Obstacle Avoidance for Power Wheelchair Using

0 downloads 0 Views 2MB Size Report
to accommodate the wheelchair dimensions, allowing the free- space to be ... training the neural network. A Bayesian ... probable value given by the training data set D during the. Bayesian .... Labview is used to monitor the Laser range data ...
Proceedings of the 29th Annual International Conference of the IEEE EMBS Cité Internationale, Lyon, France August 23-26, 2007.

SaP1D3.4

Obstacle Avoidance for Power Wheelchair Using Bayesian Neural Network Hoang T. Trieu, Hung T. Nguyen, Senior Member, IEEE and Keith Willey, Member, IEEE Abstract—In this paper we present a real-time obstacle avoidance algorithm using a Bayesian neural network for a laser based wheelchair system. The raw laser data is modified to accommodate the wheelchair dimensions, allowing the freespace to be determined accurately in real-time. Data acquisition is performed to collect the patterns required for training the neural network. A Bayesian frame work is applied to determine the optimal neural network structure for the training data. This neural network is trained under the supervision of the Bayesian rule and the obstacle avoidance task is then implemented for the wheelchair system. Initial results suggest this approach provides an effective solution for autonomous tasks, suggesting Bayesian neural networks may be useful for wider assistive technology applications.

P

I.

INTRODUCTION

eople with mobility impairments may experience barriers preventing their full participation in society. Studies have shown that both children and adults benefit substantially from access to a means of independent mobility, including the use of powered wheelchairs. For young children, independent mobility provides a foundation for early learning [1, 2]. A lack of ambulatory and control often produces a cycle of reduced motivation that leads to learning difficulties. For adults, the capacity for independent mobility raises self-esteem [3]. In addition, the vocational and education opportunities significantly increase for people with independent mobility. General obstacle avoidance is one of the most fundamental tasks of autonomous systems. It is also a very important function in designing an intelligent wheelchair system with its specific requirements such as safety, smoothness, and comfort. The major challenge is the development of a strategy for real-time obstacle avoidance. This task assists a robot to navigate and travel in unstructured and unknown environments. Many algorithms for enabling autonomous tasks have been developed. The most popular being, global map, occupancy grid [4], virtual force field and vector field histogram [5, 6]. However, most of these algorithms have difficulty successfully operating in dense environments. For example passing though a doorway or traveling along a narrow corridor. They also do not produce the smooth This work was supported in part by Australian Research Council under Discovery Grant DP0666942 and LIEF Grant LE0668541. Hoang T. Trieu is with Faculty of Engineering, University of Technology, Sydney, Broadway, NSW 2007, Australia (phone: +612-95142451; fax: +61 2 9514 2868; e-mail: [email protected]). Hung T. Nguyen is with Faculty of Engineering, University of Technology, Sydney, Broadway, NSW 2007, Australia (e-mail: [email protected]). Keith Willey is with Faculty of Engineering, University of Technology, Sydney, Broadway, NSW 2007, Australia (e-mail: [email protected]).

trajectory, and stability required in designing an assistive wheelchair system. This project designs an intelligent algorithm based on Bayesian learning techniques that aims to overcome the inherent drawbacks of the contemporary algorithms to improve the stability and robustness for an assistive wheelchair system. The algorithm primarily relies on the ability of neural network to learn how to react in certain difficult situation. After training the network acts as an “expert” when new information arrives. The Bayesian framework improves the network performance by determining the most probable structure and weight values for this application [7, 8]. This paper has the following structure. In the next section the Bayesian framework is reviewed. The wheelchair hardware structure is overviewed in the third section. An obstacle avoidance procedure is presented in the fourth section. The modification of laser signal method, and neural network controller are also presented in this section. The next section presents the research results of this project which are followed by the discussion and conclusion. II. BAYESIAN FRAMEWORK Bayesian learning of a multi-layer perceptron neural network is based on a Gaussian that provides the most probable weight set for the highest generalisation [8, 9]. The weight set w of the network H is adjusted to the most probable value given by the training data set D during the Bayesian learning process. The posterior probability of the network parameters could be estimated by: P ( D | w, H ) P ( w | H ) P ( w | D, H ) = P( D | H ) where P(D|w,H) is the likelihood that contains the information of the weight set determined from the observations, and the prior distribution P(w|H) contains the background knowledge of weight set. In comparing models the denominator, P(D|H), is known as the evidence of network H. The Bayesian approach defines the cost function S(w) as the performance index where: S(w) = βED + αEW where α, β are called hyper-parameters. The parameter β controls the variance of the noise β = 1/σ2, where σ is the standard deviation of Gaussian noise. ED is the training error and EW is the sum square of weight value. The log of the evidence of the hyper-parameters shall be estimated by equation: [7]:

4771 Authorized licensed use limited to: University of Technology Sydney. Downloaded on February 19, 2009 at 00:43 from IEEE Xplore. Restrictions apply.

ln p ( D | α , β ) = − S ( w MP ) −

III.

