7.1 C
United States of America
Friday, November 15, 2024

Classifying bodily exercise from smartphone knowledge


Introduction

On this put up we’ll describe how one can use smartphone accelerometer and gyroscope knowledge to foretell the bodily actions of the people carrying the telephones. The information used on this put up comes from the Smartphone-Primarily based Recognition of Human Actions and Postural Transitions Knowledge Set distributed by the College of California, Irvine. Thirty people had been tasked with performing numerous fundamental actions with an connected smartphone recording motion utilizing an accelerometer and gyroscope.

Earlier than we start, let’s load the varied libraries that we’ll use within the evaluation:


library(keras)     # Neural Networks
library(tidyverse) # Knowledge cleansing / Visualization
library(knitr)     # Desk printing
library(rmarkdown) # Misc. output utilities 
library(ggridges)  # Visualization

Actions dataset

The information used on this put up come from the Smartphone-Primarily based Recognition of Human Actions and Postural Transitions Knowledge Set(Reyes-Ortiz et al. 2016) distributed by the College of California, Irvine.

When downloaded from the hyperlink above, the info comprises two completely different ‘components.’ One which has been pre-processed utilizing numerous function extraction strategies resembling fast-fourier rework, and one other RawData part that merely provides the uncooked X,Y,Z instructions of an accelerometer and gyroscope. None of the usual noise filtering or function extraction utilized in accelerometer knowledge has been utilized. That is the info set we’ll use.

The motivation for working with the uncooked knowledge on this put up is to help the transition of the code/ideas to time sequence knowledge in much less well-characterized domains. Whereas a extra correct mannequin may very well be made by using the filtered/cleaned knowledge supplied, the filtering and transformation can fluctuate tremendously from job to job; requiring numerous guide effort and area data. One of many stunning issues about deep studying is the function extraction is discovered from the info, not exterior data.

Exercise labels

The information has integer encodings for the actions which, whereas not essential to the mannequin itself, are useful to be used to see. Let’s load them first.


activityLabels <- learn.desk("knowledge/activity_labels.txt", 
                             col.names = c("quantity", "label")) 

activityLabels %>% kable(align = c("c", "l"))
1 WALKING
2 WALKING_UPSTAIRS
3 WALKING_DOWNSTAIRS
4 SITTING
5 STANDING
6 LAYING
7 STAND_TO_SIT
8 SIT_TO_STAND
9 SIT_TO_LIE
10 LIE_TO_SIT
11 STAND_TO_LIE
12 LIE_TO_STAND

Subsequent, we load within the labels key for the RawData. This file is an inventory of the entire observations, or particular person exercise recordings, contained within the knowledge set. The important thing for the columns is taken from the info README.txt.


Column 1: experiment quantity ID, 
Column 2: person quantity ID, 
Column 3: exercise quantity ID 
Column 4: Label begin level 
Column 5: Label finish level 

The beginning and finish factors are in variety of sign log samples (recorded at 50hz).

Let’s check out the primary 50 rows:


labels <- learn.desk(
  "knowledge/RawData/labels.txt",
  col.names = c("experiment", "userId", "exercise", "startPos", "endPos")
)

labels %>% 
  head(50) %>% 
  paged_table()

File names

Subsequent, let’s take a look at the precise recordsdata of the person knowledge supplied to us in RawData/


dataFiles <- record.recordsdata("knowledge/RawData")
dataFiles %>% head()

[1] "acc_exp01_user01.txt" "acc_exp02_user01.txt"
[3] "acc_exp03_user02.txt" "acc_exp04_user02.txt"
[5] "acc_exp05_user03.txt" "acc_exp06_user03.txt"

There’s a three-part file naming scheme. The primary half is the kind of knowledge the file comprises: both acc for accelerometer or gyro for gyroscope. Subsequent is the experiment quantity, and final is the person Id for the recording. Let’s load these right into a dataframe for ease of use later.


