Development of a Machine Learning Based Fall Detection System for the Elderly and Disabled
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Fall accidents from accidental injury are considered one of the significant global public health concerns, and the largest proportion of fatal accidents are experienced by elderly people. Currently, there is a demand for creating an effective machine learning-based fall detection system that is portable at a low cost.
Hence, this project is aimed at undertaking a research study on the various fall detection and navigation systems designed for the elderly and disabled and developing an effective machine learning-based low-cost fall detection system. The methodology included a study on the various fall detection systems as well as general features used in machine learning for fall and ADL (Activities of Daily Living) classifications. The most suitable potential combination of machine learning algorithms that will provide the best accuracy, precision, sensitivity, specificity and lowest training time were developed via simulation models using MATLAB. The SisFall and MobiFall datasets were used for both classification and testing for input data in simulations. SisFall dataset used ADXL345 and MMA8451Q accelerometer model, while the MobiFall dataset used LSM330DLC inertial module from a Samsung Galaxy S3 smartphone.
Up to 24 algorithms were simulated, including Decision Trees, Naïve Bayes Classifiers, Support Vector Machines, KNN and available Ensemble Classifiers. Four experiments were done, with Experiment 1 using the ADXL345 accelerometer with 5-fold cross-validation, Experiment 2 using 10-fold cross-validation, Experiment 3 using the MMA8451Q accelerometer with 10-fold cross-validation and Experiment 4 using the MobiFall Dataset with 5-fold cross-validation. The classifier models were run ten times, and each iteration was saved for further processing. Classifiers showing the most accuracy (up to 99%) in both training and testing phases included Quadratic SVM, Cubic SVM, Medium Gaussian SVM and Fine KNN.
A stacked ensemble method was simulated as well utilizing classifiers such as Medium Gaussian SVM, Cubic SVM and Fine KNN. SisFall and MobiFall Datasets were further used to train and test the classifier system. Initial testing and training had shown an improvement in accuracy (up to 99% in binary classifications) when compared to the system using individual classifiers. Results had shown a remarkable increase in precision, sensitivity and accuracy when classifying data in a binary classification system compared to classifying the data based on the specific categories (the various types of falls and ADLs). Hence an effective system model was developed and trained to identify the data between both Fall Events and ADLs as well as separately identifying the actual type of Fall/ADLs.
Please use this identifier to cite or link to this item: