Research On Keypoints Detection And Data Application Based On Deep Learning

Owing to the improvement of computer performance and the application of convolutional neural networks,now more researchers are beginning to pay attention to computer vision.The key points detection task hopes to extract the features of the static images,so as to infer the posture and movement of the human body.This is the basis of other tasks such as motion recognition and motion segmentation.As the depth of neural networks increases,how to reduce the huge amount of network parameters has become a hot issue.This paper focuses on the key points detection task,designing and implementing a lightweight improvement.Knowledge distillation is added to realize knowledge transfer.Experiments prove that the amount of parameters is effectively reduced.Based on the key points detection task,this paper designs networks of action recognition and action segmention,whose accuracy has been verified through experiments.Finally,in order to achieve the goal of user-friendliness and data service,this paper studies the visualization technology,and gets visualization results based on the algorithm.Main tasks as follows:1.In view of the problem of huge amount of parameters in key points detection,this paper proposes two methods for reducing the parameter quantity.The network is designed with the Hourglass as the backbone and the knowledge distillation is added.The improved network structure effectively reduces the number of parameters while maintaining accuracy.Based on this structure,an action recognition network with LSTM unit as the core is designed.Experiments show that the network works well.2.Aiming at the problem of extracting a certain action in the video,a video action segmentation model is designed.The model includes three parts.Firstly,the keypoints detection framework can obtain all the keypoints information.Then the segmentation module with the DTW algorithm as the core is proposed,the distance function is designed,and the iterative operation using the sliding time window is finalized.Finally get the time period of the action.3.For user-friendly issues and the purpose of data service,this paper studies visualization technologies,including the OpenCV,visualization frameworks Vue.js and Echarts commonly used in the front-end field.This paper designs and implements visualization schemes.The final visualization results are intuitive and beautiful.At last,the paper summarizes the work and outlook the follow-up work.