Mänsklig rörelseklassificering: En jämförande studie : Utvärdering av maskininlärningsmodeller och funktions-extrahering för att optimera klassificering av kroppsrörelser i förstärkt verklighet

Human movement is one of the most complex elements in technical applications, as it is characterized by great variation, subtle details, and different patterns that require careful analysis. This thesis presents a system for classifying human movements based on the established training program “7-minute workout”. The work includes several stages, from data collection, cleaning, and annotation to structuring the information. A central part of the process has been the extraction of geometric features in the form of angles between different joints and body parts. Two machine learning approaches were evaluated. One was a tabular classification model built using statistical features derived from feature extraction. This model achieved a weighted average accuracy of 99.2% and used a highly compact data representation. It was approximately 250 times smaller than the skeleton data, which resulted in faster training and only 82,000 trainable parameters. The other approach used an adapted pre-trained model for skeletal movement analysis, which achieved 94.9% accuracy but required 1.45 million parameters and significantly more computational resources. The results highlight the effectiveness of lightweight, interpretable models built on well-engineered features, particularly when efficiency is a priority. The system provides a foundation for future applications in real-time feedback or augmented reality environments.