• No se han encontrado resultados

VI. RESULTADOS Y DISCUSION

6.2. Efectos de los diferentes tipos de tratamientos orgánicos y convencional

FLANN

The RGB-D SLAM System using an Adjustable BRISK-AGAST detector, a FREAK extractor and a FLANN matcher has six modifiable parameters for feature detection. In its testing, fMin-fMax ranges of 200-201 to 600- 601 were used in increments of 100, in addition to patternScale values between 16.0 and 24.0 in increments of 2.0, nOctaves values of zero and one, and normalization settings with both scale and orientation normalization disabled, only scale normalization enabled, and both terms enabled. The results of successful runs using these settings are recorded in Table 5.19 and Table 5.20. The mean ATE and RME values for each run are plotted in Figure 5.17.

2 4 6 8 10 12 14 16 18 20 22 0 0.2 0.4 0.6 0.8 1 time (s) Translation Error Translation Error (m) 2 4 6 8 10 12 14 16 18 20 22 0 0.2 0.4 0.6 time (s) Angle (rad) Angular Error

Figure 5.16: ATE translation and orientation errors for the Adjustable BRISK-AGAST, BRISK, FLANN trial run using feature bounds of 300-301, 0 nOctaves, and a patternScale of 1.5.

Table 5.19: ATE and RME values using an Adjustable BRISK-AGAST detector, a FREAK extractor, and a FLANN matcher.

Feat. Oct. Scale SN/ON Avg. RME Trans. RME Ang. ATE Trans. ATE Ang.

Bnd. Feat. (m) (rad) (m) (rad)

500-501 0 16.0 FF 515.1 0.024 ± 0.048 0.016 ± 0.018 0.196 ± 0.147 0.164 ± 0.063 600-601 0 16.0 FF 618.1 0.021 ± 0.041 0.015 ± 0.016 0.094 ± 0.069 0.161 ± 0.051 400-401 0 18.0 FF 417.3 0.030 ± 0.066 0.020 ± 0.030 0.157 ± 0.121 0.296 ± 0.094 500-501 0 18.0 FF 515.1 0.023 ± 0.046 0.015 ± 0.018 0.176 ± 0.135 0.252 ± 0.047 600-601 0 18.0 FF 618.1 0.019 ± 0.033 0.014 ± 0.012 0.122 ± 0.073 0.194 ± 0.051 500-501 0 20.0 FF 513.5 0.048 ± 0.174 0.030 ± 0.100 0.153 ± 0.142 0.164 ± 0.100 400-401 0 22.0 FF 414.2 0.040 ± 0.059 0.027 ± 0.036 0.124 ± 0.100 0.266 ± 0.028 500-501 0 16.0 TF 515.1 0.022 ± 0.039 0.015 ± 0.017 0.139 ± 0.111 0.226 ± 0.037 600-601 0 16.0 TF 618.1 0.021 ± 0.043 0.014 ± 0.015 0.100 ± 0.088 0.183 ± 0.046 500-501 0 18.0 TF 515.1 0.022 ± 0.050 0.015 ± 0.018 0.163 ± 0.122 0.156 ± 0.049 600-601 0 18.0 TF 618.1 0.025 ± 0.061 0.016 ± 0.020 0.140 ± 0.110 0.257 ± 0.046 400-401 0 22.0 TF 414.2 0.063 ± 0.197 0.049 ± 0.208 0.141 ± 0.154 0.174 ± 0.150

Table 5.20: Timing results for an Adjustable BRISK-AGAST detector, a FREAK extractor, and a FLANN matcher.

