We are using the MobiAct Dataset. It includes four different types of falls and eleven types of activities of daily living. The files were generated by 66 subjects. Their age, height, gender and weight are different.
All in all the set hast about 3300 CSV-files. We didn't import all files in our model. The files contain data of the accelerometer, gyroscope and orientation.
In this picture, you can see the distribution of the two labels. There are more activities of daily living (ADL) then falls.
Here are some example values of a fall file. The column "label" was added by our program to process the data better:
.png?raw=true)
This picture describes the data of a fall.
for a in range(1,4): # 3 tries for everyone
for i in range(1,67): # 66 subjects
if os.path.exists(path+"_{0}_{1}_annotated.csv".format(i,a)): #check if path exist
data=pd.read_csv(path+"_{0}_{1}_annotated.csv".format(i,a),usecols=["rel_time","acc_x","acc_y","acc_z","gyro_x","gyro_y","gyro_z","azimuth","pitch","roll"])
This code excerpt shows the read-in of the 3300 files in a Pandas dataframe. In the further course, we added a label to the dataframe and changed it in a Numpy Array. To ensure that every file has the same length, we used this command:
keras.preprocessing.sequence.pad_sequences(data, maxlen=train_data.shape[1], dtype='int32', padding='pre', truncating='pre', value=0.0)
In addition, we wrote an extra script, which cuts all CSV-files, which were longer than 10 seconds.
G. Vavoulas, M. Pediaditis, E. Spanakis, M. Tsiknakis, The MobiFall Dataset:
An Initial Evaluation of Fall Detection Algorithms Using Smartphones,
6th IEEE International Symposium on Monitoring & Surveillance Research (ISMSR):
Healthcare-Bioinformatics, Chania, Greece, 2013, DOI:10.1109/BIBE.2013.6701629.
| Go back to homepage | More project details | Code of the model | Code of the evaluation program | The app | Accuracy | The team | Our Video |