Automated localization of urban drainage infrastructure from public-access street-level images

ABSTRACT

Comprehensive management of urban drainage network infrastructure is essential for sustaining the operation of these systems despite stresses from component deterioration, urban densification, and a predicted intensification of rainfall events. In this context, up-to-date and accurate urban drainage network data is key. However, such data is often absent, outdated, or incomplete. In this study, a new approach to localize manhole covers and storm drains, using deep learning to mine publicly available street-level images, is presented, tested, and assessed. Thus, the time-consuming and costly acquisition of the location of these system components can be avoided. The approach is evaluated using 5,000 high-resolution panoramas covering 500 km of public roads in Switzerland. The object detection approach proposed shows good performance and an improvement over state of the art image-based urban drainage infrastructure component detection. While the geographical localization of the detected objects still contains errors, the accuracy achieved is nevertheless sufficient for some applications, e.g. flood risk assessment.

KEYWORDS:

• Deep learning
• highresolution streetlevel imagery
• urban drainage system components

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. https://developers.google.com/maps/documentation/streetview/intro.

2. https://github.com/tensorflow/models/tree/master/research/object_detection.

3. An anchor is a reference square box defined by its ratio and scale. For each image location (defined by a sliding window) the original Faster R-CNN implementation considers multiple anchors (nine in total: three scales and three ratios) (Ren et al. 2017).

4. https://maps.zh.ch/.