Download as doc, pdf, or txt
Download as doc, pdf, or txt
You are on page 1of 8

Feature Extraction for Autonomous Ro bot Navigation in Urban Environme nts

Carlos Arturo Sainz Moreno & Edgar A. M tnez - Gar c ar a Center for Autonomous Robots, Sensing and Perception Institute of Engineering and Technology Universidad Autonoma de Ciudad Ju arez

A bstract
This paper states all the basis of a robotic architecture capable of recognizing and transit throughout man-made urban environme nts. The described architecture will use a range laser, sonar and vision sensors with a feature extraction method named Adaptive Scale Sample Consensus (ASSC). Here, the authors present the background work of such architecture.

Intr oduction

Autonomous mobile robots have been widely applied in applications and research. Some of the main problems in robotics are self-localization and mapping the environme nt, both of them require to process sensor raw data in order to have means for understanding the environment. There exist many methods to understand such information by extracting features from the sensed environme nt. Some algorithms are: Split and Merge, Hough Transform or Random Sampling Consensus (RANSAC). We present the basis of a robotic architecture 95

that should be able to identify different structures, map the environme nt and avoid colli- sions. To achieve such abilities a holistic control method, a robust feature extraction algorithm and a good planned architecture must be inte- grated. In the following sections we will present the background work, highlighted by the devel- opment of a robotic architecture from scrat ch: the hardware generals (mechanical platform and gears box), electronic (sensors and connection schematics), and software (control files), which has been developed in laboratory to carry out the proposed algorithmic meth od.

Ro botic Platform

The main structure is a differential controlled based architecture, whose dimensions are 38cm 38cm 38cm. It has a middle section where the interface circuits and some sensors on-board. At the bottom it has a gear box with a 5:1 scale, the H-bridge, a coupling circuit and the batter y. Finally, at the top level, the main board and the laser sensor (Hokuyo) are fixed. The robotic architecture is equipped with

96

1st Simposium on Robotics,

Sensing and Pe rception,

N ov/29th /2008

Carlos Arturo Sainz Moreno

Figure 1: Robotic Platform various sensors, which include a sonar, laser and vision sensors. The main board is connected to control the laser and vision sensors and the interface board. Some features of the main board are: 2 AMD Turion 64 - 1.8 GHz Gentoo Linux embedded Wireless technology device (USB) 3 USB connectors As previously mentioned, the main board is Gentoo Linux embedded because the capabilities of this operating system. From all the Linux distributions, Gentoo is one of the most customizable available. Its kernel can be arranged to fit both user and hardware requireme nts. The sensors used in the architecture: 2 VEX Robotics Optical Shaft Quadrature Encoder.

1st Simposium on sensor SRF05 from Pe rception, Ultrasonic Robotics, Sensing and DEVAN-

N ov/29th /2008

TECH Ltd. URG-04LX Hokuyo range laser sensor. Logitech QuickCam. The encoders have two channels each with a 90 pulse per revolution relation. By means of software, the direction, velocity and position of each wheel can be computed, and, in consequence, the actual position of the robot can be estimated. The interface board is a Phidgets Interface Kit 8/8/8, which has 8 digital input channels, 8 analog input channels, and 8 digital output channels. This circuit board is used to control the H-bridge and the odometric sensors input. Some features of this circuit board are: 1. USB interface. It permits a direct connection to the computer. 2. It can be used under Windows, Linux, or Mac OS X operative systems. 3. It has its own libraries & Input / Output change handler modules. 4. Electronic circuits for noise filter. As shown in figure 2, between the interface board and the sensors is connected a coupling circuit. This coupling circuit board has three main purposes: 1. Regulating the battery voltage level from 12V/9V, 4A. 2. Encoders signal processing. 3. Give some neatness to the wiring

Carlos Arturo Sainz Moreno

3.1

Hoku yo Scanning Finder

Laser

Range

Figure 2: Robotic architecture functionalities and control diagram. Figure 3: URG-04LX Hokuyo range laser sensor The architecture functionalities and control This sensor offers both serial (RS-232) and diagram can be seen in figure 2. In the following sections, we will present the USB interfaces to provide extremely accurate chosen feature extraction method and the sensor laser scans. The field-of-view for this detector is 240 degrees and the angular resolution is .36 used for such work. degrees with a scanning refresh rate of up to 10H z. Distances are reported from 20mm to 3 Feature Extraction 4 meters. Power is a very reasonable 500mA at 5V . The feature extraction process is used in order to By means of Player, a hardware abstraction layer debug the great amou nt of sensor in-data into a (HAL) for robotic devices, the Hokuyo URG smaller amou nt of information with much more laser sensor is controlled by the main board. The relevant information than the original. Feature following lines state the basic configuration code extraction in robotics takes advantage of human of this sensor: capabili ty of abstraction: data compression, urglaser.cfg denoising, and distinguishabili ty [2]. driver ( name "urglaser" provides ["laser:0"] port "/dev/ttyACM0" )

3.3

Line Extraction

Algorithms

Instead of working directly with raw scan points, feature-based matching first transforms the raw scans into geometric features. Being more compact, they require much less storage and still provide rich and accurate information.

