On Road Scene Understanding Guided By High-level Information

Road-scene understanding is one of the key technologies of Autonomous LandVehicle (ALV). Vision-based road detection is one of the most popular topics in thisarea. In this paper, we studied the topic about how to introduce high-level informationabout the road to the road detection procedure. First, novel descriptions of the high-levelinformation about the known road map are proposed. Then, we use these descriptions totune the local feature detection procedure for road detection. We do experiments onstructured and unstructured road scenes respectively. Experiment results show thatintroducing high-level information to road detection is very efficient for suchapplications.In this dissertation, an osDoB (Orientation and Scale-tuned Difference of Boxes)operator is employed to deal with structured roads where land markings are available.Such an operator is used as the low-level filtering operator for the extraction of edges.Moreover the high-level information is also made use of as a global road map for tuningthe osDoB operator. And then such an operator is applied for detecting of lane markings.The global road map can roughly predict the orientation of lane markings, so that theresponses in the non-markings regions in low-level filtering process are greatlysuppressed. As a result, the robustness of the feature extraction of the lane markings isenhanced, and thus the performance of road markings detection is improved.For complicated unstructured road scenes, the high-level information is used as aroad distribution model to facilitate road segmentation. The probability of that a givenpixel belongs to the real road region is calculated, and such a probability is employed totune the segmentation process. The road distribution model can remove or alleviate theinfluence due to illumination, occlusions and so on. And the disadvantages of thesefactors are always suffered by traditional algorithms. As a consequence, the hitting rate(HTR) of road segmentation and the flexibility of the algorithm are both improved.The proposed method is valuated in both highway and plateau field environments.Results show that the proposed road detection method outperforms state of the artmethods.