Power operations involve strong electricity and high voltage, and violations by power operators can cause serious accidents. Most of the existing intelligent recognition methods for power operation violations are based on human posture estimation methods, and human skeletal key point detection is the predecessor task to complete posture estimation. Considering that electric power operators often perform electric power operations in low-light environments such as cloudy days and dark nights, this paper proposes a human skeletal keypoint detection framework considering low-light conditions, including two parts: Improved Cycle Generative Adversarial Network (ICGAN) module and the Minimal-Non-Global Dependency Integral Regression Network (Minimal-NGDIRNet). Firstly, we can get rid of the dependence on paired datasets to obtain normal brightness images by improving the generator, discriminator part of traditional recurrent generative adversarial network and introducing an identity consistency loss function. Then, based on Minimal-NGDIRNet, it is possible to significantly reduce network parameters and computational complexity while ensuring network skeletal keypoint detection performance through the Progressive Weighted Clustering Attention (PWCA) Block and Coordinate Attention Bottleneck (CABNeck) block proposed in this paper. The experimental results show that Minimal-NGDIRNet reduces the number of parameters and computation on the COCO test set by 73.12% and 62.05%, respectively, compared to the more advanced High-Resolution Net (HRNet), and the mAP reaches 73.31% without the need for additional heatmap decoding. The above experimental results show that the proposed detection framework in this paper has low deployment consumption and can effectively extract human bone keypoints.