Detection And State Analysis Of Special Targets In Dynamic Background

With the rapid development of machine vision,the use of cameras to capture video and images,and the use of image processing means of object detection and state analysis has become increasingly universal.Existing object detection algorithms have relatively mature and stable application,but there is no corresponding effective detection method for special target in dynamic backgrounds in certain application scenarios.Therefore,this thesis takes the special target such as tiny object and moving slender object in dynamic backgrounds as research object,fully analyzes the limitations of existing object detection algorithms in the face of special target detection,proposes the tiny object detection algorithm based on SPD-YOLOv5+ and the moving object detection method based on random gray fluctuation field to solve the two problems of video target-scoring and galloping detection of transmission line in dynamic backgrounds,and achieves the accuracy of bullet hole detection is 98.3% and the accuracy of galloping detection of transmission line is 95.5%.In addition,around the special target detection algorithm,this thesis also investigates the unsupervised saliency object detection algorithm and the optimization of detection model parameters based on the swarm intelligence optimization algorithm.The main works in this thesis are as follows:(1)In the study of tiny object detection,the SPD-YOLOv5+ object detection algorithm is proposed and used to solve the video target-scoring problem in dynamic background,achieving98.3% accuracy rate of bullet hole detection,and then proposing an effective and complete video target-scoring method.For the problem of insufficient YOLOv5 s tiny object detection capability,it is enhanced by adding a separate tiny object detection layer and applying the SPD-Conv module.The interference of dynamic backgrounds such as swaying trees on bullet hole detection is avoided by using template matching and tracking the target surface.For state analysis of bullet holes,Deep SORT tracking is used to distinguish old and new bullet holes,and then the analysis of the number of rings,angles,etc.of bullet holes is realized according to the design rules of bullseye,target ring and bullet hole coordinates.(2)In the moving slender object detection research,the moving object detection method based on random gray fluctuation field and the galloping detection of transmission line method based on cumulative gray model are proposed to achieve a 95.5% galloping detection accuracy.For the scenes with slow or stationary background changes captured by fixed cameras,the random gray fluctuation field is proposed to improve the plural background difference method to achieve the detection of moving objects and for the detection of transmission lines.For the transmission line is always located in the sky region,the sky region is segmented according to the texture information,and then the interference of the ground dynamic background on the transmission line detection is excluded.By dissecting the dancing nature of transmission lines,the accumulation of transmission line positions over a short period of time can reflect their movement range on a single image,and then determine the galloping status.(3)In the study of saliency object detection,we propose a detection method based on graph cut refinement and differentiable clustering.The method gradually improves the accuracy of saliency detection from "coarse" to "fine",and achieves accurate detection with only a single image,obtaining 14.3% and 23.4% MAE on the ECSSD and SOD datasets constructed under natural dynamic scenes,respectively.(4)A framework for optimizing detection models based on swarm intelligence optimization algorithms.The hyperparameter setting of swarm intelligence optimization algorithm is considered,and it is proposed to use the radius distribution of the suction basin and the local optimal number to guide it,so as to improve the convergence speed of swarm intelligence optimization algorithm and enable it to be applied to the parameter optimization of the detection model.In this thesis,the two core points of the radius distribution of the attraction basin and the number of local optima is thoroughly investigated,and the method of evaluating the radius distribution of the attraction basin based on Euclidean distance and the method of estimating the number of local optima based on resampling are proposed.