fileInfo <- data_frame(
  filePath = dataFiles
) %>%
  filter(filePath != "labels.txt") %>% 
  separate(filePath, sep = '_', 
           into = c("kind", "experiment", "userId"), 
           take away = FALSE) %>% 
  mutate(
    experiment = str_remove(experiment, "exp"),
    userId = str_remove_all(userId, "person|.txt")
  ) %>% 
  unfold(kind, filePath)

fileInfo %>% head() %>% kable()
01 01 acc_exp01_user01.txt gyro_exp01_user01.txt
02 01 acc_exp02_user01.txt gyro_exp02_user01.txt
03 02 acc_exp03_user02.txt gyro_exp03_user02.txt
04 02 acc_exp04_user02.txt gyro_exp04_user02.txt
05 03 acc_exp05_user03.txt gyro_exp05_user03.txt
06 03 acc_exp06_user03.txt gyro_exp06_user03.txt

Studying and gathering knowledge

Earlier than we are able to do something with the info supplied we have to get it right into a model-friendly format. This implies we wish to have an inventory of observations, their class (or exercise label), and the info comparable to the recording.

To acquire this we’ll scan by way of every of the recording recordsdata current in dataFiles, lookup what observations are contained within the recording, extract these recordings and return every little thing to a simple to mannequin with dataframe.


# Learn contents of single file to a dataframe with accelerometer and gyro knowledge.
readInData <- perform(experiment, userId){
  genFilePath = perform(kind) {
    paste0("knowledge/RawData/", kind, "_exp",experiment, "_user", userId, ".txt")
  }  
  
  bind_cols(
    learn.desk(genFilePath("acc"), col.names = c("a_x", "a_y", "a_z")),
    learn.desk(genFilePath("gyro"), col.names = c("g_x", "g_y", "g_z"))
  )
}

# Perform to learn a given file and get the observations contained alongside
# with their lessons.

loadFileData <- perform(curExperiment, curUserId) {
  
  # load sensor knowledge from file into dataframe
  allData <- readInData(curExperiment, curUserId)

  extractObservation <- perform(startPos, endPos){
    allData[startPos:endPos,]
  }
  
  # get statement places on this file from labels dataframe
  dataLabels <- labels %>% 
    filter(userId == as.integer(curUserId), 
           experiment == as.integer(curExperiment))
  

  # extract observations as dataframes and save as a column in dataframe.
  dataLabels %>% 
    mutate(
      knowledge = map2(startPos, endPos, extractObservation)
    ) %>% 
    choose(-startPos, -endPos)
}

# scan by way of all experiment and userId combos and collect knowledge right into a dataframe. 
allObservations <- map2_df(fileInfo$experiment, fileInfo$userId, loadFileData) %>% 
  right_join(activityLabels, by = c("exercise" = "quantity")) %>% 
  rename(activityName = label)

# cache work. 
write_rds(allObservations, "allObservations.rds")
allObservations %>% dim()

Exploring the info

Now that we now have all the info loaded together with the experiment, userId, and exercise labels, we are able to discover the info set.

Size of recordings

Let’s first take a look at the size of the recordings by exercise.


allObservations %>% 
  mutate(recording_length = map_int(knowledge,nrow)) %>% 
  ggplot(aes(x = recording_length, y = activityName)) +
  geom_density_ridges(alpha = 0.8)

The very fact there’s such a distinction in size of recording between the completely different exercise sorts requires us to be a bit cautious with how we proceed. If we prepare the mannequin on each class without delay we’re going to must pad all of the observations to the size of the longest, which would depart a big majority of the observations with an enormous proportion of their knowledge being simply padding-zeros. Due to this, we’ll match our mannequin to simply the most important ‘group’ of observations size actions, these embrace STAND_TO_SIT, STAND_TO_LIE, SIT_TO_STAND, SIT_TO_LIE, LIE_TO_STAND, and LIE_TO_SIT.

An attention-grabbing future course can be trying to make use of one other structure resembling an RNN that may deal with variable size inputs and coaching it on all the info. Nevertheless, you’d run the chance of the mannequin studying merely that if the statement is lengthy it’s almost definitely one of many 4 longest lessons which might not generalize to a state of affairs the place you had been working this mannequin on a real-time-stream of knowledge.

