The Research On Moving Object Detection Algorithms

The detection of moving object in the field of computer vision research is one of its most important subjects.Its technical principles comprehensively use the scientific technologies of computer vision,pattern recognition and artificial intelligence,and its commercial value and application prospects are mainly reflected in military applications,system monitoring,intelligent transportation,product detection and other scenes.In this thesis,the moving object detection algorithm in complex environment is discussed.Firstly,several image preprocessing methods,such as filtering denoising and edge detection,are discussed.Then,several common moving object detection methods are compared between inter-frame difference method,Gaussian model and Vibe algorithm,and the superiority of Vibe algorithm is obtained by experiments.For the shadow and ghost influence of Vibe algorithm,the parameter optimization algorithm of Vibe and YUV_Vibe fusion algorithm is proposed.Finally,the application of deep learning method to target detection is further explored.The main contents of the thesis are as follows:1.The Research on video image preprocessing.Three kinds of filtering algorithms,namely mean filter,median filter and Gaussian filter,were studied.Roberts,Prewitt,Sobel and Laplacian edge detection operators are analyzed and compared,and several common morphological processing such as opening operation and closing operation were studied.2.The Research on moving object detection.Comparing the experimental results of the inter-frame difference method,Gaussian model and Vibe algorithm,it get the worst effect of the inter-frame difference method,followed by the Gaussian model,and the Vibe algorithm worked best.It can detect moving objects in complex environment accurately and real-time,which shows the superiority of Vibe algorithm.3.An improved YUV_Vibe fusion algorithm.Aiming at the defects of Vibe algorithm,such as being easy to be influenced by illumination shadow,slow ghost removal and so on,the parameter optimization algorithm of Vibe and YUV_Vibe fusion algorithm are proposed.At the same time,Vibe parameter optimization algorithm extends the range of sample values to 24 fields,which adjusts the update factor and updates two samples at the same time,optimizes the update rate of the model,and accelerates ghost elimination;YUV color feature information and Vibe algorithmprinciple were used to construct a fusion double model by YUV_Vibe fusion algorithm,eliminating the shadow effect.4.The Research on target detection based on deep learning method.The YOLO detection method and further optimized YOLOv2 method were introduced.Firstly,the theoretical basis of the two deep learning methods are described in detail,and then the performance of the two detection methods is compared and analyzed through experiments,which verified the better detection effect of the improved YOLOv2 method.The running speed of Vibe algorithm is very fast,it can eliminate the influence of noise,dynamic background and lens shaking,and has a good target detection effect.However,there are still some deficiencies in Vibe's algorithm,which is vulnerable to the influence of light shadow,appearing false detection,and arise ghost phenomenon for objects rotating by static.In order to solve these problems of Vibe algorithm,a new improved YUV_Vibe fusion algorithm is proposed in this thesis.The improved YUV_Vibe algorithm can improve the detection accuracy and the recognition rate of small targets,which show better robustness to noise and shadow interference,achieving the expected objectives.