Research On Stereo-based Obstacle Detection With Weighted Hough Transform And Local Temporal Correlation

Due to the popularity of automobile in our daily life, developing autonomous systems that are able to assist daily driving is one of the grand challenges in the field of computer vision. Obstacle detection is a fundamental task for a variety of important applications which is used in the field of Advanced Driver Assistance Systems (ADASs) and Intelligent Vehicles (IVs).In this paper, we propose a novel robust approach to detect obstacles with the road surface profile feature by Weighted Hough Transform (WHT) and local temporal correlation, which is computed after the depth information is obtained by semi-global match (SGM) from gray-scale stereo video sequences. The piecewise linear curve is used to understand the surrounding scenes from real world by modeling the road surface profile in v-disparity domain. The final segmentation of free space and obstacles below the horizon line in the reference image is performed with the road profile extracted previously.Our approach relies on the disparity image from the Semi-Global Matching which is believed to be most efficient dense stereo matching algorithm so far in real traffic scenes. To overcome the loss of information in road profile extraction, weighted Hough Transform is introduced to find the principal linear components in v-disparity domain. Besides, temporal local correlation plays an important role of smoothing the final road surface profile in temporal dimension. The result shows that it successfully handles most case including traffic jam, urban streets, rural highway etc. In future work, we plan to optimize the performance by fusing more information from other sensors. Collision avoidance alert and obstacle segmentation will be important directions in the following research.