The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

Sonar Map Construction for a Mobile Robot Using a Tethered-Robot Guiding System Yu-Cheol Lee, Sang-Ik Nah, Hyo-Sung Ahn, Wonpil Yu Electronics and Telecommunication Research Institute (ETRI) 161 Gajeong-dong, Yuseong-gu, Daejeon, Korea {yclee, nsi, hyosung, ywp}@etri.re.kr Abstract - This study introduces a sonar grid based map construction method using Navi-guider that is one of the tethered connection systems on a mobile robot. Basic function of Navi-guider is to actively control the robot through the tension length and an orientation of the cable. Navi-guider, which is able to easily control the mobile robot, is applied to robot controller for building the map. And StarLITE, being comprised of infrared LED and CCD camera, used for providing the global robot position by an artificial landmark. In case of using Navi-guider, user can easily control the mobile robot, but the sonar sensors of the robot head occasionally return the false distance between the sensors and object. Therefore, this paper presents a filtering method, called Cell Conjunction Filter(CCF), for building a grid-based map for a mobile robot using Navi-guider. The CCF was applied to the Bayesian and orientation probability grid-map construction model, which are typical models used to build grid maps, to verify its effectiveness. Experimental results were also obtained using a mobile robot in a real world environment. Keywords - Navi-guider, Cell Conjunction Filter, Mapping, Mobile Robot

1. Introduction The vital functions of mobile robot navigation are localization, path planning, obstacle avoidance and representation of the environment. Of theses takes, exploration involving mapping and localization in an unknown environment is the most important task in mobile robots[1]. For this, we use the sonar sensor because it gives direct distance information on the location of an object. And, sonar is attractive tool of mobile robot for their economic aspects; however the range data of sonar sensor have much uncertainty due to the wide beam width of ultrasonic waves. For instance, a sonar sensor suffers from a multi-path phenomenon due to the specular reflection effect. To overcome the uncertainties of sonar sensor, this paper describes a grid-based map algorithm. To build a grid-based map, the concept of Bayesian formula introduced by Moravec was used[2]. Bayesian probability map is composed of many grids that present the robot’s workspace, and each grid has an occupancy probability of an object. The occupancy probability of each grid is determined by using Bayes conditional probability theory

according to the distance or angle of grid from sensor. The Bayesian model, however, still updates those probabilities regardless of the specular reflection effect. As a result, the occupancy probability is exceedingly decreased on the grids corresponding to real objects. To solve this problem, we developed a mapping model with ability to detect an occurrence of the specular reflection effect by evaluating the orientation probability in each grid[3]. In this model, the orientation probability is updated by conversely using the specular reflection effect. A grid is divided into some orientations. The orientation is determined by the relative angle from the sensor, and its probability is updated by Bayes formula. The vertical orientation probability of sonar sensor is increased because the sonar beam must reflect vertically from an object surface. The orientation probability can only adjust the variation of the occupancy probability according to an occurrence possibility of specular reflection effect. The orientation probability is, therefore, not a fundamental solution for the specular reflection effect because it still updates the probabilities from wrong data that should be rejected. We have developed a new map building method, called Data Association Filter (DAF), that can remove the incorrect range information such as specularly reflected data[4]. Therefore, the DAF can improve the reliability of the data for the grid-based map. This method is based on the evaluation of possibility that the acquired data are all from the same object. Namely, the data from specular reflection have very few possibilities of detecting the same object, so that they are excluded from the data cluster during the process of the DAF. Therefore, the uncertain data corrupted by the specular reflection and/or multi-path effect, are not used to update the probability map, and hence building a good quality of a grid map is possible even in a specular environment. By using the DAF, we found that the quality of the resulting grid map is considerably increased. The above mentioned grid-based mapping techniques are centered on building a good quality map using sonar sensor. This paper, therefore, develops the quick mapping methods using Navi-guider, which is one of the tethered-robot cabling systems. Navi-guider has a advantage that it is a easily control the robot[5], but the robot moving direction sonar sensors have a false measured range because Navi-guider user is most likely to include its beam aperture. These sonar data could make a

429

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

Tv

ts

Error Boundary

TOF

te

t

Fig. 1. Signal processing principle of sonar low quality grid map by falsely updating occupancy probability of a grid map. We, therefore, suggest a filtering method which called a Cell Conjunction Filter(CCF). The CCF can extract a good quality grid map how near grids mutually compensate about the falsely updated occupancy probability.

