Advanced ORB algorithm of image feature uniform distribution based on threshold adaptive

doi:10.3969/j.issn.1674-8484.2021.03.004

Abstract

Abstract:

Camera-based visual SLAM (simultaneous localization and mapping) for driverless vehicles could be used to complete the localization and mapping of driverless vehicles. In order to solve the problems that the traditional ORB (Oriented FAST and Rotated BRIEF) algorithm could easily cause clutter and concentration of distribution when extracting image feature points, an advanced ORB (A-ORB) algorithm was proposed which limited the splitting depth of quadtree algorithm. This algorithm constructed an image pyramid to solve the scale invariance problem. According to the total number of feature points extracted, the algorithm calculated the feature points that needed to be extracted for each layer of the pyramid. The image of each pyramid layer was divided into adaptive regions and the threshold of feature point extraction was calculated according to the image information. This algorithm used improved quadtree algorithm to homogenize the distribution feature points. The simulation experiment was carried out. The results show that compared with ORB, MA and S-ORB algorithms, the running efficiency of the advanced ORB is increased by more than 30%, and the matching accuracy is increased by more than 10%.

Key words: driverless cars, simultaneous localization and mapping (SLAM), movement information, ORB (Oriented FAST and Rotated BRIEF) algorithm, feature point extraction, image pyramid, uniform distribution

CLC Number:

TP391.4

SHI Peicheng, YANG Jianfeng, LIANG Taonian, QI Heng. Advanced ORB algorithm of image feature uniform distribution based on threshold adaptive[J]. Journal of Automotive Safety and Energy, 2021, 12(3): 305-313.

Figures/Tables 9

References 15

[1]	王霞, 左一凡. 视觉SLAM研究进展[J]. 智能系统学报, 2020, 15(5):825-834.
	WANG Xia, ZUO Yifan. Advances in visual SLAM research[J]. *CAAI (Chin Asso Arti Intel) Trans Intell Syst*, 2020, 15(5):825-834. (in Chinese)
[2]	姚晋晋, 张鹏超, 王彦, 等. 基于改进四叉树的ORB特征均匀分布算法[J]. 计算机工程与设计, 2020, 41(6):1629-1634.
	YAO Jinjin, ZHANG Pengchao, WANG Yan, et al. ORB feature uniform distribution algorithm based on improved quadtree[J]. *Computer Engi Des*, 2020, 41(6):1629-1634.(in Chinese)
[3]	Lowe D G. Distinctive image features from scale-invariant keypoints[J]. *Int’l J Computer Vision*, 2004, 60(2):91-110.
[4]	Bay H, Tuytelaars T, Gool L V. SURF: Speeded up robust features[J]. *Compu Visi Ima Understanding, 2008, 110*(3):346-359.
[5]	Rublee E, Rabaud V, Konolige K, et al. ORB: An efficient alternative to SIFT or SURF [C]// IEEE Intl Conf on Computer Vision, 2012: 2564-2571.
[6]	刘中晖. 基于KINECT传感器的图像特征点描述子算法研究[D]. 西安: 西安电子科技大学, 2014.
	LIU Zhonghui. Research on image feature descriptor based on KINECT sensor[D]. Xian: Xidian University, 2014. (in Chinese)
[7]	Raul M, Juan D T. ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras[J]. *IEEE T Robot*, 2017, 33(5):1255-1262. doi: 10.1109/TRO.2017.2705103 URL
[8]	Harris C G, Stephens M J. A combined corner and edge detector [C]// Alvey vision conf, 1988.
[9]	禹鑫燚, 詹益安, 朱峰, 等. 一种基于四叉树的改进的ORB特征提取算法[J]. 计算机科学, 2018, 45(S2):222-225.
	Yu Xinyi, ZHAN Yian, ZHU Feng, et al. Improved ORB features extraction algorithm based on quadtree encoding[J]. *Chin Computer Sci*, 2018, 45(S2):222-225. (in Chinese)
[10]	孙浩, 王朋. 一种基于区域划ORB算法[J]. 北京航空航天大学学报, 2020, 46(9):1763-1769.
	SUN Hao, WANG Peng. An improved ORB algorithm based on region division[J]. *J Beijing Univ Aeronaut Astronaut*, 2020, 46(9):1763-1769. (in Chinese)
[11]	潘盼. 一种改进的ORB算法[J]. 信息技术与网络安全, 2017, 36(20):23-26.
	PAN Pan. An improved ORB algorithm[J]. *Info Tech Net Sec*, 2017, 36(20):23-26. (in Chinese)
[12]	刘宏伟, 余辉亮, 梁艳阳. ORB特征四叉树均匀分布算法[J]. 自动化仪表, 2018, 39(5):52-54+59.
	LIU Hongwei, Yu Huiliang, LIANG Yanyang. ORB feature quadtree uniform distribution algorithm [J]. *Process Autom Instr*, 2018, 39(5):52-54+59. (in Chinese)
[13]	骆开庆, 杨坤, 张健, 等. 基于四叉树的改进BRISK特征提取算法[J]. 华南师范大学学报(自然科学版), 2020, 52(2):114-121.
	LUO Kaiqing, YANG Kun, ZHANG Jian, et al. The improved BRISK feature extraction algorithm based on quadtree[J]. *J South Chin Norm Univ (Nat Sci Ed)*, 2020, 52(2):114-121. (in Chinese)
[14]	王峰, 尤红建, 傅兴玉. 应用于SAR图像配准的自适应SIFT特征均匀分布算法[J]. 武汉大学学报(信息科学版), 2015, 40(2):159-163.
	WANG Feng, YOU Hongjian, FU Xingyu. Auto-adaptive well-distributed scale-invariant feature for SAR image registration[J]. *Geo Info Sci Wuhan Univ (Info Sci Ed)*, 2015, 40(2):159-163. (in Chinese)
[15]	杨弘凡, 李航, 陈凯阳, 等. 基于改进ORB算法的图像特征点提取与匹配方法[J]. 图学学报, 2020, 41(4):548-555.
	YANG Hongfan, LI Hang, CHENG Kaiyang, et al. Image feature points extraction and matching method based on improved ORB algorithm[J]. *J Graphics*, 2020, 41(4):548-55. (in Chinese)