1 W N N ln A + ln α + ln β − ln( 2π ) 2 2 2 2

where A is the Hessian matrix of the cost function, A = αC + βB, ∇∇EW = C , ∇∇E D = B . The term W is the number of network parameters, N is the number of training patterns and wMP is the most probable value of weight. The most probable values of hyper-parameters αMP, βMP can be estimated by, [7, 10]: W γ , MP N − γ , λi α MP = β = γ = ∑ MP MP 2 EW 2 ED i =1 λ i + α where λi is the eigen value of the Hessian matrix A. The values of the hyper-parameters are used to adjust the training convergence to global or good local minima that are re-estimated during the training process. They also constrain the over growth of weight values ensuring the generalization of the network during the training process. The Bayesian formula for our model, Hi, and its probability are given by the data, p( D | H i ) p( H i ) p( H i | D) = p( D) The prior of model is assumed to be the same for all models; and the term p(D) does not depend on the model. Hence, the posterior probability of the model can be determined by evidence p(D|H). The evidence of the model can be calculated by estimating the integration below over the set of network parameters. p ( D | H i ) = ∫ p ( D | w, H i ) p ( w, H i ) dw Bishop evaluated the log evidence of model, Hi, rather than the evidence itself [7] as: 1 W ln p ( D | H i ) = −α MP EWMP − β MP E DMP − ln | A | + ln α MP + 2 2 1 2 1 2 N + ln β MP + ln M !+2 ln M + ln + ln 2 2 γ 2 N −γ The different network structures are compared by estimating the evidence by the above equation. The optimal network structure is the network that has the highest evidence.

WHEELCHAIR SYSTEM HARDWARE OVERVIEW

The wheelchair is based on the commercial powered wheelchair (Roller M1). In the past few years, we have equipped the wheelchair with several kinds of sensors and support for various control input methods. The hardware of the wheelchair is depicted in figure 1. The original motor driver is used with two encoders being mounted on the motors for estimating wheelchair odometry. The information from the camera, laser systems and the encoder output is transferred directly to an operating computer (Apple-Mac mini). The driving commands are sent to the motor control system via National Instrument USB 6008. The wheelchair system supports several input methods including a conventional joystick control, headmovement input and an EEG brain computer interface. IV. OBSTACLE AVOIDANCE METHOD A. Laser based system The range data is produced by the URG laser system. The method considers the dimensions of the wheelchair. To treat the wheelchair as a point, the obstacles are enlarged to take into account the dimension of the wheelchair r and a safe distance d . The modification procedure consists of a number of steps: 1- Estimate γ = arcsin rr+d/dB, where dB is the distance from the wheelchair to an obstacle. 2- Re-calculate all the scanning distance values that are inside ±γ and store as the new range data. 3- Perform step 1 and 2 for all laser scanning data. For each scan we will achieve a set of range data that includes both original and new values. The modification values are the minimum value of this set. This procedure allows the new values of laser range to be used directly to determine the free-spaces. It removes the need for the laser sensor to be mounted in the centre of the robot. This provides a significant advantage for use with powered wheelchairs where disabled people typically sit in the middle. All spaces smaller than the wheelchair dimension, including the predetermined safety distance, are neglected. The result is shown in figure 2.

Figure 2: The free zone is limited in the shading area. Outer line is the original laser data.

Figure1: The wheelchair’s hardware system

4772 Authorized licensed use limited to: University of Technology Sydney. Downloaded on February 19, 2009 at 00:43 from IEEE Xplore. Restrictions apply.

B. Data acquisition The wheelchair is asked to follow a number of predetermined paths to gather data for training. These paths are selected by the designer to simulate the typical task required of an operational wheelchair. For example chosen paths include traveling along a wall, a corridor, passing a door way and moving in a narrow space. The collected laser range data and motion information are then used to train the neural network. For the obstacle avoidance task, the inputs of neural network are the laser range data while the outputs are the parameters to control the speed and steering. The performance of the neural network depends heavily on the network structure and the training process. The Bayesian framework finds the most probable network parameters and structure to improve the performance of the network. Figure 3 shows a network with six hidden nodes producers the highest evidence. This network is then trained using the Bayesian framework. The most probable weight values are used to set up a feed-forward network that can be used to control the wheelchair in real-time.

