Development of an Embedded 3D Robot Visual System

Development of an Embedded 3D Object Visual System

Analysis, Conclusion & Sources of Error

Analysis

In graphs of computed vs. actual distance, there is a line showing the ideal values with a slope of 1. Linear least squares regression was done on these
data to generate a straight line. With these data, the 95% confidence interval of the slope is (.71 to .98). These bounds do not include 1, so the
calculated line is slightly significantly different from the line of perfect prediction.

The graph of area values at the different distances was done similarly. Linear least squares regression is done on these data to obtain a straight line. The
95% confidence interval of the slope is (-1.8 to .17), which includes 0, showing that area does not change significantly with distance.

Similar plots were repeated for angle of rotation and tilt. Regression of the computed rotation data on the actual showed that the slope was not
significantly different from the ideal. The regression of the area values shows that area does not change significantly with the angle of rotation.

For the tilt data, the regression line indicated that the slope is significantly different from the ideal value. The graph shows that as the angle of tilt
increases, the accuracy decreases. The regression of area showed that the area does not change significantly with tilt.

Little variation was found when the vectorization output was compared with the output of picking points by hand.

Conclusion

In graphs of computed vs. actual distance, there is a line showing the ideal values with a slope of 1. Linear least squares regression was done on these
data to generate a straight line. With these data, In conclusion to the experiment described in Objective – part B, my hypothesis was only
partially correct. From these results, it does not seem that accuracy is improved when the object is closer to the camera. It also shows that the
distance is calculated quite well. Distance vs. area data shows that the area calculations are rather poor.

The graph of area values at the different distances was done similarly. Linear least squares regression. The results for angle of rotation showed similar
conclusions to the distance data. Accuracy of the calculated angle of rotation was very close to the ideal values, as it was for distance. The area
calculated after each of the rotations seems to once again be quite poor.

When angle of tilt is examined, it can be seen that accuracy starts out very good, and gradually becomes worse. The area data for angle of tilt are
poor.

Overall, I am quite pleased with the accuracy of the distance, rotation, and tilt calculations. However, the area calculations were overall poor.

Sources of Error

One main source of error is the alignment of the cameras. One of the cameras had the CCD shifted down vertically by the equivalent of 28 pixels. This was corrected using math, but showed that the accuracy of the cameras may be quite poor. A new camera will be obtained for future experimentation. Also, the orientation of the cameras may not be completely perfect, as well as
the actual position and orientation of the object. There may have been some small inaccuracies caused by poor measurement.