FPGA-enabled hand gesture recognition for smart home applications

Machine learning (ML) algorithms trained with hand gesture data have achieved accuracy rates close to 95%. However, despite this high accuracy, a trained ML algorithm requires hardware implementation to be effectively utilized in practical applications. In the proposed approach, we select one of the three investigated ML algorithms (CNN, ResNet50, and MobileNetV2) that demonstrates the best accuracy and make it compatible with loading onto a Field Programmable Gate Array (FPGA). This allows the algorithm to be effectively utilized in real-world applications. The optimized algorithm was implemented on a Zynq 7000 board to enhance performance, achieving greater efficiency, accuracy, and lower power consumption compared to traditional CPU-based or other architectures. The implementation was tested using a custom dataset containing six distinct hand gestures for detection and computation. Resource-efficient utilization led to a 2% increase in accuracy and a 5% reduction in processing time for the CNN-trained model on the FPGA compared to the CPU. Similarly, the ResNet50-trained model on the FPGA achieved 7% higher accuracy and 9% faster processing time compared to its execution on a CPU. Hand gesture recognition integrated with FPGAs can achieve real-time responsiveness, energy efficiency, and reliability, making them ideal for resource-constrained environments that demand high performance. This demonstrates the FPGA's advantage in smart home solutions, which combine accuracy, speed, and low power consumption.