Detection In Video Stream Research On Human Key Points

The detection technology of human key points has always been an important research topic in the field of computer vision and target detection.It has a wide range of applications and theoretical significance in the fields of business,security prevention,rehabilitation and medical.In the highly modernized information society,it is a significant problem how to detect key points of the human body from the video stream quickly and accurately,because of the different degrees of complexity and unstructured features in their scenes.We analyze the research status of human key points detection,and find that the detection based video stream is easily affected by external light condition,noise affection of video stream,etc.Moreover,some algorithms are not suitable for real-time detection because of the high computational complexity and poor robustness.In order to overcome the above problems,this thesis optimizes the key frame extraction based on video stream and the detection of human key points.The main work of this contents is as follows:1.A GMHKE algorithm is proposed,which uses the Gaussian kernel function combined with equal interval frame sampling to improve the motion history map to process the video stream data.It does not need to analyze each frame image,and can grayscale the motion history map MHI effectively.The value changes smoothly to ensure robust.The image features are extracted by HOG,and then the NN classifier is used to detect whether the action status tag is changed and the action key frame is extracted.Compared with original MHI,it can better capture the key information in the video stream and reduce the amount of calculation,it also has the higher calculation efficiency.2.A SN Mask R-CNN network model is proposed.The reassembly channel network of ShuffleNet is taken by Mask R-CNN network model,through using group point-to-point convolution and channel shuffle operations and combining the calculation results of the bounding regression with the mask representation.The model is lightweighted from the perspective of parameter weight reduction and computational weight reduction.Compared with the original Mask R-CNN model and the existing lightweight model mobile network on the cloud platform and low configuration machine,the experimental results are more quickly and accurately under the premise of the accuracy of the original network model.