Abstract
A new method combining multi-view matching and feature extraction techniques is developed to detect manhole covers on streets using close-range images integrated with GPS/IMU and LiDAR data. While manhole covers are critical road infrastructure, automated detection is challenged by varying shooting angles, distances, ground materials, complex street scenes, shadows, and vehicular occlusions. This paper introduces an edge detection and feature extraction method to overcome these difficulties. LiDAR data is utilized for scene segmentation to exclude non-road elements. An edge detection method based on the Canny operator, sensitive to arcs and ellipses, is applied to segmented road scenes. Extracted arcs are fitted to ellipses, resampled for perspective invariance, and matched across adjacent images. Testing on over 1,000 images demonstrates the effectiveness of this multi-sensor approach in complex urban environments.
1. Introduction
Maintaining an accurate registry of manhole covers is essential for urban infrastructure management, yet manual surveying is labor-intensive. Mobile mapping systems (MMS) provide a wealth of data, but automated identification remains difficult due to perspective distortion, visual noise from shadows or tree cover, and false positives from other circular objects like car wheels. This paper presents a robust framework that leverages 3D spatial context from LiDAR and GPS/IMU alongside high-resolution imagery to achieve reliable automated detection.
2. Methodology
The proposed pipeline integrates geometric analysis with multi-view consistency checks. The process begins with data acquisition using a Ladybug fisheye camera for panoramic coverage, synchronized with GPS/IMU for positioning and LiDAR for 3D geometry.
2.1 Edge and Geometric Verification
Instead of searching for complete circles, which are often obscured, the system employs the EDISON edge detector to identify arcs. These fragments are then fitted into ellipses using RANSAC (Random Sample Consensus). This approach allows the system to reconstruct the full shape of a manhole even when partially hidden by shadows or parked vehicles. LiDAR data serves as a critical semantic filter, ensuring that candidates are located within 0.2m of the road surface, effectively eliminating false positives like traffic signs or logos on passing buses.
2.2 Multi-View Matching and Texture Verification
Candidates are validated by checking for consistency across adjacent image frames. Using georeferencing data, the 3D position of a candidate is projected into neighboring views. If the object does not appear consistently, it is discarded. Finally, a texture verification step analyzes intensity variance within the ellipse to distinguish the specific pattern of a manhole cover from the surrounding asphalt.
3. Experimental Results
The system was evaluated using a custom dataset of 1,100 high-resolution images containing 26 distinct manhole covers under various challenging conditions, including heavy shadows and different road materials.
The integration of texture verification significantly improved performance. While geometric edge detection alone achieved a 79% detection rate (at a 0.05 false positive rate), the combined pipeline reached an 88% detection rate. Furthermore, the system achieved a locating accuracy of better than 0.2m for the center of the covers, meeting the requirements for most urban planning applications.
4. Conclusion
This multi-sensor approach provides a robust solution for automated manhole cover detection. By focusing on arc fragments and utilizing 3D height filters, the system maintains high performance despite occlusions and perspective shifts. Future work will focus on improving "weak edge" detection for low-contrast environments and extending the methodology to other road assets such as traffic lights and signs.
Summary updated