Detection of Climbing Rod Motion Changes Based on 3D Pose Estimation

Abstract

3D human pose estimation has important engineering application value in industrial scenarios such as safety monitoring of power grid pole-climbing operations and standardized action training. Existing skeleton-based human behavior recognition methods still have notable limitations in modeling long-term temporal dependencies, capturing multi-scale dynamic motion patterns, and resolving geometric ambiguities in 2D-to-3D pose lifting. This paper proposes a graph-based network integrating multi-scale temporal convolution with spatial attention, and an interactive spatio-temporal Transformer, and develops an end-to-end web-based recognition system. Experiments on public benchmark datasets fully verify the performance of the proposed methods, which can provide technical support for skeleton-based action recognition and deployment of intelligent perception systems in industrial scenarios.