INDOOR POSITIONING AND NAVIGATION BASED ON CONTROL SPHERECAL PANORAMIC IMAGES

4.2 Comparison of datasets with different flying height

Tsung-Che Huang ¹ , Yi-Hsing Tesng ²

1 Master Student, Department of Geomatics, National Cheng Kung University

2 Professor, Department of Geomatics, National Cheng Kung University

I. Abstract

When using image-based indoor positioning and navigation, artificial landmarks (Se et al., 2002) or natural visual landmarks (Hayet et al., 2006) and image databases are major strategies. In this study, the latter is chosen to achieve the indoor positioning and navigation. Suppose we have an image database, it would be possible for us to compare the captured image with the database of reference images. In this study, we used spherical panoramic image (SPI) (Lin, 2014) to get more image information and reduce the computation cost. However, because of complex geometry of SPI, there is a problem to be solved in this study. That is how to ensure the correctness of SPI matching. With the aid from epipolar geometry, lots of applications are proposed such as matching problem, pose problem and so on. Corresponding to the two raised problems, there are two objectives in this study. The first one is developing and implementing the theory of SPI matching. With the proposed method, we can increase the efficiency and reliability of matching result. The second one focuses on developing an algorithm to apply epipolar geometry for calculating the position and orientation of the SPIs.

II. Data Flow and Processes

Indoor Positioning and Navigation Based on Control Spherical Panoramic Images

IV. Transformation Model in RANSAC for Elimination Incorrect Matches

A

A

Number of

candidate matches

404

remaining matches with affine model

130(28%)

remaining matches with essential model

360(77%)

Movement test with RANSAC affine model Movement test with RANSAC essential model Comparison of two models

V. Intersection Ambiguity of SPI and Check Angle

If both rotation and translation of second camera are correct, the angle between ( 𝑅𝑅 _𝐼𝐼 ^𝑂𝑂 � 𝑟𝑟 _𝑃𝑃 ^𝐼𝐼 ) and 𝑟𝑟 _{𝐼𝐼→𝑃𝑃} ^𝑂𝑂 will be close to zero rather than close to 180 ˚. We can use this check angle to solve the ambiguity of intersection. The only one correct solution is the one has the both check angles of the two SPIs to be small. The check angle of each SPI can be computed by using inner product computation.

VI. Experiments and Results

The test field is in the indoor space of the Department of Geomatics. For validation, the POPs of the test images were determined in advance with the bundle network adjustment. Two kinds of corresponding points were applied in this experiment. The first kind involves manually measured points and the second kind involves automatic matched points, so that the effect of image correspondence can be tested.

Query SPI △X(m) △Y(m) △Z(m)

1 -0.002 0.027 -0.004

2 -0.007 0.001 0.002

3 0.045 0.007 0.014

4 0.302 0.052 0.126

5 0.040 0.038 0.002

Mean error 0.076 0.025 0.028

RMSE ±0.137 ±0.031 ±0.057

Query SPI △X(m) △Y(m) △Z(m)

1 0.011 -0.005 -0.016

2 0.448 -0.315 0.464

3 0.216 -0.030 0.050

4 -3.654 4.090 -0.097

5 0.138 -0.276 0.060

Mean error 0.203 -0.157 0.140

RMSE ±0.258 ±0.201 ±0.235

Query SPI Number of corresponding points SPI No. 1 297

154 149

28 191

175 33 72 22 79

91 32 46 20 60 SPI No. 2

SPI No. 3 SPI No. 4 SPI No. 5

Coordinate errors and mean error and RMSE

with manual measurements Coordinate errors and mean error and RMSE with image matching

Number of corresponding points between overlapped SPIs

III. Scenario of the Navigation Environment