rmakkar2018 / detectasana

Classification of Body postures using different ML algorithms and comparing their performances.

DetectAsana

Code folder

The Code folder contains the DetectAsana_ML_Project.ipynb file with the implementation of various Machine Learning Algotirhm's on PUC-RIO dataset for training and pre-processing of data.

In the Feature Extraction Section: Apart from the existing features in PUC-RIO dataset, we have manually extracted other features.We calculated norm=sqrt(x^2 + y^2 + z^2), pitch=tan_inverse(y/(x^2+z^2)) and roll=tan_inverse(-x/z) and added those to the features.

In the Resampling and Feature Selection: Class weights: We have also used Class weights technique to handle dataimbalance which punishing classes with more samples lesser for each misclassification. Oversampling: We have used the ADASYN(Adaptive Synthetic) sampling to deal with data imbalance. Undersampling: We have also used undersampling techniques like monte carlo to deal with the data imbalance.

In the "Classwise Comparison of raw data and oversampled data" section: We compare the original data wit the data we get after oversampling.

#Feature analysis After that we have perform the feature analysis by plotting all four accelerometer's reading along the three axis separately: 1-On waist, 2-On Right ankle, 3-On right upper arm, 4-On left thigh

#Raw featue vs Extracted feature analysis: We have compared the performance using the above two feature set and plotted the confusion matrix for the same with Logistic Regression and Decision Trees.

Using various ML algoritms with different preprocessing techniques and feature set:

#1-Using all the features(Extracted and Raw features) for Ml Models: Logistic Regression, SGD Classifier, Decision Tree Classifier, Random Forest Classifier, Support Vector Classification and Guassian Naive Bayes and generating Evaluation matrix for all cases.

#2-Using class weights technique to handle data imbalance and then using the data for Ml Models: Logistic Regression, SGD Classifier, Decision Tree Classifier, Random Forest Classifier, Support Vector Classification and Guassian Naive Bayes and generating Evaluation matrix for all cases.

#3-Using Raw features for Ml Models: Logistic Regression, SGD Classifier, Decision Tree Classifier, Random Forest Classifier, Support Vector Classification and Guassian Naive Bayes and generating Evaluation matrix for all cases.

#4-Using Undersampling technique to handle data imbalance and then using the data for Ml Models: Logistic Regression, SGD Classifier, Decision Tree Classifier, Random Forest Classifier, Support Vector Classification and Guassian Naive Bayes and generating Evaluation matrix for all cases.

#5-Using Monte Carlo Undersampling technique to handle data imbalance and then using the data for Ml Models: Logistic Regression, SGD Classifier, Decision Tree Classifier, Random Forest Classifier, Support Vector Classification and Guassian Naive Bayes and generating Evaluation matrix for all cases.

Final pipeline

The Code folder also contains a DetectAsana_final.ipynb file which is a complete main pipeline for the classification task after analysing their affects on the performance. The Feature set that we have used are 'weight', 'body_mass_index', 'pitch1', 'y3', 'roll4', 'y4', 'pitch2', 'z1', 'z4', 'z2', 'norm1', 'x4', 'y2', 'y1', 'x3', 'z3', 'x2', 'roll1', 'norm4', 'norm2', 'norm3', 'x1'. The Resampling technique that we have used is Bagging/Monte Carlo to deal with the data imbalance. Finally we have used the best accuracy giving ML model i.e. Random Forest for the classification task. The final accuracy that we get using the main pipeline is : 99.5%.

Other files

#Dataset folder The dataset folder contains the PUC-RIO dataset in .csv format along with a Readme file which contains the dataset's description.

#DetectAsana Proposal.pdf is the project proposal submitted for the Machine Learning Project on Sept 10.

#Mid-Progress Report.pdf is the midsem-progress report of Machine Learning Project submitted on Nov 22.

#Pipeline.pdf is the final pipeline suggested for the classification task

#DetectAsana-Poster.pdf is the poster on the literature review performed

Weights folder

It consists of the final model trained using the suggested pipeline