Research On Otsu Image Segmentation Algorithm And Its Application In Vehicle Model Recognition

As an important research interests in pattern recognition and computer vision field, image segmentation has spread practical value. In the practical application of the target detection and segmentation, the consequences such as range ambiguity brought by the detection and segmentation of image, affect the real-time applications. Based on the depth exploration of the current situation of the image segmentation, the dissertation did research on the adaptive Otsu image segmentation algorithm, and put it into the combination with virtual loop recognition of traffic video models. The main works of the dissertation can be organized as follows:1. Based on the analysis and summarize the research results and the latest technology domestic and overseas, the threshold image segmentation was chose as the keynote of the research. At the earlier stage, the paper compares several methods of the image segmentation based on abundant analysis and research.2. Aiming at the shortage of the traditional 2D Otsu algorithm, absolute difference and average deviation was brought into the design of the threshold recognition function. Firstly, statistic the absolute difference with the class of objective and background of themselves to get the total value of the each class, and then, statistic the total average deviation, at last, the new threshold recognition function can be constructed by the commerce of the total absolute difference within-class and the overall deviation between-class. Experimental results show that compared with other functions, the 2D threshold value got by the improved function reach a better segmentation results, a better reservations of the object's outline and a lower calculation.3. It is a typical difficulty that traditional genetic algorithm has a great influence on the convergence while doing the unified operation during the course of crossover and mutation because it often falls into local optimal solution. Currently, there are many scholars do research on this issue, depend on the fitness value of the stocks classified in different populations, this paper uses different methods of crossover and mutation probability, namely, cross-way based on Hamming distance criterion, and the mutation probability based on dynamic changes, to a certain extent to avoid trapping in local optimum. Experimental results also show that the optimal threshold value obtained by the improved genetic algorithm is not only faster than the traditional genetic algorithm, but also closer to the global optimal solution, and to a certain extent overcome the premature problem of the traditional genetic algorithm.4. Via the image capture card to get video sequences'location settings to design the virtual loop lane, propose a method based on multi-frame considerations according the gray scale in the loop region to detect the arriving vehicles, and use Otsu algorithm to get the geometric characteristics of the vehicle by extracting object from the specific areas of vehicle movement, and through the analysis of geometrical features to achieve the right model identification. Experiments results show that the program has a high real-time and it could identify the arriving vehicles'model accurately.