Research On Aerial Photographing Ground Target Detection Algorithm

Target detection from aerial photographing is one of the key technologies in the field of Computer Vision. It is a fusion of high-tech issues in computer vision, pattern recognition, image processing, automatic control and artificial intelligence and other advanced achievements in different fields, also the basis of intelligent monitoring systems, intelligent robot vision navigation and intelligent weapons. This paper focuses on moving target detection from aerial photographing, detection of aerial photographing targets in the city street under visible light conditions, and target detection based on infrared image sequences aerial photographing. This achievement is based on the research achievements of the predecessor researchers, which makes some improvements and innovations on object detection from aerial photographing. Generally speaking, this paper can be divided into the following three components:(1) Based on Hu moments and RM moments, propose a new moving target detection from aerial photographing algorithm ? HRM algorithm. The paper selects a set of image features of isotropic, and based on the framework of moment, extracts the Hu moment feature of seven dimensions and RM moment feature of second order 12-dimensional structure into a new image with a 19-dimensional feature. The ROC and real-time test results on ORIDS dataset show that the proposed method outperforms all state-of-the-art approaches. Moreover, the proposed model database updating can make the positive samples more abundant during the target detection process, which can improve the detection accuracy of SVM model. Experiments show that, with the accumulation of time and testing field, features library information even will get more perfect, and the detector for target detection from aerial photographing accuracy go higher.(2) Based on Fourier-HOG algorithm, this paper proposes urban streetscape target detection with the illumination normalization pretreatment from aerial photographing. This method improves the original Fourier-HOG algorithm at two aspects. Firstly, when the vehicle located in a shaded area has not high degree of detection with Fourier-HOG algorithm, the paper proposes the preprocessing algorithm based on LSH illumination normalization. Secondly, in order to reduce the calculation cost of algorithm, the paper proposes background filtering out based on the gradient information. Experimental analyses show that the improved Fourier-HOG algorithm on position accuracy has better performance compared with other four classical target detection algorithms.(3) This paper proposes a method of infrared target detection with image segmentation in the target detection by coarse and fine matches. This method has two main characteristics. At first, in the section of the coarse detecting for suspected target area, select four features as classification criteria, which are the area of the target, the average brightness, approximate elliptic short-long axis ratio, approximate ellipse filling ratio. Then, in the section of the fine target detection, the paper proposes the exact matching algorithms with the association in inter-frame-track. By matching the suspected targets from the front and rear frames by the two indexes, "appearance similarity" and "space position similarity", targets can be detected. Experimental results show that the algorithm has better performance in detection accuracy and real-time performance.