Hand gesture recognition (HGR) is one of the widely-used human-computer interaction technology. With HGR, the user can operate the interaction system without touching any devices. For a better experience, recognition accuracy and computational speed are especially important. In this work, a small-footprint HGR model and its hardware architecture design are proposed. The model first processes the hand segmentation and uses the feature to recognize the hand gesture. The model mainly consists of depthwise separable convolution to reduce the overall parameters and computations. We transfer the hand segmentation task with some features to the hand gesture recognition task as a single-stage model. Based on this hardware-efficient model, we propose the hardware architecture of the whole neural model including depthwise convolution, pointwise convolution, batch normalization, and max-pooling. We also demonstrate it on the evaluation board. The whole system is implemented on the Xilinx ZCU106 evaluation board. The implemented system can achieve the performance of 52.6 fps and 65.6 GOPS based on the evaluation board.
- attention model
- depthwise separable convolution
- field-programmable gate array (FPGA)
- Hand gesture recognition
- hardware accelerator