Create Presentation
Download Presentation

Download Presentation
## A Comparison of Algorithms for User Target Intention Recognition

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -

**A Comparison of Algorithms for User Target Intention**Recognition Gökçen Aslan Aydemir, Pat Langdon, Simon Godsill Department of Engineering, University of Cambridge, UK {ga283,pml24}@cam.ac.uk sjg@eng.cam.ac.uk International Workshop on Personalisable Media Systems& Smart Accessibility co-located at NEM Summit 2012 October 17th, 2012 in Istanbul, Turkey**Outline**• Introduction and Motivation • Intention Recognition from Bearing Angle • Intention Recognition with Kalman Filter • Comparison • Conclusion IWPMSA12,NEM SUMMIT, 17/10/2012**Introduction**• Objectives • Reduce time spent to select a target • Reduce difficulty important for motor impaired people • Smooth cursor movement IWPMSA12,NEM SUMMIT, 17/10/2012**So far…**• Angle based target prediction [Murata,1998] • Linear Regression from peak velocity – target distance relationship [Asano et.al, 2005] • Quadratic function fitting to partial trajectory to predict endpoint. [McGuffin, Balakrishnan,2005 ] However: • Cursor trajectories and velocity profiles are different for users with motor impairments IWPMSA12,NEM SUMMIT, 17/10/2012**Motivation**Prediction and smoothing is especially important for users with motor impairments. IWPMSA12,NEM SUMMIT, 17/10/2012**T1**T2 T3 T4 T5 φT3 φT2 φT4 φT1 φT5 Current cursor pos (t) Previous cursor pos (t-1) Total Angle to Bearing Algorithm IWPMSA12,NEM SUMMIT, 17/10/2012**Total Angle to Bearing Algorithm**IWPMSA12,NEM SUMMIT, 17/10/2012**A common cursor trace and prediction**Intention Recognition from Bearing Angle IWPMSA12,NEM SUMMIT, 17/10/2012**Intention Recognition from Bearing Angle**• A cursor trace with overshoot over target IWPMSA12,NEM SUMMIT, 17/10/2012**How does it perform?**• Performs well for able bodied users and relatively straight cursor paths • Cannot handle extreme cases • Cannot handle targets on the same movement direction • Performance depends on buffer size/history brings computational complexity IWPMSA12,NEM SUMMIT, 17/10/2012**Intention Recognition with Kalman Filter**• A Markov Model • Current state only depends on previous state • No need to store history • Nearly Constant Velocity Model • Velocity is driven by white noise • Noise std can be personalized • Uniform probability assumed for all targets at the beginning • Probability is penalized/updated • Angle • Distance • Angle and distance IWPMSA12,NEM SUMMIT, 17/10/2012**Target probability**• For N targets , at t = 0, uniform probability pi = 1/N • t > 0 pi+I = pi / distance_to_target pi+I = pi / angle_to_target pi+I = pi / (angle_to_target* distance_to_target) Norm(pi+I ) IWPMSA12,NEM SUMMIT, 17/10/2012**How does it perform?**• Less computational complexity • No need to keep history • Smooth and predict at the same time • Performance is not improved significantly for able-bodied users • Extreme cases are handled more robustly IWPMSA12,NEM SUMMIT, 17/10/2012**Comparison – Case 1 Prediction with bearing angle**IWPMSA12,NEM SUMMIT, 17/10/2012**Comparison – Case 1 Prediction with bearing angle**IWPMSA12,NEM SUMMIT, 17/10/2012**Comparison – Case 1 Prediction with Kalman Filter**IWPMSA12,NEM SUMMIT, 17/10/2012**Comparison – Case 2 Prediction with bearing angle**IWPMSA12,NEM SUMMIT, 17/10/2012**Comparison – Case 2 Prediction with Kalman Filter**IWPMSA12,NEM SUMMIT, 17/10/2012**Comparison – Case 2 Prediction with Kalman Filter**IWPMSA12,NEM SUMMIT, 17/10/2012**Conclusion**• 471 cases • Alternative firing mechanism Provide intended target only if the same target is predicted for k times. • Bearing algorithm failed to stay on correct target prediction in 35 cases • Prediction availability IWPMSA12,NEM SUMMIT, 17/10/2012**Conclusion**• 471 cases • Alternative firing mechanism Provide intended target only if the same target is predicted for k times. • Bearing algorithm failed to stay on correct target prediction in 35 cases • Prediction availability • Prediction accuracy correct prediction/all predictions for a task • Bearing : 64% • Kalman(a+d) : 65% IWPMSA12,NEM SUMMIT, 17/10/2012**Conclusion**• Kalman Filter provides more accurate results especially non-standard cursor traces • Kalman Filter does not provide better prediction for able-bodied users but requires less memory • Availability of prediction is greater with Kalman Filter • Using distance and angle together provides more robust prediction and higher accuracy • It is possible to provide first correct prediction quicker with Kalman Filter IWPMSA12,NEM SUMMIT, 17/10/2012**Conclusion**• Could assist target expansion algorithms • Provide user with a target suggestion • Performances should be investigated for • more complex layouts • pointing devices • effects of target size and distance • more complex process models