面向动态场景的语义视觉里程计

卢金; 刘宇红; 张荣芬

doi:doi:10.3788/LOP202158.0611001

激光与光电子学进展, 2021, 58 (6): 0611001, 网络出版: 2021-03-06

面向动态场景的语义视觉里程计下载： 788次

Semantic-Based Visual Odometry Towards Dynamic Scenes

卢金刘宇红张荣芬 ^*

作者单位

贵州大学大数据与信息工程学院, 贵州贵阳 550025

图 & 表

图 1. 本文视觉里程计的系统结构

Fig. 1. System structure of proposed visual odometry

下载图片查看原文

图 2. SegNet网络架构

Fig. 2. Network architecture of SegNet

下载图片查看原文

图 3. 语义分割效果。(a) 原图;(b) 分割结果

Fig. 3. Results of semantic segmentation. (a) Original image; (b) segmentation result

下载图片查看原文

图 4. 不同尺度RANSAC的效果对比。(a)尺度为1和0.3,内点数量为238;(b)尺度为1和0.2,内点数量为210;(c)尺度为1和0.1,内点数量为159;(d)尺度为0.1,内点数量为168

Fig. 4. RANSAC effect comparison under different scales. (a) Scale is 1 and 0.3, number of inliers is 238; (b) scale is 1 and 0.2, number of inliers is 210; (c) scale is 1 and 0.1, number of inliers is 159; (d) scale is 0.1, number of inliers is 168

下载图片查看原文

图 5. 对极几何约束。(a)p₂位于极线L₂上;(b)p₂不严格位于极线L₂上

Fig. 5. Epipolar geometric constraints. (a) p₂ is on the polar line L₂; (b) p₂ is not strictly on the polar line L₂

下载图片查看原文

图 6. 外点剔除。(a)提取了所有的特征点;(b)剔除了位于人体身上的外点

Fig. 6. Outliers removing. (a) All feature points are extracted; (b) outliers lie on people are removed

下载图片查看原文

图 7. 绝对路径轨迹误差对比。(a)ORB-SLAM2;(b)本文方案

Fig. 7. Comparison of ATE. (a) ORB-SLAM2; (b) proposed scheme

下载图片查看原文

图 8. 相对平移误差对比。(a)ORB-SLAM;(b)本文方案

Fig. 8. Comparison of relative translation error. (a) ORB-SLAM2; (b) proposed scheme

下载图片查看原文

表 1绝对路径轨迹误差典型值

Table1. Typical value of ATE

Sequency	ORB-SLAM2 /m				Proposed /m				Improvement /%
Sequency	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.
wX	0.565505	0.528696	0.515921	0.200691	0.019148	0.015891	0.013613	0.010682	96.61	96.99	97.36	94.68
wH	0.327989	0.275986	0.232572	0.177225	0.028845	0.024179	0.020507	0.015730	91.21	91.24	91.18	91.12
wR	0.817879	0.695593	0.642242	0.430206	0.407781	0.352362	0.270319	0.205247	50.14	49.34	57.91	52.29
wS	0.409268	0.369913	0.293660	0.175114	0.007302	0.006431	0.006031	0.003459	98.22	98.26	97.95	98.02
sX	0.009275	0.007939	0.007251	0.004796	0.009962	0.008540	0.007845	0.005129	-7.41	-7.57	-8.19	-6.94
sH	0.027882	0.024288	0.022784	0.013692	0.014589	0.012853	0.011611	0.006902	47.68	47.08	49.04	49.59
sR	0.021513	0.016177	0.011756	0.014181	0.016531	0.012956	0.009905	0.010268	23.16	19.91	15.75	27.59
sS	0.007698	0.006775	0.006045	0.003655	0.006142	0.005233	0.004683	0.003216	20.21	22.76	22.53	12.01

