The Research On Vehicle Identification Based On High Resolution Satellite Image

With the rapid development of social economy, the urban road traffic problems become more and more prominent. To managing and planning road transportation network scientifically by using intelligent transportation systems becomes a hot issue,and one of the key problem is how to get traffic information of region quickly.Using high resolution satellite imagery for getting vehicle information, as a new technology,has the incomparable advantage over conventional ground testing equipment.However, the automatic identification method of vehicle information from high resolution satellite images is not mature enough.To this, this article aiming at the high resolution satellite images, studied the method of identification vehicle target from the road image.In this paper, we use the cascade AdaBoost classifier to realize the recognition of vehicle targets.The training process of the classifier is: first, collecting a large number of sample images of vehicles and non-vehicles, with vehicle sample images take its symmetrical images. After normalizing all samples’ size and gray, extract the gray level and saturation information, and use Haar-like features to get the feature values,then take into the cascade classifier for training.Target recognition process is:,calculate integral figure on the target image’s grayscale images and saturation images respectively, gradually enlarge the detection window, normalize gray within the window, and then detect target and combine result.According to the characteristic that a plan view of the vehicle image are symmetrical, Two type of features were constructed in this paper: one kind is only testing half of the vehicle body, which reduce the height of detection window to half,then detect object by using all the features in the reduced window; Another kind is to detect body symmetry characteristics, in the original detection window, only use the features that are symmetrical about the symmetry axis of detection window, and describe the upper and lower part difference. Using the method, can significantly reduce the amount of characteristics, improve the training speed, while maintaining the recognition effect.Unlike most of the research uses only gray information, after contrast to several color space model, the paper add the saturation component of HSV color space so that the classifier’s selectable characteristic double the number. In the experiments foundthat the saturation component have a better ability of finding the dark vehicles on road.For different images have different lighting condition, shooting parameters and so cause the image brightness differences, this paper compares the histogram equalization, gamma correction and variance normalization. The experiment shows that using variance normalization can correct effect of light on the image well, while having a good detection effect.The traditional AdaBoost cascade classifier is adopted to train algorithm, and this method of classifier training has been able to identify the vehicle target well since it is optimized in the process of preprocessing and feature extraction. In practice, the study in this paper can be applied to data acquisition in intelligent transportation systems,and to guide the management of urban roads and future plans. It is of great application prospects.