Research On Vehicle Retrieval Algorithm Based On Attributes

With the development of economy and improvement of living standard, the number of motor vehicles is increasing rapidly. The growth of vehicles causes frequent occurrence of traffic accidents and violations which brings people’s driving serious hidden troubles. Vehicle retrieval is the technology that refers to retrieve the specified vehicles in the videos captured at intersections, and plays an important role on tracing causing-traffic-trouble vehicles.The purpose of this thesis is to study methods that retrieve the vehicles from the videos. Vehicle retrieval methods can be mainly classified into two kinds:one kind of methods is to extract attribute information of vehicles offline which is stored as key words, and then to retrieve the vehicles based on the input text information that describes vehicles’ attribute; the other kind of methods is to extract local features of vehicles offline, and then to retrieve according to input images’ content. The main works are as follows:(1) A vehicle detection algorithm combined with FPCP and Adaboost classifier is proposed.To increase the precision of vehicles detection under complex imaging condition, a vehicle detection algorithm combined with FPCP and Adaboost classifier is proposed. Firstly, FPCP algorithm is used to construct background image. Secondly, a background difference method is used to extract the vehicle candidates and then an Adaboost classifier is applied to detect the vehicles. The experimental results show that the proposed algorithm can detect vehicles well, and has strong robustness to complex imaging conditions. The accuracy of detection is about 99.25%.(2) A vehicle color recognition algorithm based on the invariance of illumination is proposed.To achieve vehicles’ color recognition under complex illumination conditions in vehicle videos, a vehicle color recognition algorithm based on the invariance of illumination is proposed. Firstly, the color feature model is built by analyzing the influence of illumination, and then the vehicle image is overlap blocked to extract the color features of every block. Secondly, the extracted color features are coded by radial basis in order to extract feature context. Finally, the features are classified by the trained SVM classifier to recognize the vehicle color. The experimental results show that the proposed algorithm achieves higher vehicle color recognition rate under a variety of illumination conditions, and the vehicle color recognition rate is about 93.90%.(3) An improved vehicle logo recognition algorithm based on HOG and SVM is proposed.To increase the precision of logos recognition under the conditions that majority voting methods can’t work well or logos are not in the training set, an improved vehicle logo recognition algorithm based on HOG and SVM is proposed. Firstly, the HOG features of the training images (logos and background images) are extracted to train SVM model. Secondly, the logos’ regions of interest are located by locating license plate accurately and prior knowledge of the positional relationship between the logo and the plate. Thirdly, the ROI of logos are scanned by multi-scale sliding windows which are predicted by SVM model. Finally, logos are recognized or rejected. The experimental results show that the proposed algorithm can greatly improve the recognition rate compared with the majority voting methods, at the same time can achieve accurate refusing recognition rate, and it can work well in the case of low quality, occlusion, incompleteness and varying light intensity. The recognition rate is about 93.34%, and the right rejection rate is about 90.11%.(4) A vehicle retrieval algorithm based on color information and Fourier-Mellin Transform is proposed.To increase the precision of vehicle retrieval under the condition that there are many similar vehicles on color, logo and model, a vehicle retrieval algorithm based on color information and Fourier-Mellin Transform is proposed. Firstly, the color of the images to be retrieved is recognized; the same color vehicle images are selected from the database to form a candidate set. Secondly, Fourier-Mellin Transform features’ similarity of the logo (include condenser) and headlight area between the query and the images in the candidate set are used to retrieve similar vehicles in the candidate set. The experimental results show that the proposed algorithm greatly reduces the time complexity and false retrieval rate, and is robust to common interference such as light, noise, etc al. Its recall rate and average accuracy rate are about 76.63%,91.33% and 62.23%,64.51% on the top 1%,5% of the database.The experimental results on the 10 actual shooting videos and the public available database show that the proposed algorithms outperform the state of the art algorithms. The precision and recall rates of the vehicle retrieval based on color information are 93.90%,93.90%; the precision and recall rates of the vehicle retrieval based on logo information are 93.34%,92.99%; the precision and recall rate are 88.17%,87.50% respectively when the inputs are combined with color and logo information. The recall rate and average accuracy rate of vehicle retrieval based on content are 76.63%,62.23% on the top 1% of the database.