The Research For Recognition Of Moving Vehicle

Recognition system of moving vehicle based on video images mainly consists of two key technologies: recognition of vehicle_license_plate (VLP) and recognition of vehicle type. It not only finds wide application in ITS, but also is a hot point of research in computer vision, image processing and pattern recognition. So its related technology is attended prevalently. On the above background, we made a deep and systematic research for recognition of moving vehicle. In recognition of VLP, the problems of VLP location and VLP character recognition are studied in this dissertation. In recognition of vehicle type, differing with other researchers who attend to the recognition of shape, size of vehicle, we pay attention to the location and recognition of vehiclelogo. The methods to deal with above problems are proposed in the dissertation and testified in the experiment. At the same time, our researches have academic significance in object recognition which is subject to illumination, noise, size, slant, deformed, shape similarity and part occlusion. The main content and innovation of the dissertation are as follows:1. Feature extraction and recognition algorithm in object recognition: The recognition of moving vehicle is a typical system of automatic target recognition (ATR), while feature extraction and recognition algorithm are key problems in ATR. Based on generalization and analysis of current methods of feature extraction and recognition, the wavelet hidden markov models (WHMM) is emphasized to be studied. WHMM is a statistical model in wavelet domain of stochastic signals, which integrates wavelet transform into hidden markov models (HMM). According to the locality, multiscality of wavelet transform and context-sensitive property of HMM, WHMM can not only characterize precisely the local feature of object multiscally, but also reflect accurately the full structure information of object. So it is a powerful tool which deals with noise, shape similarity and part occlusion in ATR, and finds its full application in recognition system of moving vehicle. The recognition and training of WHMM are optimized in the dissertation. In addition, an improved wavelet hidden markov tree models (WHMTM) is proposed to better segment texture feature of object. We also propose several perspective invariant wavelet descriptors to construct observation sequences of wavelet hidden markov chain models (WHMCM). By doing so, it has wider application in ATR.2. VLP location algorithm: VLP location in a vehicle image is a key step in VLP recognition system. Its difficulty is that because of noise, the segment threshold of norm method is hard to adjust, resulting in errors in VLP location. According to the texture feature of VLP, a fast location method which consists of an adaptive energy filter and WHMTM is proposed in the dissertation. It first segments fast VLP candidates from image by an energy filter with an initial threshold, then WHMTM is used to judge whether the candidates are true VLP or adjust threshold of energy filter to restart to segment candidates. Experimental results show that it is a fast, robust and accurate method of VLP location. 3. VLP character recognition algorithm: Because of shape similarity, strokes rupture, noise problems, the errors are often seen in VLP character recognition. Thus, conventional methods of character recognition, such as template match, cannot get satisfactory result. So the new VLP character recognition algorithm should be studied to improve correctness. A multilevel VLP character classifier which consists of feature point match and WHMCM is proposed in the dissertation. Template match is made as the first classifier to select the candidates, and then feature point match and WHMCM are made as the second classifier to get the final results from them. Experimental results show the method can efficiently deal with above problems. 4.Vehiclelogo location algorithm: While VLP has evident texture feature and regular shape, the texture feature, shape and size of every vehiclelogo are d...