Abstract-The detection of small objects from aerial images is a difficult signal processing task. To localise small objects in an image, low-complexity geometry-based approaches can be used, but their efficiency is often low. Another option is to use appearance-based approaches that give better results but require a costly learning step. In this paper, we treat the specific case of manhole covers. Currently many manholes are not listed or are badly positioned on maps. We implement two conventional previously published methods to detect manhole covers in images. The first one searches for circular patterns in the image while the second uses machine learning to build a model of manhole covers. The results show non optimal performances for each method. The two approaches are combined to overcome this limit, thus increasing the overall performance by about forty percent.

International audienceMispositioning of buried utilities is an increasingly important problem both in industrialized and developing countries because of urban sprawl and technological advances. However, some of these networks have surface access traps, which may be visible on high-resolution airborne or satellite images and could serve as presence indicators. We put forward a methodology to detect manhole covers and grates on very high-resolution aerial and satellite images. Two methods are tested: the first is based on a geometrical circular filter, whereas the second one uses machine learning to retrieve some patterns. The results are compared and combined to benefit from the two approaches

The detection of small objects from aerial images is a difficult signal processing task. To localise small objects in an image, low-complexity geometry-based approaches can be used, but their efficiency is often low. Another option is to use appearance-based approaches that give better results but require a costly learning step. In this paper, we treat the specific case of manhole covers. Currently many manholes are not listed or are badly positioned on maps. We implement two conventional previously published methods to detect manhole covers in images. The first one searches for circular patterns in the image while the second uses machine learning to build a model of manhole covers. The results show non optimal performances for each method. The two approaches are combined to overcome this limit, thus increasing the overall performance by about forty percent.