Go to previous article
Go to next article
Return to 2002 Table of Contents
Yoshinobu Maeda*, Eiichi Tano**, Hideo Makino*, Takashi
Konishi**, Ikuo Ishii**
*Faculty of Engineering, Niigata University
**Graduate School of Science and Technology, Niigata University
8050, Ikarashi-2, Niigata, 950-2181, Japan
In order to assist the navigation of visually impaired pedestrians, using devices such as GPS (Global Positioning System), there should be a capability of automatically providing building names (landmarks) and compass information, in spoken format. In this report we experimentally manipulated availability of both building names and path intersection (waypoint) information. In our evaluation, sighted participants were blindfolded and walked along a given route using our guidance system. Using a multivariate analysis of the data, we demonstrated a reduction in walking time and distance with the addition of waypoint information. Our results indicate that the landmarks alone were sufficient to travel the route, but travel was more efficient using both types of information (landmarks and waypoints).
We have developed a GPS-based speech-output guidance system for blind and visually impaired pedestrians . Our system was originally composed of two separate units: a mobile unit with GPS, carried by the visually impaired pedestrian, and the base station, a centrally-located unit containing a Geographic Information System (GIS), as shown in Fig. 1(a). Recently, we have implemented a new version of the system which integrates the GPS and GIS components into one unit (Fig. 1(b)). Both the combined and two-component versions of the guidance system are intended for blind and visually impaired people.
The numerical location data, comprising the latitude and longitude output from the GPS receiver, are conveyed to the GIS to obtain information about nearby buildings (landmarks). Also, the user's facing direction, provided by a fluxgate compass worn by the user, is conveyed to the GIS. Synthetic speech software provides speech information to orient the traveler with respect to the nearby buildings.
When a blind user with the guidance system wishes to travel to some building, he or she may easily accomplish the task. However, finding the entrance to the building is difficult, for attributes of the building, like an entrance, are not included in the GIS. Also, when he or she has to take another way owing to temporary road repairs, he or she needs path intersection information in order to make a detour. Minimal information about path intersections (waypoints) can greatly improve a guidance system.
The main purpose of this report was to investigate the importance of waypoint information for the visually impaired . In our evaluation, we measured walking times and distances of 10 blindfolded, sighted participants in two experimental conditions: with landmark information only and with both landmark and waypoint information. These data were analyzed by means of Principal Components Analysis (PCA) and Discriminant Analysis (DA) using Mahalanobis distance.
Fig. 1: (a) Two-component version of the system. (b) Integrated version of the system. "*personal computer" means a small mobile computer.
The sighted participants in the experiment were 4 undergraduate students and 6 graduate students, aged 22-24 years, of Niigata University. They were blindfolded and behaved as if temporarily blind (Fig. 2(a)). The route of the experiment was located on the Niigata University campus. The participants knew several landmarks on the campus but had never walked the route using a blindfold. In order to insure each participantfs safety, two observers accompanied the participants on the walk. These observers did not converse with the participants at any time during the experiment.
All participants walked on the same route (a bitumen walking path) three times. In Condition I, half of the participants walked along the route using a blindfold, a long cane and the integrated version of the guidance system (see hardware in Fig. 2 (b)) with access only to landmark information. Three months later, they walked the route again with a blindfold, a long cane and the guidance system, this time with both landmark and waypoint information (Condition II). The other half of the participants performed in the same two conditions but in reverse order. The observers measured the walking times of the participants to the nearest minute. Walking distances in meters were obtained from the GPS data. Finally, all participants performed in a third condition (Condition III); here they walked the same route using vision (i.e., without the blindfold, long cane, and guidance system). Condition III was included in order to obtain normal walking times as reference values. Digital maps of our GIS software are shown in Figs. 3(a) and 3(b). In all three conditions, participants walked counter-clockwise from the point labeled "Start & goal" in Fig. 3(b), returning to the same point. This route had 9 path intersections along the way, and the total length of the route was 978 meters. The guidance system was programmed so that the participant received synthesized speech every 10 seconds, even though the GPS receiver updated the location data more frequently. The landmark and the waypoint data were contained within different layers of the GIS. We gave waypoint information higher priority, so that the guidance system provided waypoint information whenever the smaller waypoint block overlapped the larger landmark block (see Fig. 3(b)).
The walking times and walking distances were normalized by dividing by the participant's reference walking time (from Condition III) and the shortest route (=978m), respectively. We supposed that the data of Condition I and II were given by two-dimensional joint Gaussian distributions, which consisted of the normalized walking time (NWT) and the normalized walking distance (NWD), respectively. Using the NWT and the NWD, we obtained mean vectors and variance-covariance matrices to evaluate theprincipal components of the distribution and the Mahalanobis distances. We applied the PCA to two data sets of Conditions I and II to obtain the correlation between the NWT and the NWD. We also applied the DA to the data to get a discriminant function (or a boundary curve in the NWT-NWD plane) between two groups of data of Conditions I and II. Every point on the boundary curve takes the same Mahalanobis distance from two mean points (or two mean vectors) of the data of Conditions I and II.
