Abstract
In modern cars, driver assistance systems support the motorist to handle the ever increasing complexity of the driving task. These systems aid to avoid accidents and alleviate their effects, thus reducing the number of casualties. To reduce the amount of rear-end collisions the ACC (Adaptive-Cruise-Control) was introduced a few years ago. It takes over the longitudinal guiding, by controlling the distance to the vehicle ahead. If the driving-lane is free, the car accelerates to the chosen velocity. The system uses radar-sensors, monitoring an area of up to 200 meters in front of the vehicle. There are two problems facing system, which use solely radar-sensors: With a narrow angle of beam of typically sixteen degrees objects in close distance can only be detected right in front of the sensor. In addition radar sensors cannot distinguish different kinds of objects. Therefore no object-specialized vehicle behaviour can be implemented. A supporting fisheye camera solves these problems. Via specialized algorithms all kinds of objects are distinguishable.
In a visual driver-assistance system, obstacle detection during driving is one of the major tasks. This paper presents a new vehicle detection method based on multi-features fusion in the images acquired by a fisheye camera. The vehicle detection algorithm can be divided into three main steps: fisheye image calibration, generation of candidates with respect to a vehicle and verification of the candidates. In this paper vehicles are detected through typical features of their front and rear perspective, like shadow and symmetry. The use of a fisheye camera with an angle of beam exceeding 180° allows the detection of objects not only for ACC or forward collision warning systems, but also enables close-up lane and pedestrian detection, as well as sensing overtaking cars earlier.
In a first step the car detection capabilities of similar systems using non-wide-angle cameras shall be reached. Due to the lower resolution in the central part of the camera picture the distance of detection is expected to be smaller. The vehicle detection was accomplished using the well established concept of hypothesis generation and verification. Whereby a larger number of vehicle hypotheses is generated in the first and filtered in the second step.
The system was tested in various weather and road conditions. Experimental results in different conditions, including sunny, rainy, snowy demonstrates that most vehicles can be detected and recognized with a high accuracy and a frame rate of approximately 16 frames per second on a standard PC.
Keywords: obstacle detection, fisheye camera, wide angle view, HOG, SVM