Filtering actions

Primarily based on our work from above, let’s subset the info to simply be of the actions of curiosity.


desiredActivities <- c(
  "STAND_TO_SIT", "SIT_TO_STAND", "SIT_TO_LIE", 
  "LIE_TO_SIT", "STAND_TO_LIE", "LIE_TO_STAND"  
)

filteredObservations <- allObservations %>% 
  filter(activityName %in% desiredActivities) %>% 
  mutate(observationId = 1:n())

filteredObservations %>% paged_table()

So after our aggressive pruning of the info we can have a decent quantity of knowledge left upon which our mannequin can be taught.

Coaching/testing cut up

Earlier than we go any additional into exploring the info for our mannequin, in an try and be as truthful as potential with our efficiency measures, we have to cut up the info right into a prepare and take a look at set. Since every person carried out all actions simply as soon as (except one who solely did 10 of the 12 actions) by splitting on userId we’ll be sure that our mannequin sees new folks solely once we take a look at it.


# get all customers
userIds <- allObservations$userId %>% distinctive()

# randomly select 24 (80% of 30 people) for coaching
set.seed(42) # seed for reproducibility
trainIds <- pattern(userIds, measurement = 24)

# set the remainder of the customers to the testing set
testIds <- setdiff(userIds,trainIds)

# filter knowledge. 
trainData <- filteredObservations %>% 
  filter(userId %in% trainIds)

testData <- filteredObservations %>% 
  filter(userId %in% testIds)

Visualizing actions

Now that we now have trimmed our knowledge by eradicating actions and splitting off a take a look at set, we are able to really visualize the info for every class to see if there’s any instantly discernible form that our mannequin might be able to choose up on.

First let’s unpack our knowledge from its dataframe of one-row-per-observation to a tidy model of all of the observations.


unpackedObs <- 1:nrow(trainData) %>% 
  map_df(perform(rowNum){
    dataRow <- trainData[rowNum, ]
    dataRow$knowledge[[1]] %>% 
      mutate(
        activityName = dataRow$activityName, 
        observationId = dataRow$observationId,
        time = 1:n() )
  }) %>% 
  collect(studying, worth, -time, -activityName, -observationId) %>% 
  separate(studying, into = c("kind", "course"), sep = "_") %>% 
  mutate(kind = ifelse(kind == "a", "acceleration", "gyro"))

Now we now have an unpacked set of our observations, let’s visualize them!


unpackedObs %>% 
  ggplot(aes(x = time, y = worth, coloration = course)) +
  geom_line(alpha = 0.2) +
  geom_smooth(se = FALSE, alpha = 0.7, measurement = 0.5) +
  facet_grid(kind ~ activityName, scales = "free_y") +
  theme_minimal() +
  theme( axis.textual content.x = element_blank() )

So not less than within the accelerometer knowledge patterns undoubtedly emerge. One would think about that the mannequin might have bother with the variations between LIE_TO_SIT and LIE_TO_STAND as they’ve an analogous profile on common. The identical goes for SIT_TO_STAND and STAND_TO_SIT.

Preprocessing

Earlier than we are able to prepare the neural community, we have to take a few steps to preprocess the info.

Padding observations

First we’ll resolve what size to pad (and truncate) our sequences to by discovering what the 98th percentile size is. By not utilizing the very longest statement size this can assist us keep away from extra-long outlier recordings messing up the padding.


padSize <- trainData$knowledge %>% 
  map_int(nrow) %>% 
  quantile(p = 0.98) %>% 
  ceiling()
padSize

98% 
334 

Now we merely must convert our record of observations to matrices, then use the tremendous useful pad_sequences() perform in Keras to pad all observations and switch them right into a 3D tensor for us.


convertToTensor <- . %>% 
  map(as.matrix) %>% 
  pad_sequences(maxlen = padSize)

trainObs <- trainData$knowledge %>% convertToTensor()
testObs <- testData$knowledge %>% convertToTensor()
  
dim(trainObs)

[1] 286 334   6

Fantastic, we now have our knowledge in a pleasant neural-network-friendly format of a 3D tensor with dimensions (<num obs>, <sequence size>, <channels>).

