Multisensor Based Obstacles Detection in Challenging Scenes

Obstacle detection is a significant task that an Advanced Driving Assistance System (ADAS) has to perform for intelligent vehicles. In the past decade, many vision-based approaches have been proposed. The majority of them use color, structure and texture features as clues to group similar pixels. However, motion blur generated by the movement of obstacles during exposure is not taken into account in most of the approaches. Generally, many visual clues could fail due to this problem. In this paper, we propose a method, which is independent to the visual clues of target obstacles, to deal with this problem. The proposed approach integrates fisheye image, laser range finder (LRF) measurements and global positioning system (GPS) data. Firstly, the road is detected in fish-eye image by a classification algorithm based on illumination-invariant grayscale image. Secondly, the corresponding geometrical shape of the road is estimated using a geographical information system (GIS). Based on the road geometrical shape, the possible regions of obstacles are then located. Finally, LRF measurements are used to check if there exist obstacles in the possible regions. Experimental results based on real road scenes show the effectiveness of the proposed method.