sirwh / pamap

Using CNN to classify PAMAP (physical activities) data

PAMAP

Using CNN to classify physical activity data

Environment: Matlab + Python 2.7 + Keras (theano)

The data preprocessed by Matlab is saved as input_data.mat, with this file and Keras installed, anyone can run PAMAP_v2.py to create a CNN and classify the data

The reference paper is "Time Series Classification Using Multi-Channels Deep CNN", by Yi Zheng et al, 2014

The whole process is:

1. get the database of PAMAP2, which is available at https://archive.ics.uci.edu/ml/datasets/PAMAP2+Physical+Activity+Monitoring

2. select the data

   Subject: #1~#7 (#8 and #9 either lack the following activities or have different dominant hand/foot.)

   Activity: 3(standing), 4(walking), 12(ascending stairs), 13(descending stairs)

   Data: there are 3 IMUs (hand, chest, ankle) and we cannot determine which one is better for classification, so we should test all three.

   We use 3d-acceleration data rather than gyroscope because gyro data is characteristic-less for at least chest position

   So the data of one subject has the format of:

    Rawdata[1,2,5,6,7,22,23,24,39,40,41]
    
    which is [time, activityID, hand_acc_x,y,z, chest_acc_x,y,z, ankle_acc_x,y,z] 
   

   Then we put the data from 7 subjects into 1 cell.

    subject_data={subject1,subject2,subject3,subject4,subject5,subject6,subject7} 
   

3. preprocessing (refer to the files in "PAMAP_Matlab" folder)

   We normalize each dimension of 3D time series as (x−μ)/σ , where μ and σ are mean and standard deviation of time series. (as in processed3.mat)

   Then we apply the sliding window algorithm to extract subsequences from 3D time series with different sliding steps.

    step size: 128, 64, 32, 16, 8 in the reference paper, we only test 128 to reduce calculation time
    
    window length: 256
   

   Now we have:

    input_data{1…7}{1…9}, where input_data{i}{j} is a matrix of Nx257, which is N observation*(1d label + 256d data)
    
     (i for subject 1~7, j for acceleration data 1~9)
   

   The result is saved as input_data.mat

4. divide training and test set (refer to the file of PAMAP_v2.py)

   To evaluate the performance of different models, we adopt the leave-one-out cross validation (LOOCV) technique. Specifically, each time we use one subject’s physical activities as test data, and the physical activities of remaining subjects as training data. Then we repeat this for every subject.

5. result

   CNN with input data of IMU (hand)

    train acc: 0.78, 0.76, 0.77, 0.79, 0.81, 0.76, 0.79
   
    test acc:  0.53, 0.77, 0.77, 0.79, 0.56, 0.74, 0.76 (avg 0.70)
   

   CNN with input data of IMU (chest)

    train acc: 0.87, 0.89, 0.88, 0.89, 0.88, 0.89, 0.87
   
    test acc:  0.86, 0.74, 0.89, 0.86, 0.90, 0.90, 0.88 (avg 0.86)	
   

   CNN with input data of IMU (ankle)

    train acc: 0.91, 0.89, 0.87, 0.88, 0.89, 0.90, 0.90
   
    test acc:  0.83, 0.89, 0.81, 0.90, 0.90, 0.85, 0.81 (avg 0.86)
   

   Only use the MLP part of the CNN, with input data of IMU (chest)

    train acc: 0.76, 0.79, 0.76, 0.76, 0.78, 0.77, 0.80
   
    test acc:  0.77, 0.20, 0.75, 0.75, 0.54, 0.74, 0.39 (avg 0.59)
   

   so the CNN model has a 0.27 accuracy boost compared to simple MLP