One-hot encoding

There’s one final thing we have to do earlier than we are able to prepare our mannequin, and that’s flip our statement lessons from integers into one-hot, or dummy encoded, vectors. Fortunately, once more Keras has provided us with a really useful perform to do exactly this.


oneHotClasses <- . %>% 
  {. - 7} %>%        # convey integers right down to 0-6 from 7-12
  to_categorical() # One-hot encode

trainY <- trainData$exercise %>% oneHotClasses()
testY <- testData$exercise %>% oneHotClasses()

Modeling

Structure

Since we now have temporally dense time-series knowledge we’ll make use of 1D convolutional layers. With temporally-dense knowledge, an RNN has to be taught very lengthy dependencies as a way to choose up on patterns, CNNs can merely stack a couple of convolutional layers to construct sample representations of considerable size. Since we’re additionally merely in search of a single classification of exercise for every statement, we are able to simply use pooling to ‘summarize’ the CNNs view of the info right into a dense layer.

Along with stacking two layer_conv_1d() layers, we’ll use batch norm and dropout (the spatial variant(Tompson et al. 2014) on the convolutional layers and normal on the dense) to regularize the community.


input_shape <- dim(trainObs)[-1]
num_classes <- dim(trainY)[2]

filters <- 24     # variety of convolutional filters to be taught
kernel_size <- 8  # what number of time-steps every conv layer sees.
dense_size <- 48  # measurement of our penultimate dense layer. 

# Initialize mannequin
mannequin <- keras_model_sequential()
mannequin %>% 
  layer_conv_1d(
    filters = filters,
    kernel_size = kernel_size, 
    input_shape = input_shape,
    padding = "legitimate", 
    activation = "relu"
  ) %>%
  layer_batch_normalization() %>%
  layer_spatial_dropout_1d(0.15) %>% 
  layer_conv_1d(
    filters = filters/2,
    kernel_size = kernel_size,
    activation = "relu",
  ) %>%
  # Apply common pooling:
  layer_global_average_pooling_1d() %>% 
  layer_batch_normalization() %>%
  layer_dropout(0.2) %>% 
  layer_dense(
    dense_size,
    activation = "relu"
  ) %>% 
  layer_batch_normalization() %>%
  layer_dropout(0.25) %>% 
  layer_dense(
    num_classes, 
    activation = "softmax",
    identify = "dense_output"
  ) 

abstract(mannequin)

______________________________________________________________________
Layer (kind)                   Output Form                Param #    
======================================================================
conv1d_1 (Conv1D)              (None, 327, 24)             1176       
______________________________________________________________________
batch_normalization_1 (BatchNo (None, 327, 24)             96         
______________________________________________________________________
spatial_dropout1d_1 (SpatialDr (None, 327, 24)             0          
______________________________________________________________________
conv1d_2 (Conv1D)              (None, 320, 12)             2316       
______________________________________________________________________
global_average_pooling1d_1 (Gl (None, 12)                  0          
______________________________________________________________________
batch_normalization_2 (BatchNo (None, 12)                  48         
______________________________________________________________________
dropout_1 (Dropout)            (None, 12)                  0          
______________________________________________________________________
dense_1 (Dense)                (None, 48)                  624        
______________________________________________________________________
batch_normalization_3 (BatchNo (None, 48)                  192        
______________________________________________________________________
dropout_2 (Dropout)            (None, 48)                  0          
______________________________________________________________________
dense_output (Dense)           (None, 6)                   294        
======================================================================
Whole params: 4,746
Trainable params: 4,578
Non-trainable params: 168
______________________________________________________________________

Coaching

Now we are able to prepare the mannequin utilizing our take a look at and coaching knowledge. Observe that we use callback_model_checkpoint() to make sure that we save solely the very best variation of the mannequin (fascinating since in some unspecified time in the future in coaching the mannequin might start to overfit or in any other case cease enhancing).


# Compile mannequin
mannequin %>% compile(
  loss = "categorical_crossentropy",
  optimizer = "rmsprop",
  metrics = "accuracy"
)

trainHistory <- mannequin %>%
  match(
    x = trainObs, y = trainY,
    epochs = 350,
    validation_data = record(testObs, testY),
    callbacks = record(
      callback_model_checkpoint("best_model.h5", 
                                save_best_only = TRUE)
    )
  )

The mannequin is studying one thing! We get a decent 94.4% accuracy on the validation knowledge, not unhealthy with six potential lessons to select from. Let’s look into the validation efficiency a little bit deeper to see the place the mannequin is messing up.

