Multi-Person Pose Estimation on Embedded Device

Tasks based on deep learning have make great progress in accuracy with the development of deep learning in recent years. However, algorithms base on deep learning are both computationally intensive and memory intensive, makes it difficult to deploy on resource-constrained devices. Therefore, it is worth studying how to compress the architecture of existing network while achieving a comparable performance, so that the model can be deployed on devices with limited resources. This paper performs model compression and acceleration in multi-person pose estimation by replacing the feature extraction network, parameter pruning and knowledge distillation, achieve a $2.6\times$ MACs reduction and $2\times$ acceleration but only 25% drop in accuracy compared with a lightweight model. The compressed model can be deployed on the embedded device Jetson Nano with a 12 FPS inference speed.