Feat. Octaves Scale SN/ON Detect Time Extract Time Match Time Run time

Bnd. (s) (s) (s) (s) 500-501 0 16.0 FF 0.007 0.002 0.024 99.938 600-601 0 16.0 FF 0.008 0.002 0.032 111.254 400-401 0 18.0 FF 0.005 0.002 0.016 85.825 500-501 0 18.0 FF 0.006 0.002 0.023 102.019 600-601 0 18.0 FF 0.007 0.002 0.030 111.960 500-501 0 20.0 FF 0.007 0.002 0.023 93.776 400-401 0 22.0 FF 0.005 0.002 0.017 84.048 500-501 0 16.0 TF 0.007 0.002 0.024 101.698 600-601 0 16.0 TF 0.008 0.002 0.032 112.666 500-501 0 18.0 TF 0.006 0.002 0.023 101.719 600-601 0 18.0 TF 0.008 0.002 0.030 114.660 400-401 0 22.0 TF 0.006 0.002 0.017 85.999

80 85 90 95 100 105 110 115 0

0.1 0.2 0.3

Detector: abrisk Extractor: freak Matcher: flann

ATE Translational Error (m)

Runtime (s) 80 85 90 95 100 105 110 115 0 0.1 0.2 0.3

ATE Angular Error (rad)

Runtime (s) (a) 80 85 90 95 100 105 110 115 0 0.02 0.04 0.06

Detector: abrisk Extractor: freak Matcher: flann

RME Translational Error (m)

Runtime (s) 80 85 90 95 100 105 110 115 0 0.02 0.04 0.06

RME Angular Error (rad)

Runtime (s)

(b)

Figure 5.17: ATE and RME values for the entries in Table 5.19. Obtained using an Adjustable BRISK-AGAST detector, a FREAK extractor, and a FLANN matcher. Figure 5.17a gives the mean absolute trajectory error compared to the total run time for each parameter setting. Figure 5.17b gives the mean relative motion error compared to the total run time.

to run to completion. Runs using bounds of 200-201 and 300-301 experienced difficulty at the 8 s mark. With these lower feature bounds, enough matching features could not be found, resulting in a failure to track the camera’s pose through the horizontal rotation. The 500-501 feature bound was the most consistent, able to function with almost all the patternScale settings. In these trials, increasing the feature bounds by increments of 100 netted an average decrease to the ATE metrics of 0.047 m and 0.006 rad, but at the cost of a 1.2 ms per frame increase to detection times, a 7.4 ms per frame increase to matching times, and a 12.6 s increase to total run time. It is best to use as small feature bounds as possible, as the improvement to pose accuracy does not warrant a greater than 10% increase to run time.

No trial runs were successful using an nOctaves setting of one. In every case, the detector was unable to supply enough matching features to recover the translation between poses, and tracking failed at the start of the run.

All patternScale values functioned adequately with the exception of 24.0. Runs using this larger value had difficulty during the first rotation, with the number of matching features dropping below acceptable levels, and localiza- tion failing. Sampling pattern scales of 16.0 and 18.0 were the most consis- tent, able to operate with most all feature bounds. In general, increasing the scale resulted in a small increase in ATE on the order of 0.03 m and 0.007 rad, and an increase to run time of 1.8 s; however with low feature bounds the greater patternScale values performed better, yielding a higher

number of correct matches, and a faster total run time due to a more rapid pose graph optimization. As such, it is best to use a scale of 16.0 with most feature bounds, but a scale of 22.0 for settings where the average number of features per frame is near 400.

Similar to the brute-force matching case, enabling orientation normaliza- tion caused all trials to fail. These runs all failed at the start of the first major turn. In each case, no matter the feature bounds, not enough match- ing features could be found to track the camera’s pose through the horizontal rotation and localization failed. In order to achieve any results at all, orien- tation normalization should be disabled.

Enabling scale normalization netted a slight overall improvement in the mean ATE with a 0.013 m decrease in the translation term, but a 0.003 rad increase in the orientation term. This improvement came at the cost of an average of a 1.38 s increase to total run times.

The best results were found using feature bounds of 400-401, a pattern- Scale value of 22.0, an nOctaves value of zero, scale normalization enabled, and orientation normalization disabled. Using these settings, mean absolute trajectory errors of 0.141 ± 0.154 m and 0.174 ± 0.150 rad were found, with a total run time of 88.99 s. Plots of the ATE terms for the entire 21 s run are shown in Figure 5.18. At the conclusion of the first horizontal rotation of the camera at the 9.5 s mark, the error spiked to 1 m and 2.5 rad when the system estimated the view of the desk was from a pose opposite of the true pose. Additional peaks in the error occurred at the 10 s and 10.75 s marks, indicating some trouble tracking the pitch of the camera.

Documento similar