Research On Vehicle Decoration Feature Based Vehicle Retrieval Method

In order to punish the traffic illegal criminality, our public security system has adopted a series of technical means to criminal forensics work, which the "Sky-net Project" acquires, transmits, displays and saves images by some equipment on the intersection of traffic for monitoring and recording. However, the equipment which installed at the section will capture tens of thousands of image data, the work for finding the target vehicle will be very difficult and the current solution is to search by the intrinsic property of cars such as the car plate number. Unfortunately, the effectiveness of the method is not satisfying for fake plate car. From our research we find that the vehicle decoration on the front windshield of the vehicle such as tags and ornaments are more obvious, therefore this article research the retrieval methods based on vehicle tags, ornaments to solve the problem.Firstly, the article studies the vehicle detection method based on the vehicle and vehicle plate symmetry features and positions the target section by the relative position between the vehicle and the front windshield. Then the article analyses the method based on the vehicle and the vehicle plate symmetry features and the methods based on the Adaboost classifier and Haar feature, the GLCM feature and SVM classifier, the HOG feature and SVM classifier comparatively. Experimental results show that the method based on the vehicle and the vehicle plate symmetry features is better than other three methods and the detection accuracy rate reaches 90.7%. Then we construct the Southeast University vehicle front windshield image sets by the detection method.Secondly, the article studies the vehicle retrieval method based on the color feature and the methods based on the local binary value model, the Gabor wavelet transform, the Contourlet Transform and the scale invariant feature are compared and analyzed. Experimental results show that the method based on the color feature is better than other four methods and the comprehensive index of retrieval is 86.7% and the average retrieval time is 29730ms.Finally, the article proposes a novel method of vehicle retrieval based on sparse coding and analyze the several methods of solving sparse vector from the under determined equation and according to the over-complete dictionary established through all the vehicle front windshield images and the input image retrieved, the solving method can obtain the sparse representation of the input image retrieved. The article conducts comparative experiments based on the constructed vehicle front windshield image sets. Experimental results show that the retrieval effect is best when the allowable reconstruction error is in the e-3, and the image retrieval method based on sparse coding is better than that of methods based on common features and the comprehensive index reaches 88.0% and the average retrieval time is 114100ms.