Blank Slate Localization: Active Global Localization from Arbitrary Locations using Semantic Floor Plans

📄 Abstract

Floor plan based active global localization refers to the problem of localizing an agent in the floor plan while giving it control commands so as to efficiently minimize the localization error and time. It is generally formulated as an optimization problem where the agent's real-time sensor data is matched against the map to minimize a cost function. However, most existing methods either assume some {\em a priori} knowledge of the starting position of the agent or consider floor plans that are created using sensors the same as the ones on the agent. In this work, we propose Blank Slate Localization (BSL): a novel methodology for architectural floor plan based active localization for indoor navigation starting from completely unknown initial locations. Semantically aware {\em a priori} map based real-time goals are created for the agent to intelligently explore the local environment while simultaneously constructing a 2D semantic point cloud to globally localize the agent. To account for the time-varying odometry drift, a method is proposed to dynamically correct it based on the floor plan without knowing the global pose of the agent. Furthermore, a novel methodology for real-time bi-directional loop closure has been used. The efficacy of the proposed pipeline has been shown by conducting several experiments on indoor environments.

📊 Results

Mean and standard deviation of the distance traveled (meters) for successful localization

Method	Candidate creation (Mean)	Candidate creation (Std)	Candidate confirmation (Mean)	Candidate confirmation (Std)
Left wall follower	13.25	10.55	15.64	11.63
Random coin toss	15.03	13.96	17.62	15.66
Our method	11.27	9.31	13.25	10.36

Mean and standard deviation of the distance traveled (meters) from start to target

Method	Distance travelled (Mean)	Distance travelled (Std)
Min. possible distance	32.92	18.29
Left wall follower	52.03	20.88
Random coin toss	51.19	21.78
Our method	44.97	20.20

📑 Abalation Study

$R_1=L_4$, $R_2=L_{12}$, $R_3=L_{10}$, $R_4=L_{16}$, $R_5=L_{7}$ from the tables at the end.

Distance traveled for localization (in meters)

	$R_1$	$R_2$	$R_3$	$R_4$	$R_5$	Avg.
Our method	4.9	4.1	18.1	16.4	9.4	10.6
No Loop Closure	5.8	4.4	22	15.5	15.2	12.6
No Drift Correction	4.1	13.5	13.6	13.2	11.1	11.1
Path to Target	16	3.8	40.4	9.5	20.3	18

Distance traveled for overall navigation (in meters)

	$R_1$	$R_2$	$R_3$	$R_4$	$R_5$	Avg.
Our method	36.6	55.3	68.7	26.4	41.8	45.7
No Loop Closure	47.2	58.5	76.4	21.5	72.8	55.3
No Drift Correction	36	23.2	-	-	-	-
Path to Target	16.5	60.2	64.7	17.4	44.5	40.6

TEST BLOCK

Experiment paths ($10$ units $=$ $1$ m)

Figure 1: Environment 1 FloorMap

Index	Starting Point (X)	Starting Point (Y)	End Point (X)	End Point (Y)	Candidate Creation	Candidate Confirmation
$L_1$	95	390	10	270	9.13	9.40
$L_2$	250	77	10	270	39.58	39.69
$L_3$	250	250	10	270	17.73	24.27
$L_4$	50	70	10	270	4.98	6.25
$L_5$	26	390	10	270	0.21	0.78
$L_6$	115	320	50	70	9.72	12.97
$L_7$	210	450	50	70	4.11	6.71
$L_8$	195	70	50	70	6.68	7.16
$L_9$	277	70	50	70	1.25	1.69
$L_{10}$	127	392	50	70	18.12	26.08
$L_{11}$	275	70	115	330	8.83	10.63
$L_{12}$	10	385	115	330	16.42	22.53
$L_{13}$	248	248	115	330	21.59	24.22
$L_{14}$	70	70	115	330	17.44	20.26
$L_{15}$	210	450	115	330	8.28	11.43
$L_{16}$	50	390	250	200	9.43	11.36
$L_{17}$	230	70	250	200	20.82	22.84
$L_{18}$	250	250	250	200	18.85	22.43
$L_{19}$	118	70	250	200	15.43	17.91
$L_{20}$	10	240	250	200	24.90	32.53

Figure 2: Environment 2 FloorMap

Index	Starting Point (X)	Starting Point (Y)	End Point (X)	End Point (Y)	Candidate Creation	Candidate Confirmation
$L_{21}$	198	135	155	470	7.24	7.67
$L_{22}$	225	465	155	470	6.01	6.79
$L_{23}$	8	131	155	470	1.17	1.39
$L_{24}$	220	133	175	75	0.99	1.27
$L_{25}$	266	382	175	75	12.85	14.03
$L_{26}$	10	255	175	75	2.59	3.45
$L_{27}$	219	490	175	75	0.01	0.01
$L_{28}$	170	135	219	490	5.82	6.26
$L_{29}$	268	219	219	490	11.95	12.71
$L_{30}$	8	415	219	490	11.80	12.47