Real-Time Human Pose Estimation on Embedded Devices Based on Deep Learning
HE Caishi, WU Sijia, Lu Yan, Zihao Dou, Yang Feng · 2023
Human posture detection plays an important role in areas such as human behavior analysis and human-computer interaction. Deep learning based methods are one of the mainstream approaches. In practice, detection models are often deployed in resource-constrained edge devices for privacy protection, increased realtime, etc. However, most of the pose detection is for dynamic scenes, which requires high detection speed and accuracy, and thus models are mostly complex in structure and decentralized computation. Aiming at the above difficulties, this paper proposes a multi-person pose detection method, implements a lightweight detection network-RepPose, and deploys it on an embedded device. Specifically, this paper proposes a branch pruning scheme based on the reparameterized structure; and designs a high-performance post-processing operator for the post-processing part; in addition, we also propose a multi-threaded parallel accelerated inference scheme. Finally, it is deployed on Huawei Atlas200DK, and the result reaches 26FPS in the multi-target detection experiment, and achieves more than 90% classification accuracy in pedestrian state detection experiment in simple traffic scenarios.