Research On Vehicle Recognition Based On Multi-feature

With the process of technology, and the launching and promoting of large securityprojects, video surveillance technology gets the wide attention of academia and industry,and is applied to real life diffusely. The video monitoring data is with long duration, bigdata volume, relatively independent of vision field and such wealth of the complexinformation. The cost of traditional manual monitoring is amazing, and the efficiency isvery low. As the foundation of traffic surveillance video data analysing technology, targetfeature extraction and matching technology get more and more attention.The recognition of vehicle based on multi-feature consists of two parts: the localfeature extraction and the matching between feature descriptors. The aspect of vehiclelocal feature extraction includes the edge detection of headlights and the surf featureextraction of vehicle front image. The deficiencies and limitations are proposed in theview of vehicle recognition system. An edge detection algorithm of current algorithms andtechnology of edge detection and local feature extraction is proposed based on thetraditional Canny algorithm with the refinement on the elimination of false edges andnoise points to improve the accuracy of image edge detection and eliminate the noisepoints. Afterwards, in aspect of matching between feature descriptors, the currentcommonly used methods are presented and analysed. Based on the characteristics of thevehicle recognition system, the improvements in position constraint and distancecalculation are proposed when the feature descriptors are matching. The improvementscan reduce the meaningless matching, and weaken the dependence on maximum featuredescriptors and enhance the sensitivity of features in smaller scale.By analysis of experimental data, the accuracy of the improved Canny edge detectionalgorithm is higher compared with the original Canny algorithm, and the results are lessaffected by the noise. The refinement in distance calculation and position constraintbetween feature descriptors matching results that the efficiency is higher, and the numberof matched feature points is more than the original algorithm. Experimental results showthat accuracy of the recognition based on multi-feature is higher and more credible thanthe recognition based on a single feature.