Research Of Object Detection And Tracking Algorithm In Video Images

Computer vision is one of the hotspots in the field of artificial intelligence,and is the most challenging and research field.This paper focuses on three core issues in computer vision: image segmentation,image recognition and target tracking.The main work is as follows:Firstly,in image segmentation,the GrabCut image segmentation algorithm is improved.This paper analyses the traditional GrabCut,and proposes an improved image segmentation algorithm cascaded with Graph-based and GrabCut algorithm to overcome the shortcomings of the algorithm,such as ignoring the connection between the foreground and background pixels and causing undersegmentation.The basic idea of the improved algorithm is: firstly,the whole image is processed by Graph-based algorithm to get a rough classification of all the pixels;then,the classification information is used to modify the image label in the traditional GrabCut algorithm,and then image segmentation is completed.So the accuracy of GrabCut segmentation is improved.Experiments show that the improved cascade algorithm is better than the traditional GrabCut algorithm in segmentation accuracy,but because of the addition of Graph-based,the complexity of the algorithm has been improved.Secondly,an image target detection algorithm based on YOLO convolution neural network and Gauss mixture model(GMM)classification is proposed.This paper analyses the traditional YOLO convolution neural network image target detection algorithm,and proposes an improved YOLO convolution neural network image target detection algorithm to overcome the shortage of the number and type of candidate frames in the traditional algorithm,which results in the low accuracy of the algorithm.The basic idea of the algorithm is: in the training process of the traditional YOLO convolution neural network target detection algorithm,GMM is used to classify and predict all rectangular frames of targets,and the prediction candidate frame of targets is added,so as to improve the accuracy of target detection.Experiments show that the improved algorithm improves the accuracy of target detection,but the complexity of the algorithm is slightly increased due to the increase of the number of candidate boxes.Thirdly,a correlation filtering video target tracking algorithm based on multi-feature fusion is proposed.In the traditional correlation filter tracking algorithm,HOG feature is used as the recognition feature of the target,which is sensitive to the deformation of the target,does not have scale invariance and rotation invariance.In order to overcome this shortcoming,a correlation filtering target tracking algorithm based on HOG,LBP and CN features is proposed,and the optimal selection of fusion parameters in the algorithm is simulated and analyzed.The simulation results show that the tracking accuracy of the improved target tracking algorithm has been greatly improved,but the complexity of the algorithm has been slightly increased due to the increase of feature dimension.At the same time,it is found that the fusion parameters of the three features have a certain impact on the accuracy of recognition.Fourthly,based on the first three works of this paper,aiming at the application target of "Animal Motion Feature Video Monitoring",a target detection and tracking system of integrated detection and tracking in animal video image is programmed and tested by simulation.