We notice in Fig. 5 that the detection ranges with respect to daylight increases in the
case of fluorescent light and decreases in the case of incandescent light.
4.2 Discussion
First, we tested the mobile application (GuiderMoi) under different conditions mainly
distance, angle of view and illumination. We noticed that the colour of the targets was
well chosen and their detection is robust especially for the red and the black targets.
Then, in the case of illumination, the detection is better in the case of the fluorescent
light since it is white and has no influence on the colours.
Moreover, for the distance between the target and the user, the detection is between
15 and 30 cm and the optimal distances are 25 cm. In addition, for the viewing angle,
the best position is 90°.
Furthermore, for time processing, the mobile application (GuiderMoi) has the
advantage of being in real time. Unfortunately, we have a loss of detection at distance
greater than 1 m. In fact, the target becomes smaller and more difficult to identify. At a
short distance, we get the reflection of the smartphone on the target that gives false
colour detection. Besides, the user has to be exactly in front of the target in order to
have a robust detection. We have a total loss of detection at the limits of the target that
is to say at an angle of 0° and 180°. The navigation tests were successful in the case of
fluorescent light, the user in front of the target with a distance equal to 25 cm.
There are some limitations to our approach. The first is that we are only using the
Hue component of the HSV color space. This means that unless the object we are
trying to track is not a single shade, then the results will likely be suboptimal.
To remedy this, we can simply extend the code to compute a 2D histogram using
both the Hue and Saturation components. However, OpenCV currently does not support 3D histograms in the back projection calculation and Cam-Shift tracking.
The second limitation is tuning the number of bins in the color histogram. This will
depend on many aspects, including the application conditions. We need to tune this
parameter for our application.
Finally, if we are looking for a more robust tracking solution that can take into
account texture and localized features, we should look into keypoints detection, local
invariant descriptors (ex. DoG and SIFT), and matching between the sets of keypoints
and their corresponding features.
5 Conclusion
In this paper, we propose a navigation system based on mobile application “GuiderMoi” to provide assistance for visually impaired in indoor environment. In fact, the
mobile application provides information about directions and helps the visually
impaired to navigate effectively and independently based on colour targets detection
and identification. In future research, it would be interesting to study how we can
improve mobile applications using augmented reality and intelligent navigation based
on deep learning.
Mobile Assistive Application for Blind People in Indoor Navigation
401
case of fluorescent light and decreases in the case of incandescent light.
4.2 Discussion
First, we tested the mobile application (GuiderMoi) under different conditions mainly
distance, angle of view and illumination. We noticed that the colour of the targets was
well chosen and their detection is robust especially for the red and the black targets.
Then, in the case of illumination, the detection is better in the case of the fluorescent
light since it is white and has no influence on the colours.
Moreover, for the distance between the target and the user, the detection is between
15 and 30 cm and the optimal distances are 25 cm. In addition, for the viewing angle,
the best position is 90°.
Furthermore, for time processing, the mobile application (GuiderMoi) has the
advantage of being in real time. Unfortunately, we have a loss of detection at distance
greater than 1 m. In fact, the target becomes smaller and more difficult to identify. At a
short distance, we get the reflection of the smartphone on the target that gives false
colour detection. Besides, the user has to be exactly in front of the target in order to
have a robust detection. We have a total loss of detection at the limits of the target that
is to say at an angle of 0° and 180°. The navigation tests were successful in the case of
fluorescent light, the user in front of the target with a distance equal to 25 cm.
There are some limitations to our approach. The first is that we are only using the
Hue component of the HSV color space. This means that unless the object we are
trying to track is not a single shade, then the results will likely be suboptimal.
To remedy this, we can simply extend the code to compute a 2D histogram using
both the Hue and Saturation components. However, OpenCV currently does not support 3D histograms in the back projection calculation and Cam-Shift tracking.
The second limitation is tuning the number of bins in the color histogram. This will
depend on many aspects, including the application conditions. We need to tune this
parameter for our application.
Finally, if we are looking for a more robust tracking solution that can take into
account texture and localized features, we should look into keypoints detection, local
invariant descriptors (ex. DoG and SIFT), and matching between the sets of keypoints
and their corresponding features.
5 Conclusion
In this paper, we propose a navigation system based on mobile application “GuiderMoi” to provide assistance for visually impaired in indoor environment. In fact, the
mobile application provides information about directions and helps the visually
impaired to navigate effectively and independently based on colour targets detection
and identification. In future research, it would be interesting to study how we can
improve mobile applications using augmented reality and intelligent navigation based
on deep learning.
Mobile Assistive Application for Blind People in Indoor Navigation
401