查看原文

表 2相对平移误差典型值

Table2. Typical value of relative translation error

Sequency	ORB-SLAM2 /m				Proposed /m				Improvement /%
Sequency	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.
wX	0.825981	0.692976	0.648657	0.449478	0.028086	0.023557	0.020073	0.015293	96.60	96.60	96.91	96.60
wH	0.502363	0.403237	0.436454	0.299615	0.040362	0.035091	0.032131	0.019942	91.97	91.30	92.64	93.34
wR	1.212279	1.006164	0.943469	0.676205	0.138982	0.087917	0.042314	0.102784	88.54	91.26	95.52	84.80
wS	0.585281	0.403727	0.157596	0.423743	0.010569	0.009465	0.008933	0.004703	98.19	97.66	94.33	98.89
sX	0.013602	0.011845	0.010865	0.006688	0.014602	0.012807	0.011729	0.007015	-7.35	-8.12	-7.95	-4.89
sH	0.040732	0.033476	0.028965	0.023205	0.020813	0.018533	0.016983	0.009471	48.90	44.64	41.37	59.19
sR	0.030898	0.025071	0.020812	0.018059	0.024480	0.020617	0.017343	0.013200	20.77	17.77	16.67	26.91
sS	0.012007	0.010637	0.009772	0.005570	0.009133	0.008002	0.007147	0.004403	23.94	24.77	26.86	20.95

查看原文

表 3相对旋转误差典型值

Table3. Typical value of relative rotation error

Sequency	ORB-SLAM2 /(°)							Proposed /(°)								Improvement /%
Sequency	RMSE	Mean		Median		S.D.		RMSE		Mean		Median		S.D.		RMSE		Mean		Median		S.D.
wX	14.812930		12.411189		11.075135		8.086117		0.717936		0.554291		0.449855		0.456282		95.15		95.53		95.94		94.36
wH	13.379170		11.226538		14.220488		7.277847		0.930220		0.806664		0.727123		0.463252		93.05		92.81		94.89		93.63
wR	22.021472		17.877791		16.175450		12.858064		2.780387		1.798397		0.922173		2.013343		87.37		89.94		94.30		84.34
wS	10.334787		7.085263		1.830560		7.523754		0.286392		0.257924		0.242004		0.124482		97.23		96.36		86.78		98.35
sX	0.578052		0.494635		0.422254		0.299133		0.590790		0.509834		0.442760		0.298500		-2.20		-3.07		-4.86		0.21
sH	1.030726		0.924055		0.857998		0.456638		0.716727		0.644671		0.600926		0.313205		30.46		30.23		29.96		31.41
sR	0.882169		0.767921		0.700110		0.434188		0.755252		0.670573		0.623672		0.347473		14.39		12.68		10.92		19.97
sS	0.336292		0.303505		0.286266		0.144834		0.316251		0.283073		0.264374		0.141012		5.96		6.73		7.65		2.64

查看原文

表 4绝对运动轨迹误差典型值

Table4. Typical value of ATE

Sequency	DS-SLAM /m				Proposed /m				Improvement /%
Sequency	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.
wX	0.022180	0.016869	0.013272	0.014402	0.019148	0.015891	0.013613	0.010682	13.67	5.80	-2.57	25.83
wH	0.032083	0.026748	0.022648	0.017715	0.028845	0.024179	0.020507	0.015730	10.09	9.60	9.45	11.21
wR	0.433820	0.368918	0.249150	0.228252	0.407781	0.352362	0.270319	0.205247	6.00	4.49	-8.50	10.08
wS	0.007709	0.006979	0.006576	0.003275	0.007302	0.006431	0.006031	0.003459	5.28	7.85	8.29	-5.62
sX	0.010339	0.008831	0.007981	0.005377	0.009962	0.008540	0.007845	0.005129	3.65	3.30	1.70	4.61
sH	0.014816	0.013229	0.011732	0.006672	0.014589	0.012853	0.011611	0.006902	1.53	2.84	1.03	-3.45
sR	0.020242	0.015779	0.011601	0.012680	0.016531	0.012956	0.009905	0.010268	18.33	17.89	14.62	19.02
sS	0.006142	0.005233	0.004683	0.003216	0.006273	0.005461	0.004728	0.003085	-2.13	-4.36	-0.96	4.07

