Research Of Video-based Vehicle Detection Algorithm For Intelligent Vehicle

With the economic development and the increase of car ownership, the trafficsafety problem is getting worse and worse in China, people property and life safetysuffered heavy losses.Using intelligent vehicles and other advanced technology toimprove road traffic safety has become the development direction of futuretransportation. This paper mainly researches vehicle detection technology based onvideo in intelligent vehicle. Using the on-board camera to get road scene video, weget environment perception information by analysising the road traffic video imagedata and privode it for drivers. Through reading some related literatures at home andabroad and then analyzing related algorithms, we put forward some correspondingimprovement algorithms. This research content mainly covers the followingrespects:(1)Vehicle hypothesis are generated through the adaptive shadow detection andSobel edge projection method. Firstly, we differentiate an area from the mid-lower ofthe video frame image, and then analysis it to get the vehicles shadow largethreshold value.Secondly, vehicle shadow position is located by segmenting imagesusing a threshold value method. Finally, we extract vehicle hypotheses combiningSobel edge projection. This method can effectively extract the vehicles region ofinteresting and reduce the number of vehicle hypothesis which is non-vehicle.(2) A new daytime vehicle detection algorithm is proposed combining SVM andGabor parameters optimization. We use the classification accuracy and punish factorand the number of support vectors to construct the fitness function, andsimultaneously optimize SVM and Gabor parameters using the Nniche GeneticAlgorithm. The best learning model and optimal Gabor filter are obtained afteroptimization. Gabor features of vehicle candidate region are extracted by using theoptimal Gabor filter and input to the SVM model to verify the vehicle candidate.Themethod can effectively reduce the number of dimensions of feature vector andimprove the accuracy of detecting system.(3) Contour four-neighborhood red level method is presented. Extracting vehicletaillights is a common means of vehicle detection at night.But the bright blocks,extracted by segmenting road scene image using threshold value method, containstreet lamps, noises and other non-taillights targets. We extract the contour of each bright region and effectively exclude false targets, such as non-rear light, by usingcontour four-neighborhood red level method.get the contour four-neighborhood redlevel by calculating the pixel red level of contour point and its four-neighborhoodpoint, and then set a proper threshold value. The method can effectively eliminatethe taillights goals.(4) A vehicle detection method using D-S evidence theory to fuse feature data ispresented. To get vehicles hypothesis, we combine the lamps using taillightsclustering algorithm.The structured feature information of taillights such as the arearatio, cross-image correlation and the rear light bounding box aspect ratio value isused to construct the basic belief distribution function, and then fusing these featureinformation get the general trust value by using D-S evidence theory. Finally we seta belief threshold value to verify validation vehicle hypothesis. Our method caneffectively reduce the number of subjective threshold and the threshold inappropriaterisk caused by inexperienced, and can improve the recognition rate.