Research On Road Scene Interpretation Techniques For Intelligent Vehicles

The development of Intelligent Vehicles (IVs) has imposing on the defense, society,economy and academy, and becomes the tactic research object of high technology of allcountries. With the promotion of urgent demand on the military, civilian and research field,IVs are changing with each passing day. Perception and interpretation of the roadenvironment is the core technology of IVs, and is the concentrative embodiment ofvehicular intelligence, and is one of hot topics in research.In the thesis, the vision technologies on IVs for road interpretation are investigated,and the topics on the vanishing point estimation and tracking of structured road and itsapplication to road interpretation, and road segmentation of unstructured road, andtraversability analysis of cross country, and local map building based on multisensor areinvolved.In the introduction, the history and development of IVs are reviewed at start and thebasic concept and key technologies of road perception are introduced, subsequently thetechnologies about road detection and recognition, on-road vehicle detection and ruggedterrain analysis are summarized also.To interpretation of the structured roads, the estimation and tracking of VanishingPoint (VP) of road images are investigated, and the estimation of multi-VP based onBayesian theory is poposed and single VP tracking based on the Gaussian PredictiveModel (GPM) is developed, and the precision and speed of VP estimation and tracking canmeet the need of autonomus navigation. Moreover, a series of algorithms for roadinterpretation based on VP are developed in this thesis, specifically, include multi-lanedetection based on VP, branch detection and extraction based on VP, preceding vehicledetection based on VP, and relative range and range rate estimation to preceding vehicles.Experiments on many data sets show that these algorithms are effective and practicable.To resolve the shadow problem of unstructured road segmentation, the new definitionof spatial relation in road images is developed, and a novel spatial partition matrix isconstructed, then a modified spatial constrained Fuzzy C-Means clustering (FCM)algorithm is proposed. The experiments on avenue images show that our algorithm is priorto the classical FCM and the similar spatial constrained FCM. Moreover, the updatemethod to cluster prototype based on Bayesian criterion makes the segmentation algorithm converge faster and more stably.To analysis the traversability of rugged terrains for cross country, the concept ofrelative elevation invariant is developed, and the features which reflect the terrain'samplitude and frequency are extracted. Specifically, the terrain roll is computed by theelevation's variance; the terrain's slope is estimated based on the numerical solution ofcurve plane fitting; the terrain's roughness is estimated by the fractal Brownian motion(fBm) model. Lastly, the traversability of rugged terrain is evaluated based on fuzzyinference. The simulation and spot experiment results indicate the algorithm's effectivityand practicability.To resolve the practical problems in the project of Autonomous Land Vehicle (ALV),some algorithms are proposed to build local map based on multisensor, specifically, theroad edge fusion algorithm based on Covariance Intersection (CI), and the avenue trackingalgorithm based on active sensors, the rugged analysis based on relative features areinvolved.