The tilt of the tower will cause serious damage to the entire power grid and threaten the lives of surrounding residents. The power inspection performed by the computer vision of the UAVs not only saves labour, but also significantly improves the inspection efficiency of the power grid. In order to get early warning before the tower falls for State Grid inspectors. In this paper, the algorithm of computer vision-based tower tilt detection in electric patrol unmanned aerial vehicles is researched. And the tilt of tower is detected using YOLOv3’s deep neural network combined with LSD line segment extraction method. Using the pole pictures of the actual inspection of the UAVs in Shanxi power grid to make the VOC2007 dataset of the pole tower and use the YOLOv3 neural network to detect the pole tower. The Bounding box obtained after the detection is fine-tuned according to the mIOU parameters after network training and used as LSD detection ROI, the detected line segment is filtered and fused, and the secondary identification of the tower is performed according to the characteristics of the tower. Finally, the outer line of the tower is used to make the center line of the tower in this direction and the inclination of the tower in this direction is calculated. The experiment uses the data provided by Shanxi State Grid Electric Power Company for verification. The tilt detection effect of the tower is more accurate under various backgrounds, and the accuracy and environmental adaptability are significantly improved compared with other algorithms. The correct recognition rate of the tower target reaches 97%, and the average error of the inclination detection is less than 0. 85°.