A Lightweight Deep Gesture Recognition Model on Embedded Computing Platform
Xiaoya Cheng, Maojun Zhang, Wei Kai Xu · 2021
With the development of big data and highperformance hardware devices, the huge structure of neural network can be fitted efficiently and strongly, therefore, it is widely used as a common algorithm in the field of gesture recognition. However, in the application of low-cost embedded devices, the deployment of model is still relatively difficult. In this paper, in order to mitigate the mismatch between the powerful computing power required by the model and the embedded hardware computing resources, we use the channel pruning method to compact the model structure and reduce the parameters of model on the basis of 3D-SqueezeNet and 3D-MobileNetV2. Without losing accuracy, we reduce the parameters by 75%. We deploy the pruned gesture recognition model on NVIDIA® Jetson Nano™. In our experimental analysis, we achieve to recognize hand gestures containing three gesture-phonemes with an accuracy of 96% (in 100 classes). The speed of model inferring maintains at up to 33fps. Our code and pretrained models are publicly available1.