Given accelerometer data from 3 activities namely: Standing, Running & Cycling, we need to classify/detect which activity is being performed from those corresponding data entries.

Some of the approaches, methodologies and ideas after surveying papers :-

1. Each time series A_i, with i = {x, y, z} can be filtered with a digital filter in order to separate low frequencies components and high frequencies component. The cut-off frequency has been set to 1 Hz, arbitrarily. In this way, we obtain for each time series, three more time series A_ij with j = {b, dc, ac}, where b, dc, ac represent respectively the time series without filtering, the time series resulting from a low pass filtering and the time series resulting from a high pass filtering.

2. A successful technique for extracting features from sequential motion data has been demonstrated to be windowing with overlapping or continuous non-overlapping windows.

3. We can extract the following features: root mean squared value of integration of acceleration in a window, mean value of Minmax sums, mean value, standard deviation, coefficient of variation, zero crossing rate, skewness, kurtosis , squared sum of data below 25 & 75 percentile, mean of mean of CWC, variance of CWC and correlation between each pairwise of accelerometer axis.

#NOTE : Papers that we have gone through while literature review for collecting ideas, approaches and methodologies are:-
1. A Study on Human Activity Recognition Using Accelerometer Data
2. Feature Selection and Activity Recognition
3. Physical Activity Recognition from Accelerometer

Click here to download the required paper in PDF format.

Step 1 : Accelerometer Data Collection and Data-Set Creation.
Step 2 : Feature Extraction.
Step 3 : Training the Classifier & testing and finding accuracy.

Section 1 describes the Motivation for our Project, Problem Statement and the Literature we reviewed.
Section 2 describes the detailed approach followed to extract feature out of the given dataset.
Section 3 shows our classification results and the data set used for verifying.
In section 4, a summary of our whole project work is presented.

Data is collected from ADXL335 which is a tri-axial accelerometer worn on the waist of the user. The data set is obtained by recording 9 sets of data which are performed by a number of individuals. Total duration of each set being 7.5 minutes corresponding to 2.5 minutes for each activity. These activities are performed in succession i.e. running, then cycling followed by standing. The total duration of data for each of the activity being [(2.5 minutes) * 9] = 22.5 minutes. The sampling rate used is 2 Hz and hence a total of [22.5 * (60 seconds) * (2 samples/second)] = 2700 samples for each activity.
Acceleration data corresponding to the three axes are therefore collected. This data is segregated into different categories i.e. Running, Cycling and Standing and the data set is created.

The following classifiers are used in our Project :-

1. Logistic Regression Model -

This is a classfication technique which fits a linear model to the feature space involving a probabilistic view of classification. This model involves a vector ß in d – dimensional feature space, the points in feature space are projected on to ß and converted into a real number in the range spanning the whole real line. This real number, thus obtained, is mapped to a value in the range 0 to 1 using the standard logistic function. The prediction using this model can be treated as a probability of class membership. By applying threshold to probability, class assignment can be done. Threshold represents the decision boundadry in the feature space. ß is optimized to give the best possible reproduction of training set labels. This is usually done by numerical approximation of maximum likelihood. On really large datasets, stochastic gradient descent may be used. There are a number of advantages to this :-

1. Makes no assumptions about distributions of classes in feature space.
2. Easily extended to multiple classes (multinomial regression)
3. Natural probabilistic view of class predictions.
4. Natural probabilistic view of class predictions
5. Quick to train
6. Very fast at classifying unknown records
7. Good accuracy for many simple data sets
8. Resistant to overfitting
9. Can interpret model coefficients as indicators of feature importance

When used on the data set at hand, it gives a training accuracy of 98.33% and a test accuracy of 97.41%.

2. Multinomial Regression -

Softmax regression (or multinomial logistic regression) is a generalization of logistic regression to the case where we want to handle multiple classes. In logistic regression the labels were binary, for this case the labels are more than two (three in our case i.e. Running, Cycling, Standing).In the softmax regression setting, we are interested in multi-class classification (as opposed to only binary classification), and so the label can take on K (the number of classes in question) different values, rather than only two. we want to estimate the probability of the class label taking on each of the K different possible values. Thus, our hypothesis will output a K-dimensional vector (whose elements sum to 1) giving us our K estimated probabilities.In the special case where K = 2, one can show that softmax regression reduces to logistic regression. This shows that softmax regression is a generalization of logistic regression.

