Deep neural network for human falling prediction using log data from smart watch and smart phone sensors

The purpose of this research was to conduct a prediction human falling using deep learning and dimensionality reduction techniques in human activity recognition and behavioral prediction using smart watch and smart phone data. The deep learning-based techniques combined with multiple sensor data aim to classify daily activities. Previous work in human falling has focused on using multiple accelerometers placed on different parts of the body, with more recent work focused on sensors embedded in smartphones to classify activities. This research classifies activities from utilizing the data from the following sensors-accelerometer, gyroscope and magnetometer. In addition to comparing these evaluation metrics, a comparison of each network's confusion matrix, feature importance and multisensory fusion analysis is performed-to evaluate which network best suits the data and successfully classifies the daily activities in question. Another intriguing aim of this research is to compare two data clustering techniques for visualizing the smart watch and smart phone dataset. This research aims to present the best visualization technique by conducting a comparative study on the two-visualization techniques. The result of this research found that all six-machine learning classification algorithms consistently outperformed State-of-the-Art baselines. Deep Neural Network (99.97% accuracy) and MLP (90.55%) accuracy performed excellently on the data, with very little misclassified instances. All six-classification algorithms produced more insightful, predictive results than existing baselines, while DNN successfully clustered and visualized the data. The results show that each algorithm is suited to the smart watch and smart phone dataset, with high performance results achieved throughout. The DNN model does not struggles in distinguishing between the falling activity and running activity with 7% of the activity misclassified. DNN outperforms MLP in this aspect as it misclassifies 3% of the activities between jogging and running. A solution to this would be to place an extra sensor on the thigh to distinguish between both activities. This sensor would lead to detection in a greater acceleration and range of motion in the upper thigh area when the subject is running in comparison to falling.