z. Multi-level fusion based 3D object detection from monocular images[C]//Proc. of the IEEE/CVF Confe-rence on Computer Vision and Pattern Recognition, 2018.
21	MOUSAVIAN A, ANGUELOV D, FLYNN J, et al. 3D Bounding box estimation using deep learning and geometry[C]Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.

模型	汽车-AP/%	行人-AP/%	自行车-AP/%	货车-AP/%	mAP/%	FPS
SSD	85.7	41.6	51.4	86.8	66.4	125.7
SSD+不确定性	88.9	46.8	59.9	89.7	71.4	103.8

模型	车	行人	自行车
SMOKE	10.27	2.35	3.48
SMOKE+不确定性	12.08	1.96	3.56

模型	简单-AP%	中等-AP%	困难-AP%	FPS
SMOKE	13.29	9.55	7.26	32.30
SMOKE+不确定性	14.52	10.57	8.76	30.50
文献[20]	10.00	7.70	5.30	55.20
文献[21]	10.50	5.70	5.40	-

模型	简单-AP%	中等-AP%	困难-AP%	FPS
SMOKE	19.65	14.77	14.32	32.30
SMOKE+不确定性	21.28	15.16	14.45	30.50
文献[20]	10.00	7.70	5.30	55.20
文献[21]	22.00	13.60	11.60	-

模型	简单-AP%	中等-AP%	困难-AP%	FPS
SMOKE	87.25	82.61	65.45	32.30
SMOKE+不确定性	86.79	83.27	68.72	30.50