Recognizing human activities using light-weight and effective machine learning methodologies

Introduction

The task of recognizing the present physical action done by one or more users from a series of observations is known as human activity recognition. These observations are made while the user is doing any action in the defined environment. The rise of omnipresent, wearable, and persuasive computing has spawned a variety of new uses. Human activity recognition has evolved into a significant technology that is altering people's everyday routines. The notion of human activity recognition has recently gained popularity, resulting in a variety of applications such as assistive technology, health and fitness tracking, elder care, and automated surveillance. Furthermore, activity recognition research has progressed at such a quick pace that it is already being used for purposes other than activity recognition like monitoring elderly people or prison inmates.

Mobile phone sensors play an important part in making smart phones more useful and aware of their surroundings such as space, the presence of lamp, rubbish, among others. As a result, most smart phones have a variety of integrated sensors. It is feasible to get a large quantity of information on a person's everyday life and activities using this method. Among these sensor devices are accelerometer and gyroscope sensors. We can capture the motion of an item with the right usage of these sensors. As a result, machine learning algorithms may be used to anticipate human action based on sensor data. The information is constantly monitored. Human activity recognition is capable of identifying a variety of actions, both basic and complicated.

Methods

The methodology for human activity recognition comprises four phases: input, data cleaning, data splitting, and classification and validation. Human activity recognition structure is shown in Figure 1.

Figure 1. Human activity recognition structure.

Input: The dataset downloaded from the UCI dataset repository was the input of our study. The dataset contained information about triaxial acceleration from an accelerometer and triaxial angular velocity from a gyroscope. The dataset contains totally 10,299 data entries and 561 features.

Data cleaning: Data cleaning means removing the unnecessary columns and removing or replacing the inaccurate records from the dataset. In data cleaning, an encoding technique is applied to convert categorical data into numerical data. In our dataset, all feature values were numerical, hence there was no need to apply an encoding technique. Data cleaning prevents an error from happening again.

The first step in data cleaning is removing unwanted observations from the dataset. These unwanted data include irrelevant and duplicate data. Duplicate data most frequently arises when we combine data from multiple resources; irrelevant data refers to data that do not fit to the problem.

The second step is to handle the missing data. This can be done in two ways:

Removing the record and filling the missing value manually were not fully optimal because we were losing some information; for example, range might be replaced with another value. Hence, for missing numerical data we had to take the following two steps:

By using the technique of flagging and filling, the algorithm can estimate the missing value optimally.

Data splitting: In this phase the data is segmented into two components

Training data: The data which was used to train the model. We took 60% of the data as the training data.

Testing data: The data which was used to test our model is called testing data. We have taken 20% of the data from the dataset as testing data.

Classification and validation: The classification and validation was further comprised of two phases

Classification: Classification is the process predicting the target class of new observation based on the training data. The dataset we selected contains the target class hence we implemented the model using classification algorithms. We used the following algorithms

Decision tree: This classification uses tree representation to solve a problem. Each internal node represents the features, branches represent the outcome of the feature and leaf node represent target class labels.

Attribute selection measures: We had to select the best attribute for node at each level. This can be done by two measures

Information gain is mathematically defined as Equation 1

Hence from entropy, information gain is mathematically represented as

Gini index: It measures how often a randomly chosen element would be incorrectly classified.

As our dataset were unbalanced, meaning there were multiple classes, a decision tree made use of information gain to select the best attribute.

Logistic regression: Logistic regression is a statistical method used for classification which falls under supervised learning. It is used for predictions based on probability concepts. It analyses data set and take discrete independent input values and give a single output.

Logistic function: Logistic function is also known as the sigmoid function, which takes a real-value number as input and maps it into a value between 0 and 1, but not exactly at those limits. It maps predictions to probabilities.

By using the sigmoid function, if we give an input value on x, then it predicts the target value on the y axis.

Here k = 1, $x_0=0,L=1$

Naïve Bayes classifier: Naïve is a probabilistic classification algorithm which is based on Bayes theorem. It assumes that the occurrence of an instance is independent of other instances.

The conditional probability of an object with feature vector $x_1,x_{2,}$… …., $x_n$ belongs to a particular class $C_i$, and it is calculated with Equation 7.

Gaussian naïve Bayes: As values of each feature of our data set are continuous, we use the Gaussian distribution, which is also called the normal distribution.

Where $\mu$ is the mean of all the values a feature and $\sigma$is the standard deviation.

K-nearest neighbours is one of the essential classification algorithms. It falls under the category of supervised learning. It assumes that similar data are close to each other. Here the data is classified into groups based on an attribute. It is used in applications like pattern recognition, intrusion detection and data mining.

Random forest: Random forests are ensembles, which means combination of two or more models to get better results. Random forests create a forest of decision trees on data samples. Each decision tree gives a target class as output and the final target class is identified by performing some measures on outputs from each decision tree of corresponding input record.

Artificial neural network (ANN): ANN are made up of nodes that are connected to one other. The values of the property are stored in these nodes. The input values for ANN are collected from the training example's characteristics. The weighted values are subsequently delivered to the next set of nodes. This weighted total is subjected to a non-linear activation function, and the resulting value is transmitted to the next layer, where the process is repeated until the final output is reached. Figure 2 depicts this process.

Figure 2. Neural network model structure.

By adjusting the weights in ANN, the anticipated output value gets closer to the observed value. The back-propagation algorithm is the most often used method for altering weights. The complete training set is applied repeatedly due to the nature of ANN learning, where each application is unique.

