Abstract:In recent years, unmanned aerial vehicle (UAV) based vision inspections have been widely applied in electricity systems for both efficiency improvement and labor cost saving. Cable detection is essential for both navigation and flight safety of aerial vehicles. However, power cable detection is widely regarded as a challenging task since the targets are very weak and very easy to be confused with cluttered backgrounds. Traditional line and edge detectors are lack of robustness to scene variations. Recent deep learning based methods also can not support fast and stable power cable detection well for onboard applications . In this paper, a new convolutional neural network (CNN) based cable detection method is proposed. First of all, we encode cables by groups of evenly distributed key points, which reduce the complexities of detection tasks. By this approach, the proposed model detect grouped key points of cables from aerial images directly and the detailed pixels of cables can be restore with the curve equations which are implicitly behind those grouped key points. Subsequently, new methods of data labeling and augmentation, sample matching, post clustering, and performance evaluation for cable key points detection are presented. Finally, comprehensive experimental results demonstrate the efficiency and accuracy of our proposed cable detection method.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **How to quickly and accurately detect the position and direction of transmission power cables in the vision - based power system inspection by unmanned aerial vehicles (UAVs)**. This problem is crucial for ensuring the safe navigation and flight of UAVs, because transmission power cables are not only one of the most severe obstacles faced by UAVs, but also their position and direction information can be used for navigation and finding the next target object.
### Problem Background
1. **Limitations of Traditional Methods**
- Traditional edge and line detection methods assume that there is a significant contrast between power lines and the background, and that power lines are relatively straight in shape. However, in practical applications, this contrast is often low, and power lines are not always straight, resulting in these methods being prone to false detections.
- Although semantic segmentation methods in deep learning perform well on large - scale public datasets, when dealing with power cables, due to large computational requirements and the inability to provide detailed shape and direction information, it is difficult to meet the requirements of real - time navigation of UAVs.
2. **Deficiencies of Existing Deep Learning Methods**
- Existing object detection methods based on CNN (such as YOLO, SSD, etc.) perform well in general object detection tasks, but they mainly focus on predicting bounding boxes and lack attention to the internal relationships and global features of power cables, such as endpoints and quadratic polynomial parameters.
### The Method Proposed in the Paper
To overcome the above problems, the paper proposes a new power cable detection method based on convolutional neural network (CNN). The main innovations of this method include:
1. **Key - point Representation Method**
- Represent power cables as a set of evenly distributed key - points, thereby simplifying the detection task. Through this method, the model can directly detect these key - points from aerial images and restore the detailed pixel information of power cables through the curve equations implicit behind these key - points.
2. **New Sample Matching Method**
- A new sample matching mechanism is introduced, allowing the network to directly use pixels along power cables as positive samples for training. This enables the model to adapt more flexibly to power cables of different shapes and sizes.
3. **Post - processing Clustering Method**
- Use the DBSCAN algorithm to cluster the detected key - points to reduce redundant predictions and improve the accuracy of the final results.
4. **Data Labeling and Augmentation**
- Propose a set of data labeling and augmentation methods for key - points of power cables, including operations such as random translation, scaling, and flipping, to generate more training samples.
### Summary
This paper aims to solve the complexity and real - time problems in power cable detection. By introducing the key - point representation method, new sample matching mechanism, and post - processing clustering method, it significantly improves the speed and accuracy of power cable detection, providing strong support for the application of UAVs in power system inspection.