130 likes | 272 Views
Multimodal and Sensorial Interfaces for Mobile Robots course task. Nicola Piotto a.y. 2007/2008. Specifics about the task. Robertino has been positioned at different distances from an obstacle (i.e. 0.125, 0.25, 0.5, 1, 2, 3 [meters]).
E N D
Multimodal and Sensorial Interfaces for Mobile Robots course task Nicola Piotto a.y. 2007/2008
Specifics about the task • Robertino has been positioned at different distances from an obstacle (i.e. 0.125, 0.25, 0.5, 1, 2, 3 [meters]). • For each step several measurements from the frontal IR sensor has been collected. • The final goal is to define a function to map the noisy sensor data to the real object distance.
Initial data observations • More the object distance increases the noise increases as well.
Employed solution • The solution to the problem can be reached using a linear regression over the acquired sensor data. • In this way it is possible to analitycally define a linear function throughout least squared error minimization (data fitting). • The derived function maps sensor data to object distances.
Considerations • It has been tried to retrieve an higher degree interpolating function (e.g. polinomial, quadratic) using a Support Vector Regression (SVR) procedure: however, due to the noise in the observed data it has not been possible to successfully end the task (the final result was unreliable).
Some specification • The linear regression has been implemented in Matlab environment. • [b,c]=regress(x,y) • Y=c+b*X • Y is the estimated object distance. • X is the sensor measurement • x is the training data • y is the related distance
Results • The matrix z includes in 2 columns all the training sensor data(1) and the distance they refer to(2) (bracketed results refre to different set of data considered). • b=0.013;(0.014);(0.016); • MSE=0.0849;(0.033);(8.9269*10^-5)
Results(2) 0.125,0.25,0.5 0.125,0.25,0.5,1 0.125,0.25,0.5,1,2,3
Considerations • Including also the noisy data from the bigger object distance leads to a calibration function not particularly precise (high MSE) . • Instead, considering only the less noisy data from the smallest object distance (up to 0.5-1 m) leads to a more reliable calibration function.
Considerations(2) • Given the impossibility in processing the information from the more noisy distances, it may be suggested to employ the IR sensor to estimate object distances up to 1 meter. • For bigger distances is not achieved an sufficient precision so it may be better employ different kind of sensor.