When used on the data set at hand, it gives a training accuracy of 98.61% and a test accuracy of 98.05%.

3. Support Vector Machine -

Support Vector Machines fall under the category of supervised learning models that analyze data used for classification and regression analysis. Given a set of training examples, each marked as belonging to one or the other of two categories, an SVM training algorithm builds a model that assigns new examples to one category or the other, making it a non-probabilistic linear classifier (although methods such as Platt Scaling exist to use SVM in a probabilistic classification setting). An SVM model is a representation of the examples as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible. New examples are then mapped into that same space and predicted to belong to a category based on which side of the gap they fall. Other SVM classfiers used are Polynomial SVM and Gaussian (RBF) SVM, the difference being that the kernel used are polynomial and Gaussian respectively.

# Plots with different features mapped on 3 axes for Standing, Running and Cycling.

Standing corresponds to Red points.
Running corresponds to Green points.
Cycling corresponds to Black points.

In each of the case:-
1st Feature is mapped in X direction.
2st Feature is mapped in Y direction.
3rd Feature is mapped in Z direction.

Plot of Mean, Standard Deviation and Coefficient of Variation:

Plot of Crest Factor, Average and Squared Sum of Data below 25 Percentile:

Plot of Squared Sum of Data below 25 Percentile, Peak Frequency and Number of peaks in Spectrum of Y-axis data below 2 Hz:

Plot of Lag One Auto-Correlation, Skewness and Kurtosis:

Plot of Zero Crossing Rate, Continuous Wavelet Transform Mean, Continuous Wavelet Transform Variance:


Results from various classifiers:-

#Logistic Regression Train Accuracy :: 0.983333333333
#Logistic Regression Test Accuracy :: 0.974110032362

#Multinomial Logistic Regression Train Accuracy :: 0.986111111111
#Multinomial Logistic Regression Test Accuracy :: 0.980582524272

#Linear SVM Train Accuracy :: 0.988888888889
#Linear SVM Test Accuracy :: 0.980582524272

#RBF SVM Train Accuracy :: 0.991666666667
#RBF SVM Test Accuracy :: 0.961165048544

#Polynomial SVM Train Accuracy :: 0.994444444444
#Polynomial SVM Test Accuracy :: 0.987055016181

All the classifiers have performed reasonably well and are showing a high level of accuracy. Polynomial SVM performed the best on both the training and testing dataset.

The classifiers that we have used are Logistic Regression, Multinomial Regression and SVM trained on the features as previously mentioned i.e. Mean, Standard Deviation, Skewness, Kurtosis, Coefficient of Variation, Crest Factor, Zero Crossing Rate, 25 Percentile & 75 Percentile, Data below 25 and 75 percentile, Peak Frequency Component below 2 Hz, Number of peaks in y-axis data below 2Hz and Lag One Autocorrelation. The Multinomial Regression Classifier performed best while the other classifiers i.e. Logistic Regression and SVM Classifier performed reasonably well.

The methodology used classifies activities performed by differnet individuals at differnet speed and style with high accuracy. The system developed, thus, provides a foundation towards a more robust system that will require minimum training of the users and provide least errors due to orientation and positioning offsets of the accelerometer. With extension of the given system to include classification of more activities performed in real time it remains to test the system on more subjects. Also with the high accuracy achieved, it becomes worthwhile to deploy the system for real time use.

1. More efficient algorithms for the classification purpose are to be investigated obtaining better accuracy. Keeping in view that the classification if can be done in real time would be a great asset. Thus, appropriate feature vectors and algorithms have to be explored for segmenting the data and classifying these segments into different activities.

2. The sampling rate used in the present approach is quite low, this in itself causes complications in the extraction of features. Hence, for a future extension it becomes worth investigating the ideal sampling rate so that data obtained does not become redundant, all the while making it easier to select and extract features.

3. Since an individual performs many more activities than just the three considered it is worthwhile to investigate methods for classification among more number of activities.

1. A Study on Human Activity Recognition Using Accelerometer Data by Akram Bayat, Marc Pomplun, Duc A. Tran
2. Feature Selection and Activity Recognition by Piyush Gupta and Tim Dallas
3. Physical Activity Recognition from Accelerometer by Yonglei Zheng, Weng-KeenWong, Xinze Guan