The Research Of Key Technologies On Object Detection Based On Vision Navigation For Intelligent Vehicle

Intelligent vehicle is an important component of Intelligent TransportationSystem (ITS); and also it is a key step in achieving intelligent transportation.Intelligent vehicle is a synthesis body which can percept environmental informationand make a decision automaticly. It involves many theories and techniques such aspattern recognition, image processing, automatic control theory and othertechnologies. It represents the latest achievements of information science andartificial intelligence.Among the many key technologies of intelligence vehicle,vision navigation is base of perception environmental information and making thenext step decision. Since visual information such as color image has advantages ofinformation-rich, passive non-polluting and non-invasive detection, so it has beenattentioned widely, and has become the commonly used technology for realization ofautonomous navigation of intelligent vehicles. Especially road detection and trafficsigns detection and obstacle detection are an essential function in vision basednavigation, and it is the important precondition to prevent dangerous condition and insave driving. But there are many factors such as the complexity and environmentuncontrollable of intelligent vehicle while working in the outdoor environment, thevolume of color images in scene and the computational complexity are very high,which result in poor performance of system’s real-time, accuracy and robustness invision navigation. To improve the performance, enhance the effectiveness of objectdetection system, based on the existing techniques and theories, the paper researchedfollowing key techniques in detail.1. In order to overcome the computational complexity and enhance the speed ofwhole system, the paper proposed an effective color image preprocessing algorithm ofintelligent vehicle. As H component is instable when saturation is too low, so select athreshold T on the saturation component, and divided S component into the regions ofhigh saturation in low saturation, then project H component in the high saturation areaand V component in low saturation area, the projected components is stretched in S.while the S-component contains not only necessary color information but alsograyscale information, at the same time, the images reduced from the originalhigh-dimension to two-dimension. This algorithm not only overcomes thesegmentation inaccuracy caused by traditional method which lost color information,reduce the amount of data in follow-up treatment significantly, improve real-time, andalso improve the robustness against light and shadow effect significantly.2. In order to resolving the shortcoming of region imperfection in segmentationand increasing the fault-tolerant, the paper proposed a novel road detection algorithmbased on semantic model and multi-neural network adaptive detection. The algorithmsegment road image using multi-threshold method on histogram firstly, then projectsegmented area by multi-neural network into the semantic model. The algorithm caneffective overcome the problem of self-adaptive during the road occurred change, andcan realize the full road region extraction in unstructured road.3. Aimed at overcoming the deficiency of high computational complexity and difficulty of extracting complex background in active traffic sign detection, the paperproposed a rapid passive traffic signs detection algorithm. Firstly, the algorithmconstructs roughness histogram for olor which included in traffic signs. If there aretraffic signs in road can bedeterminated according to color matching table, theroughness and it’s background and foreground color. If there are no traffic signs in theroad image, the algorithm do not do following work, if there are traffic signs in theroad image, it extract the region according to roughness statistics in RSH, so thealgorithm avoid to blind detection and repeat judgment during the whole process, andcan dramatically enhance the effectiveness.4. In the aspect of obstacle detection, in order to overcome the problem of bothlow accuracy in color and edge detection and low comprehensiveness in informationextraction, the paper proposed a spatial vision obstacle detection algorithm based onquestionable area match techniques. The algorithm segment questionable area firstly,then match it at the first time, during the results of it, if matched success, thealgorithm matched second time. Thus the algorithm not only decreased the specimenspatial, but also offered much useful information.