Classifying bodily exercise from smartphone knowledge


On this put up we’ll describe tips on how to 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-Based mostly Recognition of Human Actions and Postural Transitions Information Set distributed by the College of California, Irvine. Thirty people had been tasked with performing varied fundamental actions with an hooked up smartphone recording motion utilizing an accelerometer and gyroscope.

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

library(keras)     # Neural Networks
library(tidyverse) # Information 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-Based mostly Recognition of Human Actions and Postural Transitions Information Set(Reyes-Ortiz et al. 2016) distributed by the College of California, Irvine.

When downloaded from the hyperlink above, the information accommodates two completely different ‘components.’ One which has been pre-processed utilizing varied function extraction methods corresponding to fast-fourier remodel, 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 information set we are going to use.

The motivation for working with the uncooked knowledge on this put up is to assist 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 differ tremendously from job to job; requiring a lot of guide effort and area information. One of many stunning issues about deep studying is the function extraction is discovered from the information, not exterior information.

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"))

Subsequent, we load within the labels key for the RawData. This file is an inventory of all the observations, or particular person exercise recordings, contained within the knowledge set. The important thing for the columns is taken from the information 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(
  col.names = c("experiment", "userId", "exercise", "startPos", "endPos")

labels %>% 
  head(50) %>% 

File names

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

dataFiles <- checklist.information("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 accommodates: 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) %>% 
    experiment = str_remove(experiment, "exp"),
    userId = str_remove_all(userId, "person|.txt")
  ) %>% 
  unfold(kind, filePath)

fileInfo %>% head() %>% kable()

Studying and gathering knowledge

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

To acquire this we are going to scan by means of every of the recording information current in dataFiles, search for what observations are contained within the recording, extract these recordings and return every thing to a straightforward 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")
    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 courses.

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

  extractObservation <- perform(startPos, endPos){
  # get commentary 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 %>% 
      knowledge = map2(startPos, endPos, extractObservation)
    ) %>% 
    choose(-startPos, -endPos)

# scan by means 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 information

Now that we’ve all the information loaded together with the experiment, userId, and exercise labels, we are able to discover the information set.

Size of recordings

Let’s first have 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 practice the mannequin on each class directly 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 are going to match our mannequin to only the biggest ‘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 could be trying to make use of one other structure corresponding to an RNN that may deal with variable size inputs and coaching it on all the information. Nevertheless, you’ll run the danger of the mannequin studying merely that if the commentary is lengthy it’s most definitely one of many 4 longest courses which might not generalize to a situation the place you had been operating this mannequin on a real-time-stream of knowledge.

Filtering actions

Based mostly on our work from above, let’s subset the information to only be of the actions of curiosity.

desiredActivities <- c(

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

filteredObservations %>% paged_table()

So after our aggressive pruning of the information we may have a decent quantity of knowledge left upon which our mannequin can study.

Coaching/testing break up

Earlier than we go any additional into exploring the information for our mannequin, in an try and be as honest as potential with our efficiency measures, we have to break up the information right into a practice and check set. Since every person carried out all actions simply as soon as (excluding one who solely did 10 of the 12 actions) by splitting on userId we are going to be sure that our mannequin sees new individuals solely once we check 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’ve trimmed our knowledge by eradicating actions and splitting off a check set, we are able to truly visualize the information for every class to see if there’s any instantly discernible form that our mannequin could possibly decide 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) %>% 
    dataRow <- trainData[rowNum, ]
    dataRow$knowledge[[1]] %>% 
        activityName = dataRow$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’ve 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 no less than within the accelerometer knowledge patterns positively emerge. One would think about that the mannequin might have hassle 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.


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

Padding observations

First we are going to resolve what size to pad (and truncate) our sequences to by discovering what the 98th percentile size is. By not utilizing the very longest commentary size it will assist us keep away from extra-long outlier recordings messing up the padding.

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


Now we merely must convert our checklist of observations to matrices, then use the tremendous helpful 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()

[1] 286 334   6

Great, 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 very last thing we have to do earlier than we are able to practice our mannequin, and that’s flip our commentary courses from integers into one-hot, or dummy encoded, vectors. Fortunately, once more Keras has provided us with a really useful perform to just do this.

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

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



Since we’ve temporally dense time-series knowledge we are going to make use of 1D convolutional layers. With temporally-dense knowledge, an RNN has to study very lengthy dependencies to be able to decide up on patterns, CNNs can merely stack a number of convolutional layers to construct sample representations of considerable size. Since we’re additionally merely searching for a single classification of exercise for every commentary, we are able to simply use pooling to ‘summarize’ the CNNs view of the information right into a dense layer.

Along with stacking two layer_conv_1d() layers, we are going to 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 study
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 %>% 
    filters = filters,
    kernel_size = kernel_size, 
    input_shape = input_shape,
    padding = "legitimate", 
    activation = "relu"
  ) %>%
  layer_batch_normalization() %>%
  layer_spatial_dropout_1d(0.15) %>% 
    filters = filters/2,
    kernel_size = kernel_size,
    activation = "relu",
  ) %>%
  # Apply common pooling:
  layer_global_average_pooling_1d() %>% 
  layer_batch_normalization() %>%
  layer_dropout(0.2) %>% 
    activation = "relu"
  ) %>% 
  layer_batch_normalization() %>%
  layer_dropout(0.25) %>% 
    activation = "softmax",
    identify = "dense_output"


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        
Complete params: 4,746
Trainable params: 4,578
Non-trainable params: 168


Now we are able to practice the mannequin utilizing our check and coaching knowledge. Notice that we use callback_model_checkpoint() to make sure that we save solely one of the 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 bettering).

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

trainHistory <- mannequin %>%
    x = trainObs, y = trainY,
    epochs = 350,
    validation_data = checklist(testObs, testY),
    callbacks = checklist(
                                save_best_only = TRUE)

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


Now that we’ve a educated mannequin let’s examine the errors that it made on our testing knowledge. We are able to load one of the best mannequin from coaching based mostly upon validation accuracy after which have a look at every commentary, 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)) %>% 

# 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) %>% 
    highestProb = max(prob),
    predicted = label[prob == highestProb]
  ) %>% 
    reality = testData$activityName,
    appropriate = reality == predicted

predictionPerformance %>% paged_table()

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

predictionPerformance %>% 
  mutate(end result = ifelse(appropriate, '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 simply 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(reality, predicted) %>% 
  summarise(depend = n()) %>% 
  mutate(good = reality == predicted) %>% 
  ggplot(aes(x = reality,  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) +

We see that, because the preliminary visualization instructed, the mannequin had a little bit of hassle with distinguishing between LIE_TO_SIT and LIE_TO_STAND courses, together with the SIT_TO_LIE and STAND_TO_LIE, which even have related visible profiles.

Future instructions

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