Several Key Problem Research Of The Intelligent Vehicle

In 1999, about 800,000 people died globally in road related accidents, causinglosses of around US $ 518 billion that equaled to 1.7% of the global GNP.Recently, developing the intelligent vehicle is one of the important motives ofthe governments, institutes and vehicle manufactures. Computer vision basedroad scene recognition and the vehicle state estimate is one the key problems,including lane recognition and tracking, vehicle recognition and tracking,etc.Although the many years of studies, the road scene recognition is still an openproblems. The problems studied in this paper is as follows: object trackingin clutters, lane detection and tracking, vehicle detection, vehicle tracking andbehavior prediction, and lane departure warning method.The contents and main contribution of this paper are as follows:Propose robust visual tracking methods by integrating multiple cues. First,a new intensity gradient based shape observation model is proposed. Comparedwith generative model, the new model has less parameters and preserve mostof information in images. In clutters, single cue based tracking method oftenfails. To solve this problem, this paper proposes two cue integrating methodunder the particle filter framework. The first algorithm calculate the observationmodel by fusing the color and shape cue, and it use information entropy tomeasure the cue reliability and to tune the cue weight. The second algorithm, the samples are drawn respectively from the color importance density and theshape importance density. The resulting tracker is robust to rapid, heavy clutter,occlusion, appearance changing.A MAP and PSO based lane detection method is proposed. The paramet-ric model of the projective projection of the lane borders on the image plane isintroduced, and the model parameters completely determine the position of thevehicle inside the lane, its heading direction, and the local structure of the road.The proposed lane detection approach makes uses of the parametric model, amaximum a posteriori (MAP) formulation of the lane detection problem, and aparticle swarm optimization algorithm to maximize the posterior density. Ex-periments show that the proposed algorithm can handle the situations where thelane boundaries in an image have relatively weak local contrast, or where thereare strong distracting edges. The percentage of correct lane detection is over96%, depended on the real road conditions.Develop a lane tracking approach to estimate the vehicle position inside thelane, lane local geometry structure. The largest challenge facing lane trackeris the variability likely caused by differing road conditions, weather conditionsand the types (and quality) of lane markers. Such condition makes feature ex-traction difficult and prone to fail, in particular the presence of background fea-tures incorrectly classified as lane markers. The techniques of data association,robust line fitting and averaging used in the past work are efforts to reduce theeffects of these distracters prior to lane shape estimation. By contrast, the pro-posed method is to diminish the effects of the distracters by using particle filter,that was designed specifically with estimation in presence of clutter in mind. The new approach deals with clutter at higher level of estimation, rather thanwith lower lower level image processing. Experiments show that the percent ofcorrect lane tracking is over 98%.Develop a nearby-vehicle detection method and a distant-vehicle detectionmethod. The proposed nearby-vehicle detection approach is as follows: first,sample the image, perform an edge detection, and use planar parallax modelto predict what that edge image will look like after traveling a certain distance.Next, capture an image after traveling our assumed distance, and compare it tothe prediction. For each point in the predicted image, the method verify thatthere is a corresponding edge point in the actual image. If there is a match, thenour prediction which is based on a ?at earth assumption is verified. Otherwise,the cause of the horizontal line in the predicted image is an obstacle which isabove the ground plane. Detecting vehicles in the mid-range/distance regionis as follows: using the shadow feature to find candidate vehicle locations inan image, then tests are performed to verify the correctness of a hypothesisby using texture feature and symmetric feature. Finally, the vehicle locationis precisely determined using a MAP methods. Experimental results show thepercent of correct vehicle detection is over 95%.Develop a vehicle state prediction method. Two vehicle tracking methodsare developed: The first algorithm fuses the color and shape cues to track ve-hicle, the second algorithm is based on multi-model switching method, andits output is qualitative and quantitative description of the vehicle state. Theoutputs of the vehicle tracking module are the trajectories of vehicles and thefeatures of vehicles such as size. These outputs form the sample data for learn- ing activity patterns. After obtaining enough sample data, we can learn thevehicle- activity patterns from the sample data using a fuzzy self organizingneural network. The activity patterns can be thought as the classification ofvehicles'activities. In prediction step, the sub-trajectories of the vehicle iscompare with the learnt activity model, and the trajectory represented by theneuron with the highest probability is chosen as the most probable one alongwhich the object will move in the future.Propose a virtual lane boundary based lane departure warning system(VLWM). VLWM allows the driver to drift beyond the physical lane bound-ary by adding a virtual lane boundary. Accounting for the driving habit, lanegeometry, and the local driver behavior changes, the virtual lane width is de-termined using a fuzzy logic based method. When the vehicle is predict toexceed the virtual lane boundary, an alarm is triggered. Real world driving dataare used to test the lane departure warning systems (LDWS) The results show:compared with RRS system, the warning time increase about 0.7s, while thefalse warning rate is almost similar, less than 4 per hour. The warning time ofVLWM is 0.1s shorter than that of TLC based methods, but the false warningrate of the VLWM is less than 4 per hour, which is much less that 15 times perhour of TLC based methods.