Comparison of hardware-optimized CNN and SVM models for human activity recognition using the HARTH and HAR 70 + datasets
Kei Palabasan, Ramona Rajagopalan, Jean Marriz Manzano, Marc Driz Rosales, Maria Theresa de Leon, John Richard E. Hizon · 2023
Human activity recognition (HAR) has been of interest for decades as an application for machine learning (ML) in low-power and profile embedded applications such as in FPGAs. This study compares the model performance and FPGA resource utilization estimates after HLS synthesis of representative classical (SVM) and deep (CNN) models using the publicly available HARTH and HAR 70 + datasets along with common optimizations. In resource-constrained environments, CNNs perform better overall than SVMs, with axis reduction (using only the x-axis of the sensors) and data format reduction (using int16) providing substantial improvements in resource utilization without significant performance penalties. For SVMs, loop perforation also provides an improvement although these models necessitate external processing for feature extraction.