indoor location detection by arezou pourmir ece 539 project instructor: professor yu hen hu

11
Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Upload: wilfred-small

Post on 14-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Indoor Location Detection

By

Arezou Pourmir

ECE 539 project

Instructor: Professor Yu Hen Hu

Page 2: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Problem Definition

• Definition:

This project is a new idea of using the acoustic signal in indoor location detection.

• What do we mean by “Indoor Location Detection”?– It is a location “Detection” problem, not location Estimation!– “Indoor” location detection no GPS signal is available!

• Differences from indoor Localization:– Indoor localization answers the question: What is my location in a particular

place, like a mall. – But this problem answers this question: Which place, out of my N interested

places (detection), am I in?

Page 3: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

An Example

• Suppose there are 3 places that you want your mobile device to detect them automatically after entering them!

• Maybe you want your mobile device to switch

to different modes in different places.

Page 4: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Steps in Solving the Problem• We are using acoustic waveform:

– We send acoustic waveform from the speaker of the mobile device and then try to extract all the reflected components of the received signal from the microphone.

• Extracting the feature vectors:– Delays and attenuation factors of the reflected waveforms

• Preprocessing on feature vectors: • Normalizing the attenuation factors• working with “sum of pairs of delays” instead of the delays in all the above cases!• ignoring irrelevant features (like the reflections from the mobile device itself)• ignoring redundant features (like close reflections)

• Applying some classification methods to the extracted feature vectors. I tried:– K-Nearest-Neighbors – ML– MAP

Page 5: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Training and Testing Phase

• Training Part:– For each place of our interest we need to train the mobile device. I

sent some waveforms (5 pulses (periods) of sine wave at frequency of 2KHz) and then record the received signal immediately heard after that, and extract and save the feature vectors.

– We repeated this at 10 different random points in each place and saved the index of that room.

• Testing Part:– To detect a new place we do the same thing. Send the same

waveform and record the reflected signal, and extract the feature vector from the reflected waveforms.

Page 6: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Feature Extraction

• I used correlation for extracting the feature vectors.

• I correlated the original Sine wave with the first 0.1 ms of the signal recorded immediately after the sent waveform.

• Then I calculated the peaks of the envelope of the absolute value of the correlator output. The locations of the peaks tells us the delays and the reciprocal of their amplitudes shows their attenuation factors.

Page 7: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Preprocessing on Feature Vectors

• Before passing the extracted features to the classification block, I applied:

– Normalizing the attenuation factors in each sample.– I also worked with “sums of pairs of delays” instead of

the delays!– Ignoring irrelevant features: for example the reflections

from the mobile device itself– Ignoring redundant features: for example the close

reflections

Page 8: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Classification Section

• Applying some classification methods to the extracted feature vectors. I tried:

– K Nearest Neighbors (K=5)–ML–MAP

For the distribution extraction part I applied Parzen window on the extracted samples.

Page 9: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Results

• Experiment setup:– There number of interested places is 3.– For each of them in the training phase there are 10

feature vectors, each of them consists of 10 delays with their corresponding normalized attenuation factor.

– Maximum delays considered: 0.1 ms

• The classification percentage is around 80% to 90%. • The simple KNN (K=5) works a little better than ML in

this problem.

Page 10: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Future works

• Source of errors:– It is obvious that there is no perfect solution for this problem with zero

probability of error, (like distinguishing rooms of exactly similar dimensions and decorations.)

– Feature extraction method. We can use Golay code instead of Sine wave and hopefully it will improve the accuracy of the feature extraction part a lot. (That will also help a lot in determining the exact start point in the received signal.)

– This problem (like many real cases) can be very sensitive to the training phase.

– Changes in the decoration of the rooms can affect on the result so updating the trained data is required.

– Other classification methods might work better than the currently tested ones, like SVM or MLP.

– Maybe a better feature transformation can be found!

Page 11: Indoor Location Detection By Arezou Pourmir ECE 539 project Instructor: Professor Yu Hen Hu

Question?