Autopilot control unmanned aerial vehicle system for sewage defect detection using deep learning

Abstract

This work proposes the use of an unmanned aerial vehicle (UAV) with an autopilot to identify the defects present in municipal sewerage pipes. The framework also includes an effective autopilot control mechanism that can direct the flight path of a UAV within a sewer line. Both of these breakthroughs have been addressed throughout this work. The UAV's camera proved useful throughout a sewage inspection, providing important contextual data that helped analyze the sewerage line's internal condition. A plethora of information useful for understanding the sewerage line's inner functioning and extracting interior visual details can be obtained from camera‐recorded sewerage imagery if a defect is present. In the case of sewerage inspections, nevertheless, the impact of a false negative is significantly higher than that of a false positive. One of the trickiest parts of the procedure is identifying defective sewerage pipelines and false negatives. In order to get rid of the false negative outcome or false positive outcome, a guided image filter (GIF) is implemented in this proposed method during the pre‐processing stage. Afterwards, the algorithms Gabor transform (GT) and stroke width transform (SWT) were used to obtain the features of the UAV‐captured surveillance image. The UAV camera's sewerage image is then classified as “defective” or “not defective” using the obtained features by a Weighted Naive Bayes Classifier (WNBC). Next, images of the sewerage lines captured by the UAV are analyzed using speed‐up robust features (SURF) and deep learning to identify different types of defects. As a result, the proposed methodology achieved more favorable outcomes than prior existing approaches in terms of the following metrics: mean PSNR (71.854), mean MSE (0.0618), mean RMSE (0.2485), mean SSIM (98.71%), mean accuracy (98.372), mean specificity (97.837%), mean precision (93.296%), mean recall (94.255%), mean F1‐score (93.773%), and mean processing time (35.43 min).