Fig. 2: (a) Configuration. (b) Guidance system
Fig. 3: Digital maps in our GIS. (a) Landmark information (Condition I). (b) Landmark and waypoint information (Condition II). The small blocks labeled with "Start & goal" and with single digits are the waypoint information.
All participants completed the three conditions. Figure 4 shows the GPS plots of one of the participants. This participant temporarily got lost in the vicinity of both waypoints "3" and "5", which are depicted in Fig. 3(b). Figure 5 gives the mean and the standard deviation of NWT (Fig. 5(a)) and NWD (Fig. 5(b)). In each graph, the left represents Condition I, and the right Condition II.
In the PCA, the data of Conditions I and II were approximately scattered along a positively-slope line in the NWT-NWD plane, respectively (not shown in figure). These lines represent the eigenvectors corresponding to the maximum eigenvalues of the variance-covariance matrices. The fact that the data lie on a positively-slope line indicates that NWT has a positive correlation with NWD.
In the DA, the discriminant function (thick solid curve in Fig. 6(a)) in the NWT-NWD plane is shown as the ellipse, where unfilled and filled circles represent the data of Conditions I and II, respectively (Fig. 6(a)). Figure 6(b) is a magnification in the vicinity of the ellipse of Fig. 6(a).
Fig. 4: GPS trajectories. The labels "Start&goal", "3" and "5" correspond to those in Fig. 3(b). (a) Condition I. (b) Condition II.
Fig. 5: Mean and standard deviation. (a) Condition I. (b) Condition II.
Fig. 6: DA. (a) Unfilled (white) and filled (black) circles represent the data of Conditions I and II, respectively. Unfilled (white) and filled (red) squares represent the mean points of Conditions I and II, respectively. The ellipse represents the discriminant function. (b) A closeup of the data in the vicinity of the ellipse of (a). Abscissa, the normalized walking time (NWT); ordinate, the normalized walking distance (NWD).
It was demonstrated in Fig. 4 that the blindfolded pedestrian lacking waypoint information in the GIS got lost in the vicinity of two intersections. We can also see it statistically in the NWT and the NWD of Fig. 5. These mean values were reduced in half when waypoint information was available.
The results of the PCA confirms that Figs. 5(a) and 5(b) are the correct results. If the participant had stopped walking intentionally in the experiment, the NWT would be overestimated relative to NWD. Thus, the positive correlation between the NWT and the NWD would differ from the results of the PCA.
The discriminant function of the DA (thick solid curve in Fig. 6) was elliptical, where the inside of the ellipse corresponded to the statistical group of the data of Condition II, and the outside Condition I. Indeed, the mean point of Condition II (filled square) was inside the ellipse, and conversely, the mean point of Condition I (unfilled square) was outside. Generally, if the two groups were to have the same statistical variances, the discriminant function would be linear. When the difference of the variances is large, the discriminant function is quadratic nd changes from hyperbolic to elliptic. Therefore our result indicates that, when the participants had waypoint information in the GIS, not only the mean values of the NWT and the NWD were reduced by so were the variances. This means that waypoint information is necessary for efficient travel, though navigation differs considerably from one individual to the next.
It is expected, from the diameter of the ellipse in Fig. 6(b), that visually impaired people will take no more than 3.3 times (and no less than 1.8 times) as much time to traverse a path than sighted people, if they have both landmark and waypoint information. For distance, these corresponding maximum value is 1.3.
In our guidance system for the visually impaired pedestrian, we divide guidance information into two categories: building names (landmarks), and path intersections (waypoints). These two types of information were provided in different layers in the GIS of our guidance system. We programmed the guidance system to give priority to the waypoint information. In order to assess the importance of waypoint information, we experimented using 10 blindfolded sighted participants and analyzed the walking times and distances. The results indicate that providing both landmark and waypoint information improves effectiveness of the system.
We are going to investigate the effect of the other information such as paths and districts on the guidance system. Furthermore, we are going to improve the guidance system so that it has an input module by means of the speech recognition .
The authors would like to thank Prof. Jack M. Loomis for helpful discussions.
 H. Makino, I. Ishii and M. Nakashizuka (1996) Proc. 18th Conf. IEEE/EMBS, Amsterdam, The Netherlands, Oct.31-Nov.3
 E. Tano, Y. Maeda, H. Makino, T. Konishi and I. Ishii (2001, in press) Theory and Applications of GIS, GISA, Vol.9, No.2 (in Japanese)
 R. G. Golledge, R. L. Klatzky, J. M. Loomis, J. Speigle and J. Tietz (1998) Int. J. Geographical Information Science, Vol.12, No.7, pp.727-749
 J. M. Loomis, R. G. Golledge and R. L. Klatzky (2001) pp.429-446, In: W. Barfield and T. Caudell (eds.) Fundamentals of Wearable Computers and Augmented Reality, Mahwah NJ: Lawrence Erlbaum Associates.
Go to previous article
Go to next article
Return to 2002 Table of Contents
Return to Table of Proceedings