图像数据集	均匀度,u				运行时间,t / ms
图像数据集	算法ORB	MA	S-ORB	A-ORB	算法ORB	MA	S-ORB	A-ORB
Bark	198.65	156.23	212.34	160.58	20.45	23.06	9.67	10.73
Bikes	201.54	162.35	205.62	170.64	22.36	26.36	10.35	12.56
Boat	196.25	140.35	200.36	157.23	19.56	25.41	9.96	13.56
Leuven	192.33	153.24	198.76	155.33	21.47	27.68	10.36	11.39
Ubc	168.56	126.58	189.62	155.67	18.36	24.59	9.37	10.95
Graf	202.36	145.69	210.56	165.69	20.67	27.53	10.69	14.72
平均值	193.28	147.40	202.86	160.86	20.47	25.77	10.06	12.32

图像数据集	均匀度,u				运行时间,t / ms
图像数据集	算法ORB	MA	S-ORB	A-ORB	算法ORB	MA	S-ORB	A-ORB
Bark	198.65	156.23	212.34	160.58	20.45	23.06	9.67	10.73
Bikes	201.54	162.35	205.62	170.64	22.36	26.36	10.35	12.56
Boat	196.25	140.35	200.36	157.23	19.56	25.41	9.96	13.56
Leuven	192.33	153.24	198.76	155.33	21.47	27.68	10.36	11.39
Ubc	168.56	126.58	189.62	155.67	18.36	24.59	9.37	10.95
Graf	202.36	145.69	210.56	165.69	20.67	27.53	10.69	14.72
平均值	193.28	147.40	202.86	160.86	20.47	25.77	10.06	12.32

图像数据集	算法	总匹配数	正确匹配数	正确率 / %	t / ms
Leuven	ORB	500	256	51.2	1.36
	MA	500	354	70.8	1.43
	S-ORB	506	324	64.0	1.23
	A-ORB	502	418	83.3	1.31
Bikes	ORB	500	282	56.4	1.41
	MA	500	363	72.6	1.53
	S-ORB	502	304	60.7	1.32
	A-ORB	504	413	81.9	1.33
Ubc	ORB	500	412	82.4	1.37
	MA	500	442	88.4	1.44
	S-ORB	506	398	78.7	1.23
	A-ORB	507	478	94.3	1.28
Bark	ORB	500	221	44.2	1.43
	MA	500	246	49.2	1.53
	S-ORB	508	197	38.8	1.24
	A-ORB	504	300	59.5	1.35
Boat	ORB	500	232	46.4	1.43
	MA	500	287	57.4	1.55
	S-ORB	502	267	53.2	1.29
	A-ORB	510	313	61.4	1.41
Graf	ORB	500	241	48.2	1.41
	MA	500	286	57.2	1.53
	S-ORB	509	239	47.0	1.31
	A-ORB	505	345	68.3	1.36

图像数据集	算法	总匹配数	正确匹配数	正确率 / %	t / ms
Leuven	ORB	500	256	51.2	1.36
	MA	500	354	70.8	1.43
	S-ORB	506	324	64.0	1.23
	A-ORB	502	418	83.3	1.31
Bikes	ORB	500	282	56.4	1.41
	MA	500	363	72.6	1.53
	S-ORB	502	304	60.7	1.32
	A-ORB	504	413	81.9	1.33
Ubc	ORB	500	412	82.4	1.37
	MA	500	442	88.4	1.44
	S-ORB	506	398	78.7	1.23
	A-ORB	507	478	94.3	1.28
Bark	ORB	500	221	44.2	1.43
	MA	500	246	49.2	1.53
	S-ORB	508	197	38.8	1.24
	A-ORB	504	300	59.5	1.35
Boat	ORB	500	232	46.4	1.43
	MA	500	287	57.4	1.55
	S-ORB	502	267	53.2	1.29
	A-ORB	510	313	61.4	1.41
Graf	ORB	500	241	48.2	1.41
	MA	500	286	57.2	1.53
	S-ORB	509	239	47.0	1.31
	A-ORB	505	345	68.3	1.36