Research On Real-time Detection Algorithm Of Athletes

In this thesis,based on the players in the sports field,the deep learning algorithm has improved,and the related algorithms for real-time detection of high-resolution and small targets are studied.The main research contents are as follows:In the part of image input,this paper proposes a new image segmentation and merging algorithm to process the input image,I divide the entire image of high resolution into many small tiles to improve the image resolution.By storing image block indexes to merge image tiles and using the IoU-like method to merge boxes which have higher than the threshold.At the same time,I try to use the improved Retinanet and single-stage Yolov2 network to compare the speed performance and explore new directions.Yolov2 can improve the speed with a slight decrease in accuracy.In the part of feature extraction,based on the Retinanet structure network,this paper intend to design Xception as a feature extraction network using depthwise and pointwise convolution.The k-means++clustering algorithm is used to calculate the size of players in the data set.Instead of the manual extraction of candidate boxes using RPN,the number of convolutional network layers is reduced.Optimizing the basic network framework can ensure the speed is improved without reducing the accuracy.The image width and height clustered by k-means++can help us set the aspect ratio of the anchor and choose feature map in the feature pyramid structure,which are selected to extract candidate frames,so that accurate positioning can be achieved,and the FCN layer can be reduced to improve the speed.In the part of loss function,this paper proposes to change the loss function in Yolov3,adding the Reploss loss,Reploss can enhance the constraint on the candidates,so that the candidate for a certain detection object is closer to its real target,is more farther away from other candidates and other real targets,by compressing the feature extract layer of Yolov3,writing I realize real-time detection of high-resolution small targets,which has high accurate and recall.