Car detection in low frame-rate aerial imagery of dense urban areas

Abstract

Knowledge about quantity and position of moving and stationary vehicles is essential for traffic management and planning. This information can be used, for instance, for security of mass events or to support rescue crews in disaster situations. In order to get this information, large areas have to be examined quickly and completely. Very suitable for this task are airborne optical sensors. However, a reliable automatic method to locate vehicles in aerial images is necessary. In the present work a method for automatic extraction of vehicles in urban areas is presented. The work mainly covers three key fields of car detection. The first is related to the extraction of ground areas. On the assumption that trafficable areas are often ground areas in densely populated cities, disparity maps are calculated using the semi-global matching algorithm (SGM). Subsequently, a threshold is automatically determined to separate ground from non-ground regions (Minimum Error Thresholding). The second field concerns the introduction of a object-based method for extracting car candidates. In order to do this, the image is smoothed using the mean curvature flow, and a region-growing algorithm is then applied. The regions obtained are considered autonomous regions and are filtered multiple times with regard to their geometric properties. The third field is the examination of the remaining candidate regions by a classifier based on gradients (HOG features), which is trained by a machine learning algorithm (AdaBoost). However, the classifier is trained using only a few training samples. The goal is to minimize the manual effort and to provide a high degree of generalization. Thus, a strategy is presented which combines object-based and gradient-based techniques. The strategy is tested with five urban images from the 3K+ camera system and the UltraCam Eagle camera system, with 13 cm and 20 cm GSD, respectively. Through the use of disparity maps, it is shown that the car detection quality in densely populated inner-city areas can be enhanced. Objects on the top of buildings are now accurately excluded from the detection process. Furthermore, the car detection approach presented is able to detect cars in different datasets without adjustment of parameter settings (different sensors and different resolution). The results of detection show that a completeness of 80% leads to a correctness of 65% to 95%.