Research On Key Techniques Of Visual Navigation In Unstructured Road Environment

As an important research field in Intelligent Transportation System, visual navigation techniques could be used to improve on the efficiency of urban traffic, increase the driving safty and decrease the unnecessary hunman death and damage. Compared with the autonomous navigation based on the fusion of multi-sensors, visual navigation techniques supply a low-cost resolution and therefore possess more potential applications.In terms of the unstructured road, this paper mainly handled the contradiction between real-time and robustness caused by huge image data in complicated road environments. Based on the consistency hypothesis within the road region and the inconsistency hypothesis between the road and the background, this paper focused its researches on the road region segmentation and roadside detection.This paper first improved on the classical Otsu thresholding method by peak searching in grey histogram so that it could be available in double-threshold sementation for unstructured road. By the correlation of segmentation result between current images and past image, the feature similarity between current image and reference region, the segmentation quality could be improved greatly.Roadside detection method was then researched in the following step by using a heuristic probabilistic Hough transform to increase the precision and the robustness in unstructured roadside detection. This paper then models the roadside by double line segment so that the realtime motion planning in both line and curv road environment could be realized.A roadside detection method based on edge matching was proposed in this paper to combine the road region segmentation with roadside detection. It optimized the segmentation threshold by matching the boundary between road region and non-road region produced by double-threshold Otsu method with the weighted Canny edges. Subsequently, it would refresh the weighted Canny edges in the produced road and non-road region. Based on the evaluated weighted Canny edges, the edge matching method realized unstructured roadside tracking by using adaptive Monte Calo Method and modeling the roadside with double-line segement. By virtue of the adaptive selection for particle size and the mechanism of particle sampling in perception model, it could improve the precision of the roadside detection and overcome the particle degradation lies in traditional Monte Carlo method.To overcome the problem of grey inconsistency in certain regions on actual road, this paper proposed a 3d rebuilding method for monovision, which realizes optical flow detection and 3d reconstruction for the region of interest. By constructing the layered structure for the scale invariant features transform and corner fertures, the obstacles in road region could be detected in time.The research on above problems supplied important theory and technique support for visual navigation and auxiliary driving for intelligient vehicle, therefore can be extensively used in intelligient transportation, autonomous driving for intelligient vehicle, industry product transportation and military field. The key techniques above can be directly used in auxiliary safe driving for intelligient vehicle, decrease the quantity of traffic accident caused by the negligence of the driver and therefore have huge social and economy benefit.