Free Space Detection Using Stereo Confidence Metrics and Obstacle Position Probability Maps

This paper proposes a novel free space detection method in complex traffic scenes. First, stereo confidence cues are adopted to construct a pixel-wise outlier detection model, which can reduce the false detecting rate caused by inaccurate disparity values. Second, in order to detect slender and farther obstructions, our approach gives a global optimization model which estimates free space by calculating the obstacle position probability map. An obstacle position probability map is a 2D map which shows the probability that each pixel belongs to an obstacle position. The proposed method can stably detect free space in real time in a variety of complex traffic situations, such as urban scenes, highways and secondary structured roads. An in-depth assessment of the actual traffic database validates the effectiveness of our approach. We show that the new confidence-based detection model is robust to the environment and it can greatly improve the free space detection effect.