Figure 4: Hokuyo Laser Sensor Detection Area

3.2

Sensing Model

The presented range laser sensor returns a set of polar points which have the sensor as origin. In Figure 6: Extracted Line figure 4 the scanning area is prese nted. The scan area is 240 semicircle with maximum radius 4000mm. Pitch angle is 0.36 and sensor outputs the distance measured at every point (683 3.3.1 Line-building Algorithms steps). Laser beam diameter is less than 20mm Line-building algorithms directly compute a subat 2000mm with maximum divergence 40mm at optimal parametric solution of a multiple struc4000mm. ture, exploiting the property that a single scan is an open chain of sorted points with no loops. However, this methods, requires a prior knowledge of inliers scale. Conseque ntly, basic heuristic conditions work well with simple pseudooutliers, but crowded environme nts and gross outliers require complicated rules being highly dependable of the applicatio n. 3.3.2 Hough Transform

Figure 5: Laser scan observation

The purpose of the method is to find imperfect instances of objects within a certain class of shapes by a voting procedure. This voting procedure is carried out in a parameter space,

from which object candidates are obtained as local maxima in accumulator space that is explicitly constructed by the algorithm for computing the Hough transform. The main drawback of the method is that the parametric space must be discretized; consequently, the accuracy is highly affected in a real time application. As a single majority voting strateg y, it can intrinsically deal with both outliers and pseudo-outliers. 3.3.3 Random Sampling (RANS AC) Consensus

RANSAC assumes the inliers1 are the largest cluster for a predefined scale, which can be an heuristic threshold or obtained from sensor calibration. Conseque ntly, the optimization criteria consist of maximizing the number of inliers. RANSAC is able to cope with large amou nts of outliers, by providing a good scale estimate for every cluster. Nevertheless, the prior scale knowledge is sometimes unavailable or inaccurate.

of being a heuristic 3 threshold. Kernel density estimation and mean shift method are based on the smoothing of the sample distribution. Wide kernels produce oversmoothed pdf, while narrow kernels yield peaked pdf. The performance of the mean shift vector is related to the kernel width. ASSC has the capabili ty to handle multiple scale models even in a single scan. Because of this, it needs a stop criteria to avoid selecting outlier data after having clustered all inliers. As shown in [2], the main advantage of this method, is the capabili ty to compute a data driven scale, which does not require any previous calibration of the sensor or tuning of the parameters. It also showed robustness dealing with different objects in the same scene.. Usually, line extraction algorithms tend to include the doors as a part of the wall segment because the elements are collinear according to the error model or the algorithm cannot deal with multiple structures, in contrast, ASSC is able to distinguish between texture and multiple close structures.

Adapti ve Scale Sample sensus (ASSC)

Con-

Conclusions Work

and

Future

Adaptive Scale Sample Consensus (ASSC) is a modification of RANSAC involving an adapti ve scale estimation. The data driven scale estimate is computed using mean shift meth od2 . The main advantage of ASSC is that the scale factor is computed for every cluster separatel y, instead
A distinct area or formation completely surrounded by another. Merriam- Webster 2 The mean shift paradigm is a nonparametric meth od to find maxima and minima of an unkn own pdf; the pdf is defined by a sample.
1

This paper has presented the basis of a mobile robotic architecture made from scratch. Future work will involve experime nts with the ASSC method in both simulation and in the robotic platform, the development of a minimal circuit to manage various sensors and to control the motors velocity, and the addition of a ring of sonar sensors and bumpers.
3 The process of solving problems by evaluating each step in the progress, searching for satisfactory solution s rather than optimal solutio ns.

References
[1] Sebastian Thrun, Wolfram Burgard, Diete r Fox, Probabilistic Robotics, The MIT Press. 2006. [2] Ruben Martinez-Ca ntin, Jose A. Castellanos, Juan D. Tardos and J. M. M. Montiel, Adapti ve Scale Robust Segmentation for 2D Laser in Proceedings of the 2006 IEEE/RSJ International Confe rence on Intelligent Robots and Systems, Beijing, China, pp. 796-801. [3] Fulvio Mastrogi ovanni, Antonio Sgorbissa and Renato Zaccaria, The More the Better? A Discussion about Line Features for Self-Localization in Proceedings of the 2007 IEEE/RSJ International Confe rence on Intelligent Robots ans Systems, San Diego, CA, USA, pp.3199-3204. [4] Chunlin Chen, Han-Xiong Li, and Daoyi Dong, Hybrid Control for Robot Navigation in Robotics & Automatization Magazine, IEEE/RAS Vol. 15, No. 2 June 2008, Pp.37-47. [5] Viet Nguyen, Agostino Martinelli, Nicola Tomatis, Roland Siegwart, A Comparasion of Line Extraction Algorithms using 2D Laser Rangefinder for indoor Mobile Robotics. [6] Johan Forsberg, Ulf Larson and Ake Wernersson, Mobile Robot Navigation using the Range- Weighted Hough Transformin Robotics & Automatization Magazine, IEEE/RAS March 1995, Pp.18-26.

You might also like