Extracting traffic information from the air
The detection of vehicles in aerial images is important for various applications e.g. traffic management, parking lot utilization, urban planning, etc. Collecting traffic and parking data from an airborne platform gives fast coverage over a larger area. Getting the same coverage by terrestrial sensors would need the deployment of more sensors, more manual work, thus higher costs. A good example for an airborne road traffic measuring system is the one in the project VABENE of the German Aerospace Center (DLR). In this real-time system aerial images are captured over roads and the vehicles are detected and tracked across multiple consecutive frames. This gives a fast and comprehensive information of the traffic situation by providing the number of vehicles and their position and speed.
The vehicle detection is a challenging problem due to the small size of the vehicles (a car might be only 30x12 pixels) and the complex background of man-made objects which appear visually similar to the cars. Providing both the position and the orientation of the detected objects supports the tracking by giving constraints on the motion of the vehicles. This is particularly important in dense traffic scenes where the object assignment is more challenging. The utilization of roads and parking lots depends also on the type of the vehicle (e.g. a truck impacts the traffic flow different as a personal car). A system having access to this richer information can manage the infrastructure better. In a real-time system, as in Vabene, the processing time (and computing power) is limited. Therefore the processing method should be as fast as possible.
To address these challenges we apply computer vision and machine learning techniques and test our methods both in experimental and operational settings.
Dataset over Munich
To help the research we provide a dataset of aerial images with vehicle annotations. This can be found in the downloads.
Figure 1: The detected cars are highlighted in an aerial image from our dataset. The green and cyan rectangles show different vehicle types, while the black rectangles show not detected cars.
Figure 2: The detected cars are highlighted in an aerial image from our dataset. The green and cyan rectangles show different vehicle types, while the black rectangles show not detected cars.