查看原文

表 5相对平移误差典型值

Table5. Typical value of translation

Sequency	DS-SLAM /m				Proposed /m				Improvement /%
Sequency	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.
wX	0.032488	0.025585	0.020938	0.020022	0.028086	0.023557	0.020073	0.015293	13.55	7.93	4.13	23.62
wH	0.045461	0.039412	0.035235	0.022658	0.040362	0.035091	0.032131	0.019942	11.22	10.96	8.81	11.99
wR	0.148749	0.094158	0.045830	0.112832	0.138982	0.087917	0.042314	0.102784	6.57	6.63	7.67	8.91
wS	0.010977	0.009989	0.009499	0.004552	0.010569	0.009465	0.008933	0.004703	3.72	5.25	5.96	-3.32
sX	0.014969	0.013095	0.012071	0.007252	0.014602	0.012807	0.011729	0.007015	2.45	2.20	2.83	3.27
sH	0.021379	0.019180	0.017768	0.009444	0.020813	0.018533	0.016983	0.009471	2.65	3.37	4.42	-0.29
sR	0.028873	0.024268	0.020204	0.015643	0.024480	0.020617	0.017343	0.013200	15.21	15.04	14.16	15.62
sS	0.009217	0.008144	0.007363	0.004316	0.009133	0.008002	0.007147	0.004403	0.91	1.74	2.93	-2.02

查看原文

表 6相对旋转误差典型值

Table6. Typical value of rotation

Sequency	DS-SLAM /(°)				Proposed /(°)				Improvement /%
Sequency	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.	RMSE	Mean	Median	S.D.
wX	0.768973	0.583035	0.460229	0.501387	0.717936	0.554291	0.449855	0.456282	6.64	4.93	2.25	9.00
wH	0.983289	0.862411	0.769254	0.472340	0.930220	0.806664	0.727123	0.463252	5.40	6.46	5.48	1.92
wR	3.013413	1.909141	0.996452	2.320810	2.780387	1.798397	0.922173	2.013343	7.73	5.80	7.45	13.25
wS	0.285163	0.261389	0.250628	0.113990	0.286392	0.257924	0.242004	0.124482	-0.43	1.33	3.44	-9.20
sX	0.577467	0.493059	0.420298	0.300600	0.590790	0.509834	0.442760	0.298500	-2.31	-3.40	-5.34	0.70
sH	0.778858	0.698167	0.649726	0.345227	0.716727	0.644671	0.600926	0.313205	7.98	7.66	7.51	9.28
sR	0.863546	0.760713	0.701008	0.408689	0.755252	0.670573	0.623672	0.347473	12.54	11.85	11.03	14.98
sS	0.308551	0.276659	0.259871	0.136614	0.316251	0.283073	0.264374	0.141012	-2.50	-2.32	-1.73	-3.22

查看原文

表 7运动一致性检测时间消耗

Table7. Time consuming of moving consistency check

Sequency	DS-SLAM /ms	Proposed /ms	Reduced /%
wX	0.019487	0.013030	33.14
wH	0.018344	0.010929	40.42
wR	0.017183	0.010957	36.23
wS	0.016712	0.015695	6.09
sX	0.017424	0.014163	18.72
sH	0.018717	0.013182	29.58
sR	0.016412	0.011832	27.90
sS	0.014110	0.013517	4.20

查看原文

卢金, 刘宇红, 张荣芬. 面向动态场景的语义视觉里程计[J]. 激光与光电子学进展, 2021, 58(6): 0611001. Lu Jin, Liu Yuhong, Zhang Rongfen. Semantic-Based Visual Odometry Towards Dynamic Scenes[J]. Laser & Optoelectronics Progress, 2021, 58(6): 0611001.

面向动态场景的语义视觉里程计下载： 788次

图 1. 本文视觉里程计的系统结构

Fig. 1. System structure of proposed visual odometry

图 2. SegNet网络架构

Fig. 2. Network architecture of SegNet

图 3. 语义分割效果。(a) 原图;(b) 分割结果

Fig. 3. Results of semantic segmentation. (a) Original image; (b) segmentation result

图 4. 不同尺度RANSAC的效果对比。(a)尺度为1和0.3,内点数量为238;(b)尺度为1和0.2,内点数量为210;(c)尺度为1和0.1,内点数量为159;(d)尺度为0.1,内点数量为168

Fig. 4. RANSAC effect comparison under different scales. (a) Scale is 1 and 0.3, number of inliers is 238; (b) scale is 1 and 0.2, number of inliers is 210; (c) scale is 1 and 0.1, number of inliers is 159; (d) scale is 0.1, number of inliers is 168

