Improved OpenPose for multi-person pose estimation and behavior recognition

In response to the sluggishness of multi-person pose estimation algorithms, rendering real-time pose estimation unattainable and restricting the output to human keypoints alone, practical behavior recognition applications remain unfeasible. Propose a lightweight attitude estimation method that introduces a deep separable convolutional structure and improves its pre-feature extraction module to improve target detection speed. Simultaneously, an adopted dual-stream adaptive graph convolutional network, Two-Stream Adaptive Graph Convolutional Networks (2s-AGCN) , leverages the key point information extracted by OpenPose for skeleton-based action recognition. On the MSCOCO dataset, in comparison to prevalent algorithms like OpenPose, AlphaPose, and HRNet, this algorithm attains an 85.2% accuracy rate in human key point recognition. While upholding detection precision, it reduces the computational complexity of the base model by 33.7×109 and trims down the parameter count by 83.1%. On the NTU-RGB+D dataset, the algorithm achieves a 95.3% accuracy rate on the X-View benchmark set, demonstrating a slight advantage over its counterparts. And it can effectively detect behavior recognition for human pose estimation in real-time.