Analysis

Now that we now have a skilled mannequin let’s examine the errors that it made on our testing knowledge. We are able to load the very best mannequin from coaching based mostly upon validation accuracy after which take a look at every statement, what the mannequin predicted, how excessive a chance it assigned, and the true exercise label.


# dataframe to get labels onto one-hot encoded prediction columns
oneHotToLabel <- activityLabels %>% 
  mutate(quantity = quantity - 7) %>% 
  filter(quantity >= 0) %>% 
  mutate(class = paste0("V",quantity + 1)) %>% 
  choose(-number)

# Load our greatest mannequin checkpoint
bestModel <- load_model_hdf5("best_model.h5")

tidyPredictionProbs <- bestModel %>% 
  predict(testObs) %>% 
  as_data_frame() %>% 
  mutate(obs = 1:n()) %>% 
  collect(class, prob, -obs) %>% 
  right_join(oneHotToLabel, by = "class")

predictionPerformance <- tidyPredictionProbs %>% 
  group_by(obs) %>% 
  summarise(
    highestProb = max(prob),
    predicted = label[prob == highestProb]
  ) %>% 
  mutate(
    fact = testData$activityName,
    right = fact == predicted
  ) 

predictionPerformance %>% paged_table()

First, let’s take a look at how ‘assured’ the mannequin was by if the prediction was right or not.


predictionPerformance %>% 
  mutate(end result = ifelse(right, 'Right', 'Incorrect')) %>% 
  ggplot(aes(highestProb)) +
  geom_histogram(binwidth = 0.01) +
  geom_rug(alpha = 0.5) +
  facet_grid(end result~.) +
  ggtitle("Possibilities related to prediction by correctness")

Reassuringly it appears the mannequin was, on common, much less assured about its classifications for the inaccurate outcomes than the proper ones. (Though, the pattern measurement is just too small to say something definitively.)

Let’s see what actions the mannequin had the toughest time with utilizing a confusion matrix.


predictionPerformance %>% 
  group_by(fact, predicted) %>% 
  summarise(depend = n()) %>% 
  mutate(good = fact == predicted) %>% 
  ggplot(aes(x = fact,  y = predicted)) +
  geom_point(aes(measurement = depend, coloration = good)) +
  geom_text(aes(label = depend), 
            hjust = 0, vjust = 0, 
            nudge_x = 0.1, nudge_y = 0.1) + 
  guides(coloration = FALSE, measurement = FALSE) +
  theme_minimal()

We see that, because the preliminary visualization steered, the mannequin had a little bit of bother with distinguishing between LIE_TO_SIT and LIE_TO_STAND lessons, together with the SIT_TO_LIE and STAND_TO_LIE, which even have comparable visible profiles.

Future instructions

The obvious future course to take this evaluation can be to aim to make the mannequin extra basic by working with extra of the provided exercise sorts. One other attention-grabbing course can be to not separate the recordings into distinct ‘observations’ however as an alternative maintain them as one streaming set of knowledge, very like an actual world deployment of a mannequin would work, and see how properly a mannequin might classify streaming knowledge and detect adjustments in exercise.

Gal, Yarin, and Zoubin Ghahramani. 2016. “Dropout as a Bayesian Approximation: Representing Mannequin Uncertainty in Deep Studying.” In Worldwide Convention on Machine Studying, 1050–9.

Graves, Alex. 2012. “Supervised Sequence Labelling.” In Supervised Sequence Labelling with Recurrent Neural Networks, 5–13. Springer.

Kononenko, Igor. 1989. “Bayesian Neural Networks.” Organic Cybernetics 61 (5). Springer: 361–70.

LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. 2015. “Deep Studying.” Nature 521 (7553). Nature Publishing Group: 436.

Reyes-Ortiz, Jorge-L, Luca Oneto, Albert Samà, Xavier Parra, and Davide Anguita. 2016. “Transition-Conscious Human Exercise Recognition Utilizing Smartphones.” Neurocomputing 171. Elsevier: 754–67.

Tompson, Jonathan, Ross Goroshin, Arjun Jain, Yann LeCun, and Christoph Bregler. 2014. “Environment friendly Object Localization Utilizing Convolutional Networks.” CoRR abs/1411.4280. http://arxiv.org/abs/1411.4280.

Related Articles

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Latest Articles