Options
2015
Conference Paper
Titel
Object-based detection of vehicles in airborne data
Abstract
Robust detection of vehicles in airborne data is a challenging task since a high variation in the object signatures - depending on data resolution - and often a small contrast between objects and background lead to high false classification rates and missed detections. Despite these facts, many applications require reliable results which can be obtained in a short time. In this paper, an object-based approach for vehicle detection in airborne laser scans (ALS) and photogrammetrically reconstructed 2.5D data is described. The focus of this paper lies on a robust object segmentation algorithm as well as the identification of features for a reliable separation between vehicles and background (all none-vehicle objects) on different scenes. The described method is based on three consecutive steps, namely, object segmentation, feature extraction and supervised classification. In the first step, the 2.5D data is segmented and possible targets are identified. The segmentation progress is based on the morphological top-hat filtering, which leaves areas that are smaller than a given filter size and higher (brighter) than their surroundings. The approach is chosen due to the low computational effort of this filter, which allows a fast computation even for large areas. The next step is feature extraction. Based on the initial segmentation, features for every identified object are extracted. In addition to frequently used features like height above ground, object area, or point distribution, more complex features like object planarity, entropy in the intensity image, and lineness measures are used. The last step contains classification of each object. For this purpose, a random forest classifier (RF) using the normalized features extracted in the previous step is chosen. RFs are suitable for high dimensional and nonlinear problems. In contrast to other approaches (e.g. maximum likelihood classifier), RFs achieves good results even with relatively small training samples.