Abstract: The recognition of dangerous behaviors at power operation site requires high spatial coordinates of the image. Due to the influence of lighting environment, the point cloud histogram is easily mixed with a lot of bad information, resulting in offset or feature information overlap in the spatial key point coordinates of the target pose estimation. It is impossible to match the feature quantities with high correlation degree with the risk behavior, which affects the accuracy of the output result of dangerous behavior recognition. Therefore, an identification method of hazardous behaviors in power operation sites based on laser radar and convolutional neural network(CNN) is proposed. The proposed method collects point cloud data from power operation sites by using laser radar equipment and processes by using statistical filtering. Based on this, a CNN model for identifying dangerous behaviors is constructed, which converts the feature map into a two-dimensional projection image form, and matches it with dangerous behaviors to determine that the feature map belongs to the category of dangerous behaviors, achieving effective recognition of dangerous behaviors. The test results show that the dangerous behavior recognition results obtained by the proposed method are consistent with the tested dangerous behavior level. The maximum sensitivity of dangerous behavior recognition is 0.98, and the maximum specificity of dangerous behavior recognition is 0.94, which fully confirms that the design method has excellent dangerous behavior recognition performance.