Automatic Pipeline Surveillance Air-Vehicle

Abstract

This thesis presents the developments of a vision-based system for aerial pipeline Right-of-Way surveillance using optical/Infrared sensors mounted on Unmanned Aerial Vehicles (UAV). The aim of research is to develop a highly automated, on-board system for detecting and following the pipelines; while simultaneously detecting any third-party interference. The proposed approach of using a UAV platform could potentially reduce the cost of monitoring and surveying pipelines when compared to manned aircraft. The main contributions of this thesis are the development of the image-analysis algorithms, the overall system architecture and validation of in hardware based on scaled down Test environment. To evaluate the performance of the system, the algorithms were coded using Python programming language. A small-scale test-rig of the pipeline structure, as well as expected third-party interference, was setup to simulate the operational environment and capture/record data for the algorithm testing and validation. The pipeline endpoints are identified by transforming the 16-bits depth data of the explored environment into 3D point clouds world coordinates. Then, using the Random Sample Consensus (RANSAC) approach, the foreground and background are separated based on the transformed 3D point cloud to extract the plane that corresponds to the ground. Simultaneously, the boundaries of the explored environment are detected based on the 16-bit depth data using a canny detector. Following that, these boundaries were filtered out, after being transformed into a 3D point cloud, based on the real height of the pipeline for fast and accurate measurements using a Euclidean distance of each boundary point, relative to the plane of the ground extracted previously. The filtered boundaries were used to detect the straight lines of the object boundary (Hough lines), once transformed into 16-bit depth data, using a Hough transform method. The pipeline is verified by estimating a centre line segment, using a 3D point cloud of each pair of the Hough line segments, (transformed into 3D). Then, the corresponding linearity of the pipeline points cloud is filtered within the width of the pipeline using Euclidean distance in the foreground point cloud. Then, the segment length of the detected centre line is enhanced to match the exact pipeline segment by extending it along the filtered point cloud of the pipeline. The third-party interference is detected based on four parameters, namely: foreground depth data; pipeline depth data; pipeline endpoints location in the 3D point cloud; and Right-of-Way distance. The techniques include detection, classification, and localization algorithms. Finally, a waypoints-based navigation system was implemented for the air- vehicle to fly over the course waypoints that were generated online by a heading angle demand to follow the pipeline structure in real-time based on the online identification of the pipeline endpoints relative to a camera frame.