图 5. 对极几何约束。(a)p₂位于极线L₂上;(b)p₂不严格位于极线L₂上

Fig. 5. Epipolar geometric constraints. (a) p₂ is on the polar line L₂; (b) p₂ is not strictly on the polar line L₂

图 6. 外点剔除。(a)提取了所有的特征点;(b)剔除了位于人体身上的外点

Fig. 6. Outliers removing. (a) All feature points are extracted; (b) outliers lie on people are removed

图 7. 绝对路径轨迹误差对比。(a)ORB-SLAM2;(b)本文方案

Fig. 7. Comparison of ATE. (a) ORB-SLAM2; (b) proposed scheme

图 8. 相对平移误差对比。(a)ORB-SLAM;(b)本文方案

Fig. 8. Comparison of relative translation error. (a) ORB-SLAM2; (b) proposed scheme

表 1绝对路径轨迹误差典型值

Table1. Typical value of ATE

表 2相对平移误差典型值

Table2. Typical value of relative translation error

表 3相对旋转误差典型值

Table3. Typical value of relative rotation error

表 4绝对运动轨迹误差典型值

Table4. Typical value of ATE

表 5相对平移误差典型值

Table5. Typical value of translation

表 6相对旋转误差典型值

Table6. Typical value of rotation

表 7运动一致性检测时间消耗

Table7. Time consuming of moving consistency check

关于本站 Cookie 的使用提示

全站搜索

面向动态场景的语义视觉里程计 下载： 788次

图 1. 本文视觉里程计的系统结构

Fig. 1. System structure of proposed visual odometry

图 2. SegNet网络架构

Fig. 2. Network architecture of SegNet

图 3. 语义分割效果。(a) 原图;(b) 分割结果

Fig. 3. Results of semantic segmentation. (a) Original image; (b) segmentation result

图 4. 不同尺度RANSAC的效果对比。(a)尺度为1和0.3,内点数量为238;(b)尺度为1和0.2,内点数量为210;(c)尺度为1和0.1,内点数量为159;(d)尺度为0.1,内点数量为168

Fig. 4. RANSAC effect comparison under different scales. (a) Scale is 1 and 0.3, number of inliers is 238; (b) scale is 1 and 0.2, number of inliers is 210; (c) scale is 1 and 0.1, number of inliers is 159; (d) scale is 0.1, number of inliers is 168

图 5. 对极几何约束。(a)p2位于极线L2上;(b)p2不严格位于极线L2上

Fig. 5. Epipolar geometric constraints. (a) p2 is on the polar line L2; (b) p2 is not strictly on the polar line L2

图 6. 外点剔除。(a)提取了所有的特征点;(b)剔除了位于人体身上的外点

Fig. 6. Outliers removing. (a) All feature points are extracted; (b) outliers lie on people are removed

图 7. 绝对路径轨迹误差对比。(a)ORB-SLAM2;(b)本文方案

Fig. 7. Comparison of ATE. (a) ORB-SLAM2; (b) proposed scheme

图 8. 相对平移误差对比。(a)ORB-SLAM;(b)本文方案

Fig. 8. Comparison of relative translation error. (a) ORB-SLAM2; (b) proposed scheme

表 1绝对路径轨迹误差典型值

Table1. Typical value of ATE

表 2相对平移误差典型值

Table2. Typical value of relative translation error

表 3相对旋转误差典型值

Table3. Typical value of relative rotation error

表 4绝对运动轨迹误差典型值

Table4. Typical value of ATE

表 5相对平移误差典型值

Table5. Typical value of translation

表 6相对旋转误差典型值

Table6. Typical value of rotation

表 7运动一致性检测时间消耗

Table7. Time consuming of moving consistency check

相关论文

相关资讯

关于本站 Cookie 的使用提示

全站搜索

面向动态场景的语义视觉里程计下载： 788次

图 5. 对极几何约束。(a)p₂位于极线L₂上;(b)p₂不严格位于极线L₂上

Fig. 5. Epipolar geometric constraints. (a) p₂ is on the polar line L₂; (b) p₂ is not strictly on the polar line L₂