An Intelligent Unmanned Aerial Vehicle For Detection of Pipeline Faults built on Microsoft Airsim
Leakages on Pipeline continues to be a problem in the petroleum sector, but this research presents the use of state of the art techniques in Robotics and Artificial Intelligence to tackle this menace. This project makes use of an intelligent drone enabled by computer vision algorithm to detect pipeline faults. The Drone is able to perform the following
- Real time 3D Reconstruction of the environment
- Detection of Pipeline Faults using Computer Vision algorithm
- Obstacle avoidance during drone navigation
- Tracking of Pipeline path using Path Dijkstra's and A* algorithm
Video Demo :
- Unreal Engine 4.27
- Ubuntu 18.04 (WSL2)
- Q Ground Countrol on Windows
- Airsim on Windows
- PX4 SITL
python version >= 3.8
pip install roboflow
pip install open3d
pip install msgpack-rpc-python
pip install airsim
pip install pandas pillow tqdm
The Drone has a pay load of Depth Camera, RGB Camera and a LIDAR sensor
The simulation was carried out using Microsft Airsim on Windows and on Ubuntu 18.04 (WSL2)
The simulation environment was built from scratch on the Unreal Engine 4.27 Platform. The Unreal engine provides a photorealistic simulation environment with High Visual Fidelity
The 3D Map recosntruction was carried out using the fusion of the sensor readings from lidar sensor and the depth sensor. The reconstructed map when compared to ground truth comprises of rgb image and segmentation point cloud. The segmentation point cloud gives better seen understanding of the environment by segmenting objects in the scene using colour mapping. The point cloud data can be visualized using CloudCompare
The You-Only-Look-Once v8 and Detection Transformer object detection model was used on the real-life dataset gotten from Roboflow . The syntehtic dataset was generated from the simulation environment and was trained on the Yolo v8 model. The weight from the training was tested in the simulation environment.
Training on Real life Dataset
Training on Synthetic Dataset
The DETR was used on the Real life dataset but because of its performance it was not used on the synthetic dataset.
Training on Real life Dataset
The weight of the Yolov8 model being trained on the synthetic dataset was tested in the flight simulation environment
The pipeline tracking was carried out using a combination of A * motion planning algorithm and Djikstra's algorithm
[1] Microsoft, “Airsim,” Microsoft, Jul. 20, 2023. https://microsoft.github.io/AirSim/ (accessed Jul. 20, 2023).
[2] B. Alvey, D. T. Anderson, A. Buck, M. Deardorff, G. Scott, and J. M. Keller, “Simulated Photorealistic Deep Learning Framework and Workflows to Accelerate Computer Vision and Unmanned Aerial Vehicle Research.” Accessed: Jul. 21, 2023. [Online]. Available: https://github.com/MizzouINDFUL/UEUAVSim
[3] Roboflow, “Pipe Burst Dataset,” 2021.https://universe.roboflow.com/zzi-driha/pipe-burst