2. Cell Conjunction Filter The Orientation Model and DAF can reduce the sonar sensor uncertainty[4]. In here, Navi-guider equip for quick and accurate mapping at the front of the robot. Navi-guider user, however, can make against the sonar sensor because Navi-guider user obstruct sonar sensor’s field of beam aperture. The sonar sensors in front of mobile robot measure the wrong locations of the object. The locations of the object falsely draw on the grid map when these data used for building a grid map. This research, therefore, developed the CCF for compensating the wrong updated occupancy probability of the grid map. 2.1 Basic Principle of Cell Conjunction A sonar sensor can measure the distance from the nearest object to the sensor. The transmitter emits an ultrasonic wave in front of the sensor and the receiver detects the beam reflected back by the first contacted object. The sonar sensor then calculates the time of flight (TOF), or the elapsed time between transmission and detection of the sonar beam, and converts the result from time to distance using the speed of sound. At this time, distance unit of sonar sensor was decided according to digital signal resolution. In Fig. 1, the resolution of a sonar sensor is rather bigger than one grid size. In that case, it is difficult to know where a reflected grid is in the map because one grid size include into error boundary of sonar information. On this occasion, sonar’s reflected beam was definitely returned from a grid into error boundary of sensor resolution. An occupancy probability of one grid, therefore, has to be evaluated in conjunction with surrounding grids. This is a basic principle of the CCF based on physical sonar sensor characteristics.

information about the object location. Metal, wood, concrete, and glass can be detected well because they reflect almost all ultrasonic waves, but cloth is hardly detected at all because the material absorbs ultrasonic waves. The specular reflection effect is another problem that occurs frequently in the real world, in which the reflected beam is redirected away from the receiver by an object with a surface that is not perpendicular to the sensor orientation. Specular reflection of the sonar beam gives rise to two types of problems, often referred to as the multipath phenomena of sonar sensors [6]. The entire distance traveled by the beam after multiple reflections is longer than the maximum detection range of the sensor. This frequently occurs at concave corners, which are difficult to detect using sonar sensors. Another case of the multipath effect is that the total distance of the beam path due to multiple reflections is shorter than the maximum detection range. In this case, a phantom object is detected instead of the nearest real object. If data resulting from multipath phenomena are used to update the sonar map, its quality will markedly deteriorate. To solve this problem, the Orientation Model was developed with the ability to detect specular reflection by evaluating the orientation probability of each grid. The possibility that the specular reflection effect or multipath phenomena occurred for each set of sonar range data can be probabilistically considered using the orientation probabilities. Data with a high possibility of the specular reflection have little influence on the occupancy probability updates while data with a low possibility of specular reflection effect have much influence on the occupancy probability updates. Thus, the orientation model can construct a good quality map despite the presence of specular reflection. The orientation model, however, is not a fundamental solution for the specular reflection effect problem because it still updates the grid map probabilities using false data that should be rejected. Therefore, we have developed a new map building method based on a DAF that can remove the incorrect range information caused by specularly reflected data. By combining grid-based mapping algorithms obtained using the DAF, we will show that the quality of the resulting probability map improves considerably. In this sonar map construction method, the update occupancy probability is determined using the Bayes conditional probability theory according to the distance of the grid from sensor. It is an independent relation between emptied and occupied possibilities of a grid based on Bayesian formula which defined in Eq. (1).

Pi ( o | A) + Pi ( o | A) = 1

(1)

where Pi ( o | A ) is an occupancy probability which is possibility of an existing object in a grid. And Pi ( o | A ) is

2.2 Update Probability Using Cell Conjunction Filter A sonar sensor sometimes cannot find the nearest object in front of it and instead gives incorrect distance

empty probability that is possibility of not existing object. The grid based probability map is updated using Bayes conditional formula whenever a new sensing operation of

430

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

i+n

i + n +1

i

i +1

P ′ ( oi )

row map size = n

Fig. 2. Revaluation of occupancy probability using grid association filter

Fig. 3. 3-DX mobile robot with a sonar ring and Navi-Guider

the mobile robot takes place. The object being and not probabilities can be calculated by Eq (2) and (3).

P(o M ∩ A) =

P( M o ∩ A) × P (o A)

(2)

P( M o ∩ A) × P (o A) + P( M o ∩ A) × P(o A)

P(o M ∩ A) =

P( M o ∩ A) × P (o A)

(3)

P ( M o ∩ A) × P (o A) + P ( M o ∩ A) × P(o A)

where P(o M ∩ A) is estimated occupancy probability of

a grid given accumulated map information A and a new sensing distance M . P( M o ∩ A) is the probability that a new sensing distance M can be obtained assuming that a grid is occupied and A is given and P ( o A) is the occupancy probability of a grid. Division of Eq (2) by Eq. (3) gives

Odd (o) =

P ( o M ∩ A)

P ( o M ∩ A)

=

P ( M o ∩ A)

×

P ( o A)

P ( M o ∩ A) P ( o A)

(4) Fig. 4. Configurations of objects for the experiment

In Eq. (4), because P ( o A) can be retrieved from old map information, only P ( M o ∩ A ) is left side unknown value for calculation P ( o M ∩ A ) . Occupancy probability of a grid can be computed by Eq. (5) which was given by using Eq.(1) and Eq.(4). P ( oi ) =

Odd ( oi )

1 + Odd ( oi )

Updated occupancy probability with Eq. (5) is reevaluated by a CCF. As shown in Fig. 2, left-bottom corner of a grid defines the origin of the object. A grid i ’s occupancy probability is again updated by related with near 3 grids’ occupancy probabilities using Eq. (6). where n is a row map grid size.

