Trampoline jump classification using sparse data in machine learning

Sammanfattning: Within the last few years, machine learning has grown rapidly and become widely used when classifying data from inertial measurement units (IMU). Sensor data, and in particular accelerometer data, are used in different types of trick classifications for various sporting exercises. As the field of machine learning is growing, more tendency is from the industry to incorporate AI into their products. The difficulties in working with machine learning are the importance of having enough data. This thesis will study the effectiveness in terms of accuracy of four different algorithms, Random Forest, Decision Tree, Logistic Regression and a Deep Neural Network to classify trampoline jumps from non-jump high accelerated events like a kick using accelerometry data. This experiment is done using a small dataset of 250 samples. The best result showed an accuracy of 96% in an artificial neural network with Synthetic Minority Oversampling Technique (SMOTE). Using SMOTE is shown to be a very successful approach to deal with the small dataset and get a more predictive model for all algorithms.