Akademik Veri Yönetim Sistemi
5th International Artificial Intelligence and Data Science Congress (ICADA), İzmir, Türkiye, 24 - 25 Nisan 2025, sa.50, ss.147-157, (Tam Metin Bildiri)
Human activity recognition (HAR) using millimeter-wave (mmWave) radar data has gained significant attention due to its robustness under challenging environmental condi- tions. Building on previous work that employed the VIDAR (Video of Radar) approach— transforming raw radar signals into multidimensional grayscale images for classification via CNN and SVM models—this study proposes a background subtraction-based en- hancement technique to further improve recognition accuracy. Static environmental data were collected across multiple sessions, averaged pixel-wise to create a generalized background model, and subtracted from each action frame to generate motion-salient radar representations. Experiments conducted under identical model configurations reveal an approximate 0.5% increase in classification accuracy compared to the original VIDAR baseline. These findings indicate that environmental noise removal can meaningfully en- hance motion features critical for HAR tasks. Future work will investigate the generalizabil- ity of the proposed method across diverse settings and neural architectures.