Figure 4: The real-time software for SAM wheelchair.

V. EXPERIMENT RESULTS The wheelchair was tested for a number of tasks necessary for any autonomous mobile system, including door passing, following a corridor, entering a narrow area, and following a wall. Labview is used to monitor the Laser range data, velocity, and turning rate in real-time figure 4. Figure 5 shows how the wheelchair performed the “entering the narrow place” task by turning to a narrow corner. The laser range data is noisy as a result of the geometry and obstacles within the testing environment, for example boxes, drawers, tables, chairs. The performance result shown in figure 5 demonstrates the potential of the system. The wheelchair not only keeps a safe distance from all obstacles but its trajectory is smooth. Figure 5: The “entering the narrow space” task realtime experiment.

Figure 3: Evidence versus number of hidden nodes: the solid line is the average evidence of ten training times over the network structures.

The “wall following” task was tested along by the longest wall of the laboratory as depicted in figure 6. The obstacles were positioned in such a way to allow only one free-space option within which the wheelchair may move. The wheelchair travels smoothly from A to B moving slightly away from the wall to remain in the centre of the available path as the free space area increases. The speed of the wheelchair reduces as it approaches point B as the free space narrows. The most difficult tasks are the narrow “corridor traveling” and “door passing”. The doorway and the corridor are narrow compared to the dimensions of the wheelchair being approximately 1m and 1.7m respectively. The wheelchair is 0.7m wide and 1m long.

4773 Authorized licensed use limited to: University of Technology Sydney. Downloaded on February 19, 2009 at 00:43 from IEEE Xplore. Restrictions apply.

VI. DISCUSSION AND CONCLUSION The results demonstrate that a neural network controller provides a promising solution for the task of obstacle avoidance and more generally in semi/autonomous systems. By utilising the learning ability of a neural network this method may overcome many of the inherent shortcomings of contemporary autonomous algorithms. After training the network has the potential to provide satisfactory real-time performance. Compared to VFH, a well-known obstacle avoidance algorithm [6], our neural network algorithm provides better performance in critical tasks such as door passing. A semi-autonomous controller based on the Bayesian neural network will be designed to share control of the wheelchair between a user and the wheelchair sensors. This controller will decide on the most suitable command in a given situation. The Bayesian neural network has demonstrated its potential to be an effective tool in developing smart wheelchair systems. REFERENCES

Figure 6: The “wall following” task experiment.

1. 2. 3. 4. 5.

6.

7. The door 8. Figure 7: The “corridor traveling” and “door passing” tasks real-time experiment.

9.

The wheelchair has some difficulty finding the optimum path around the corner but travels smoothly along the corridor. The velocity of the wheelchair is decreased when it passes the doorway. The wheelchair trajectory when passing the doorway is stable and smooth. All the tasks were performed successfully.

10.

Rosenbloom, L. and K.S. Holt, Consequences of impaired movement: A hypothesis and review. Movement and Child Development, 1975. Simpson, R.C., Smart wheelchairs: A literature review. Journal of Rehabilitation Reasearch & Development, 2005. 42(4): p. 423-436. Wright, B.A., Physical Disability—A Psychosocial Approach. 1983: New York: Harper & Row. Elfes, A., Using occupancy grids for mobile robot perception and navigation. Computer, 1989. 22(6): p. 46 - 57. Borenstein, J. and Y. Koren, Real-time obstacle avoidance for fast mobile robots in cluttered environments. IEEE International Conference on Robotics and Automation, 1990. 1: p. 572 - 577. Borenstein, J. and Y. Koren, The vector field histogram-fast obstacle avoidance for mobile robots. IEEE Transactions on Robotics and Automation, 1991. 7(3): p. 278 - 288. Bishop, C.M., Neural networks for pattern recognition. 1995, Oxford: Oxford University Press. MacKay, D.J.C., Bayesian interpolation. Neural Computation, 1992a. 4(3): p. 415–447. MacKay, D.J.C., A practical Bayesian framework for backpropagation networks. Neural Computation, 1992b. 4: p. 448-472. Penny, W.D. and S.J. Roberts, Bayesian neural networks for classification: how useful is the evidence framework? Neural Networks, 1999. 12(6): p. 877-892.

4774 Authorized licensed use limited to: University of Technology Sydney. Downloaded on February 19, 2009 at 00:43 from IEEE Xplore. Restrictions apply.