A Real Time Object Detection Approach Applied to Reliable Pedestrian Detection
Guanglin Ma, Su-Birm Park, Alexander Ioffe, Stefan Müller-Schneiders, Anton Kummert · IEEE Intelligent Vehicles Symposium · 2007
This paper presents a robust real time obstacle and pedestrian detection algorithm, which is capable of handling the challenges of stationary as well as moving objects, utilizing a single car mounted monochrome camera. First, the system detects obstacles above the ground plane by obtaining a "virtual stereo system" through the usage of inverse perspective mapping. A fast digital image stabilization algorithm is used to compensate erroneous detections whenever the flat ground plane assumption is an inaccurate model of the road surface. Finally, a low level pedestrian segmentation algorithm is developed to extract bounding boxes of potential pedestrians. Furthermore a novel approach called the pedestrian detection strip is used to improve the calculation time by a factor of six compared to previous attempts. Experiments have been carried out by applying the proposed algorithm on prerecorded sequences as well as within a test vehicle and thus in a closed loop environment. The experimental results indicate a promising detection performance. Obstacles and pedestrians up to 50 meters away from the vehicle have been detected reliably at 64 frames per second on a 3 GHz PC.