In order to ensure the safety of personnel and equipment during normal live working of distribution network, abnormal behavior recognition is an indispensable technical means. However, the existing live working behavior recognition methods have problems such as low accuracy, few identifiable types, and missed detection and false detection caused by background interference. An abnormal behavior recognition method is proposed for 10 kV live working video based on improved spatial temporal graph convolutional networks (ST-GCN). Firstly, the method of target detection and tracking is used to add a mask in the video human region to eliminate the influence of complex background. Then, the human skeleton is obtained by using the lightweight improved pose estimation model, and the spatio-temporal graph is constructed with multi-frame skeleton sequence. Finally, the spatial posture and timing information of the spatio-temporal map are extracted by ST-GCN, and the squeeze-and-excitation networks (SENet) module is introduced to strengthen the action features to complete the live working behavior recognition. The behavior dataset is constructed with some live working videos, and five typical behaviors such as insulation detection are selected for verification. The experimental results show that this method improves the speed of human pose estimation, constrains the skeleton detection area, and reduces the false detection rate and missed detection rate of limbs. It can effectively identify the live working behavior and reach 88% average accuracy rate. It has good generalization in complex environment and provides an effective reference for the intelligent safety monitoring of live working.