Classifier Module of Types of Movements Based on Signal Processing and Deep Learning Techniques

Human Activity Recognition (HAR) has been widely addressed by deep learning techniques. However, most prior research applied a general unique approach (signal processing and learning) to deal with different human activities including postures gestures. These types of activity typically have highly diverse motion characteristics, which could be captured wearable sensors placed on the user’s body. Repetitive movements such as running or cycling repetitive patterns over time generate harmonics in frequency domain, while sitting lying are characterized fixed position, some positional changes gestures non-repetitive being based an isolated movement usually performed limb. This work proposes classifier module perform initial classification among these movements, would allow for applying afterwards appropriate terms signal techniques each type movement. evaluated using PAMAP2 OPPORTUNITY datasets subject-wise cross-validation methodology. contain recordings from inertial hands, arms, chest, hip, ankles, collected non-intrusive way. In case PAMAP2, baseline classifying 12 5-s windows domain obtained accuracy 85.26 ± 0.25%. distinguish between reaching higher performances. Afterward, specific window size, format, architecture were used module, obtaining final 90.09 0.35% (an absolute improvement 4.83%).