(5)

431

P ′ ( oi ) =

P ( oi ) + P ( oi +1 ) + P ( oi + n ) + P ( oi + n +1 ) 4

(6)

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

(a)

(a)

(b) Fig. 6. Constructed grid map after grid association filtering

(b) Fig. 5. Constructed grid map from the experimental data

432

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

If any grid’s occupancy probability is wrong updated by sonar uncertainty data, it was compensated with the CCF by using near grids’ occupied information.

Acknowledgements This work was supported in part by MIC & IITA through IT Leading R&D Support Project

3. Experiment and Results

References

3.1 Experimental Setup and Method The mobile robot used in this experiment was the Pioneer 3-DX, built by the ActiveMedia Robotics Company and is shown in Fig. 3. This robot consisted of two driving wheels and one non-driving wheel. The distance moved was measured with two encoders mounted on each driving wheel. A gyroscope was used to measure the rotational angle of the robot. A sonar ring composed of 16 sonar sensors at uniform intervals of 22.5 was mounted on the top of the robot. The sonar sensor was a Murata MA40B8R/S model with a beam aperture of approximately 45 , minimum detection range of 20cm, and maximum detection range of 6m. Figure 4 shows the experimental environment, which consisted of sofas, tables, chairs, desks, televisions, and drawers.





3.2 Experimental Results Figure 5 shows the results of the experiment. The size of the grid was 8 × 8 cm, and the sonar beam aperture was assumed to be 45°. The total number of data points was 24057, collected from 2187 different positions. In the maps shown in the Fig. 5, the gray level represents the occupancy probability: the darker the color, the higher the occupancy probability. Figure 5(a) draws a whole probability values. And Fig. 5(b) shows threshold maps of the Fig. 5(a) with 0.8 occupancy probability value. As shown in Fig. 5, the occupancy probability of some grid is falsely updated in the middle of grid map because Navi-guider user is affect to sonar sensor as an obstacle. Figure 6 illustrate the map obtained using the CCF. Figure 6(a) draws a whole probability values. And Fig. 6(b) shows threshold maps of the Fig. 6(a) with 0.8 occupancy probability value. Comparing to the results shown in Fig. 5(a), wrong updated occupancy probability in the grid map is compensated by the CCF. This implies that the incorrect sonar range data effect by Navi-guider user and sonar uncertainty were attenuated by the CCF.

[1] J. J. Leonard and H. F. Durrant-Whyte, “Mobile Robot Localization by Tracking Geometric Beacons,” IEEE Trans. on Robotics and Automation, Vol. 7, No. 3, pp. 513-523, 1991. [2] H. P. Moravec and A. Elfes, “High Resolution Maps from Wide Angle Sonar,” Proceedings of the IEEE International Conference on Robotics and Automation, St. Louis, pp. 116-121, 1985. [3] J. H. Lim and D.-W. Cho, “Multipath Bayesian Map Construction Model from Sonar Data,” ROBOTICA , Vol.14, pp. 527-540, 1996. [4] S.-J. Lee, Y.C Lee, D.-W. Cho, W.-K. Chung, J.-H. Lim, C.-U. Kang, W.-S. Yun, "Evaluation of Features through Grid Association for Building a Sonar Map", Proceedings of the 2006 IEEE International Conference on Robots and Automation, Hilton, Orlando, Florida, USA, pp. 2615-2620, May 15~19, 2006. [5] E. F. Fukushima, N. Kitamura and S. Hirose, "A New Flexible Component for Field Robotic System ", Proceedings of the 2000 IEEE International Conference on Robots and Automation, San Francisco, CA, USA, pp. 2583-2588, April, 2000. [6] J. H. Lim and D.-W. Cho, “Specular Reflection Probability in the Certainty Grid Representation,” Transactions of ASME Journal of Dynamic System, Measurement and Control, Vol. 116, pp. 512-520, 1994.

4. Conclusions We developed a cell conjunction filter that could compensate incorrect sonar range data by interrupt of Navi-guider user and sonar sensor uncertainty. The method was based associations among occupancy probabilities of the near grids. The CCF was applied to the grid based mapping method which called the Orientation model that typically used to build a grid map to verify its effectiveness. The experimental results showed that the quality of the resulting map was quite good even through incorrect sonar range data were occurred by Navi-guider user and sonar uncertainty. Consequently, the CCF is an effective means of compensating wrong updated probability and has potential to become widely used in the area of Navi-guider and mobile robot research.

433