| As one of the important research directions in the field of computer vision,moving object detection and tracking technology has been widely used in many fields such as intelligent monitoring,traffic system and human-computer interaction.However,in the actual application scene,there are many factors such as dynamic background,illumination change,occlusion,scale change and fast motion,which bring great difficulties and challenges to the detection and tracking of moving objects.In this paper,the elimination method of ghost and shadow,feature fusion method and model update strategy are studied.The specific research work is as follows:(1)Aiming at the problem of ghost and shadow in the process of moving target detection with Visual Background Extractor(ViBe) algorithm,an improved ViBe moving target detection algorithm was proposed to eliminate ghost and shadow.In order to solve the ghost problem,a method was proposed to detect and eliminate the ghost area by calculating the similarity of foreground and neighborhood background pixel gray histogram based on Euclidean distance and Tanimoto coefficient,and the pixel value of ghost area was used to update the background model to speed up the ghost elimination in the subsequent frame detection process.In order to solve the shadow problem,the shadow area is determined in YCbCr color space,and then the mixed Gaussian shadow model is established to further detect and eliminate the shadow.The simulation results show that the improved algorithm can effectively and quickly eliminate the ghost and shadow problems existing in ViBe algorithm in moving target detection.The average detection speed of the algorithm is 29.8 FPS,which can meet the requirements of real-time tracking.(2)Aiming at the fact that Kernelized Correlation Filter(KCF)algorithm could not adequately represent the target by a single feature,and that the model update method of linear interpolation could easily introduce the wrong target representation information into the model,a feature fusion model adaptive Kernelized Correlation Filter target tracking algorithm was proposed.The algorithm adaptively fuses the feature response graph with the Histogram of Oriented Gradient(HOG)and Color Names(CN)according to two indexes,namely,the product of the peak side lobe ratio and the response difference between adjacent frames.In order to solve the problem of linear interpolation model updating method,an adaptive model updating mechanism is proposed,which determines the update state of the model according to the target tracking confidence,and then adaptively updates the model.The simulation results on OTB-2013 data set show that the tracking accuracy and success rate of the improved algorithm are greatly improved,the tracking accuracy is 0.822 and the success rate is 0.596.The average tracking speed of the algorithm is 28.6 FPS,which can meet the requirements of real-time tracking.(3)Aiming at the problem that Siam RPN algorithm is easily disturbed by similar objects and lacks model online update,a Siam RPN target tracking algorithm combining attention mechanism and model adaptation is proposed.In the process of off-line training,the algorithm increases the positive sample pairs and negative sample pairs in the training samples,and introduces the feature space attention mechanism to improve the discrimination ability of the training model to the moving target and the similar objects in the background.In the process of online tracking,an online adaptive updating mechanism is proposed,which calculates the cosine value of the target template eigenvector and the tracking target eigenvector as the tracking confidence index to judge the current frame model updating state,and then the model is updated online.The simulation results on OTB data set show that the improved algorithm greatly improves the comprehensive performance of the algorithm,the tracking accuracy is 0.891 and the success rate is 0.670.The average tracking speed of the algorithm is 36.8 FPS,which can meet the requirements of real-time tracking. |
PDF Full Text Request