dhoepelman / in4254

IN4254 Smart Phone Sensing

in4254

IN4254 Smart Phone Sensing

in4254's Issues

Calculate & Send in raw & processed data

Calibrate training data

If we have extra time, calibrate the phone to the training data set

Offload normal distribution calculation

Because most of the work needs to be done for #22 anyway, make a simple REST interface to get them from the "cloud" so the profs are satisfied

Make a test fragment

We need a test fragment for that uses our (nonexistant) classifier.

Make a way to measure Received Signal Strength (RSS)

• Get RSS data from Wifi radio per access point
• Put samples into windows (suggested 1 sample per second, window size 120)
• Store raw samples in database
• Store mean + std. deviation in database

Write report 1

• Train and test our app a few times to generate a confusion matrix
• Wite about K-nn
• Write about feature vectors used + overlapping window technique

Make a K-NN classifier

Implement a K-NN classifier

1. [no CSV parsing] Collect samples with activity and remember them as training data set
2. Collect a measurement
3. [just x-axis] Calculate the distance to each of the points in the training data set
   • There are numerous possibilities for distance, probably eucledian or manhattan would be most suitable
4. Select the k points with the least distance (k=sqrt(size of training data set))
5. The measurement gets the activity that is most common among those k points
   • There needs to be some tie-breaking solution (but ties should be logged)

Finetune activity monitoring

• Add Mahalanobis distance
• Add "automated" testing & generation of confusion matrix according to holdout (2N/a;N/a) method. a somewhere between 3 and 10
• Try removing some features, collapsing multiple features into 1 or adding more (see original paper)
• Replace STAIRS with LIFT

Implement measuring and feature vectors

• Measuring
• Mean
• StdDev
• Correlation
• FeatureVectors

Measure RSS Data on 9th floor

Dependent on #10 and #9

Gather new training data

We should gather new training data and store it in the db (#9)

final-report-and-demo

Final report

Make a GUI for localization

• Make a fragment with a map of 9th floor EWI
• Allow user to select cell
• Training button

& GUi elements from lecture

Make measurements windows overlap during testing

Currently our windows overlap during training, as is intended. However, during testing we do not overlap our windows. Not only might this skew results but it also increases the interval of measurements from 2,5s to 5s.

Gather Test data

We need to gather test data for our report

• Make something that allows us to easily gather test data (e.g. modify TrainFragment or TestFragment)
• Gather a large enough training set (10 samples per activity?)
• Take measurements for each activity (~50 measures per activity?) and compile them into a table

Make a classifier for localization

Particle filters if we have enough time

Rotation bug training fragment

If the smartphone it turned while training, the training stops.

Solution is probably to create a background taks (headless fragment, AsyncTask) to do the measuring in, instead of directly from the training fragment. This might also be beneficial for the testing fragment.

Not all activity icons are fully visible on Galaxy S

5th icon is only 20 % visible. It can still be selected though. Icons should be scaled or moved to a second row.

Screen information: 4" 800×480 px WVGA (233 ppi)

Write report 2

Save training data in SQLite

I was thinking about sqlite, as this still would enable easy file transfers to analyse data and sqlite also is reasonably easy to use in e..g gnuplot (plot < 'sqlite x.db SELECT * FROM training') while still being a full-flegded SQL database

I was thinking about the following tables:

measurements (ID, timestamp, activity)
measurements_raw_samples (timestamp, measurement_id, X, Y, Z)
features (ID, feature_name)
measurements_features (measurement_ID, feature_ID, value)

I realised there isn't really difference between training data en testing data, as we currently also store an expected activiy with the testing data. Thus we can use our testing data as training data and vice versa :)

What do you think about the features? Should we save them like this or in a flat table?

Feature branches voor de site

Make NSA fragment

Follow someone else with Google Cloud API