The opening windows of architectural curtain wall are easily blown off when strong winds come, which seriously endanger personal and property safety. In order to discover the opening windows promptly and accurately, this paper proposes an opening windows detection approach based on computer vision. The panorama image of curtain wall is captured by wide angle camera which locate in bottom of the building, use U-Net semantic segmentation model to obtain facade mask image, and extract reference points from the mask image which are used in perspective transform, the perspective transform is used to correct the facade distortion which caused by elevation angle shooting, the corrected open windows position distribution are regularly which can be directly cropped by the preset area,then use the ResNet-18 classification model to judge the opening and closing status. The proposed method is verified on a high-rise building, the results demonstrate that this approach has a high level of effectiveness with various brightness images, and the detection time is less than 2 seconds. Therefore, this approach meets application requirements in accuracy and timeliness.