Validation: Evaluation metrics were used to verify how well the model fits to our data. There are many metrics to evaluate the model; some of the measures are accuracy, precision, recall, F1 score.

Precision: Precision is also known as positive predicted value which is a measure of accuracy. It is mathematically defined as true positives divided by the sum of true positives and false positives.

Recall: Recall is also referred toas sensitivity, which is a measure of accuracy. It is mathematically defined as true positives divided by sum of true positives and false negatives.

F1 score: The F1 score is calculated with equation 11

Accuracy: Accuracy defines how well the model predicts the class. It is mathematically defined as the number of truly predicted classes divided by total number of classes.

Results and discussion

The Results section comprises two parts:

Exploratory data analysis means gaining insights from the data before applying any model to it. It can be done using mathematical functions and visualizations. The total number of rows and columns in a dataset are identified by using the ‘shape’ function. It is shown in Figure 3.

Figure 3. Shape of the dataset.

From this we know that the dataset contains 10299 instances and 562 attributes.

Out of 562 attributes 561 are independent attributes and one is dependent variable.

We used graphs to explore the data. We used a bar graph, box plot and heatmap.

A bar graph shows the comparison between different target classes. The bar graph showing how many persons had thesame classification is shownin Figure 4. The following observations were made from Figure 4: LAYING--1944, STANDING--1906, SITTING--1777, WALKING--1722, WALKING_UPSTAIRS--1544, WALKING_DOWNSTAIRS--1406.

Figure 4. Bar graph for the attribute ‘Activity’.

Box plot: Box plots show the distribution of data by using five measures, namely minimum, first quartile(Q1), median, third quartile(Q3), and maximum. The box plot is drawn for the target class ‘Activity’ and the feature ‘tBodyAcc-max()-X’ in Figure 5.

Figure 5. Box plot.

From Figure 5, we made the following observations:

Confusion matrix for naïve Bayes classifier is shown in Figure 6. From Figure 6, the following observations were made:

Figure 6. Confusion matrix for naïve Bayes classification.

Precision, recall and F1 score values of naïve Bayes classifier are depicted in the bar plot shown in Figure 7. From Figure 7, we can conclude that the class Laying was most correctly predicted among all other classes. Among all the classes, Standing had a low number of correctly predicted instances.

Figure 7. Precision, recall and F1-Score for naïve Bayesclassifier.

Confusion matrix for the decision tree classifier is shown in Figure 8. From Figure 8, the following observations were made:

Figure 8. Confusion matrix for decision tree.

Precision, recall and F1 score values of decision tree classifier are depicted in the bar plot shown in Figure 9. From Figure 9 we can see that the class Laying was predicted correctly without any incorrectly classified classes.

Figure 9. Precision, recall and F1 score for decision tree classifier.

Confusion matrix for K-nearest neighbours classifier is shown in Figure 10. From Figure 10, the following observations were made:

Figure 10. Confusion matrix for K-nearest neighbour.

Precision, recall and F1 score values of the K-nearest neighbour classifier are depicted in the bar plot shown in Figure 11. From Figure 11, we can see that the activity Laying was predicted correctly without any incorrectly classified classes; hence, all metrics had a 100% score.

Figure 11. Precision, recall and F1 score for K-nearest neighbour classifier.

Confusion matrix for the random forest classifier is shown in Figure 12. From Figure 12, the following observations were made:

Figure 12. Confusion matrix for random forest.

Precision, recall and F1 score values of random forest are depicted as a bar plot shown in Figure 13.From Figure 13 we can see that all instances having Laying as target were correctly predicted.

Figure 13. Precision, recall and F1 score for random forest classifier.

Confusion matrix for logistic regression is shown in Figure 14. From Figure 14, the following observations were made:

Figure 14. Confusion matrix for logistic regression.

Precision, recall and F1 score values of logistic regression are depicted as a bar plot shown in Figure 15. From Figure 15, we can see that Laying, Walking, Walking_Downstairs instances were correctly classified.

Figure 15. Precision, recall and F1 score for logistic regression classifier.

Neural network: Epoch versus training loss for the human activity dataset is represented in Figure 16. From Figure 16, the testing loss decreased and hence the accuracy increased. Of all the algorithms we applied, neural networks predicted more accurately, with an accuracy of 98.93%

Figure 16. Loss graph for neural network.

Comparison of different algorithms: Different algorithms used to train the model were considered as variable x and their corresponding accuracy as variable y. The bar graph drawn for these x and y variables is shown in Figure 17. The accuracy results are shown in Table 2. From Figure 17, when comparing all algorithms, neural network predicted the data with the highest accuracy, i.e.98.93%.

Figure 17. Comparison of algorithms.

Conclusions

For this work, human activity recognition was used is to detect the activity of a person. The data set was collected by sensors from triaxial accelerometer and gyroscope. The data set consisted of 561 features and 10,299 records, and six classes, namely Sitting, Standing, Laying, Walking, Walking_Downstairs, Walking_Upstairs. In this paper we used six algorithms, namely naïve Bayes, decision tree, random forest, K-nearest neighbours and neural network. For neural networks, the input layer consisted of 561 nodes and three hidden layers (1024,512,64);output layers consisted of six nodes, each representing a target class. From the experimental results, naïve Bayes classifier achieved 76.89% accuracy, the decision tree classifier achieved 93.39%, random forest classifier achieved 96.89%; K-nearest neighbours achieved 96.40%;logistic regression classifier achieved 98.05%. Among all these models,the neural networks model predicted the target class with an accuracy of 98.93%.