Hardware/Software-Codesign for Hand Gestures Recognition using a Convolutional Neural Network
Sarah Said, Lester Kalms, Diana Göhringer, Mohamed A. Abd El Ghany · 2019
Hand Gesture (HG) recognition has been an interesting and challenging scientific problem in computer vision and machine learning. However, achieving high performance and high accuracy is a challenging task in gesture recognition. The intensive computation, the complex backgrounds, conditions of lights and positions are the most remarkable challenges regarding performance and accuracy. Those two aspects are taken into consideration in this paper. It proposes a convolutional neural network (CNN) model that can recognize three static HGs using a simpler architecture and fewer weights than previously proposed networks. The proposed CNN is completely implemented and several optimizations are performed to reduce runtime for frames with 28x28 pixels. The total runtime is 42 ms. Moreover, one of the layers is implemented on hardware(HW) level on ARM-FPGA SoC and it is faster than the software(SW) by 39.8x speedup. The total run time is 23.75 ms in C++/ VHDL. A comparison of the proposed model is made between our accelerated implementation and Caffe, which achieves a speedup of 1.17. The proposed model has an accuracy of 94%.