Pixels, Stixels, and Objects

Dense stereo vision has evolved into a powerful foundation for the next generation of intelligent vehicles. The high spatial and temporal resolution allows for robust obstacle detection in complex inner city scenarios, including pedestrian recognition and detection of partially hidden moving objects. Aiming at a vision architecture for efficiently solving an increasing number of vision tasks, the medium-level representation named Stixel World has been developed. This paper shows how this representation forms the foundation for a very efficient, robust and comprehensive understanding of traffic scenes. A recently proposed Stixel computation scheme allows the extraction of multiple objects per image column and generates a segmentation of the input data. The motion of the Stixels is obtained by applying the 6D-Vision principle to track Stixels over time. Subsequently, this allows for an optimal Stixel grouping such that all dynamic objects can be detected easily. Pose and motion of moving Stixel groups are used to initialize more specific object trackers. Moreover, appearance-based object recognition highly benefits from the attention control offered by the Stixel World, both in performance and efficiency. Fig. 1. Dense disparity image and the corresponding Stixel World representing the three-dimensional environment in front of the vehicle. The colors encode the distances, with red = close, green = far away, and gray = freespace. The arrows show the motion vector of the tracked objects. This medium-level representation achieves a reduction of the input data from hundreds of thousands of single depth measurements to a few hundred Stixels only. 2 David Pfeiffer, Friedrich Erbs, and Uwe Franke