HRNet deep learning-based posture and behavior analysis module and analysis method

Abstract

The invention discloses a HRNet deep learning-based posture and behavior analysis module and analysis method, and relates to the technical field of human body posture and behavior analysis. The invention comprises a video stream acquisition module, a human body detection module, a human body picture processing module, a human body key point detection module, a key point information processing module and a posture and behavior analysis module; the video stream acquisition module is used for acquiring a video stream; the human body detection module is used for performing human body detection on an image in the video stream; and the human body picture processing module is used for cutting the human body picture meeting the requirement. According to the method, a deep learning network model adopts a repeated multi-scale fusion mode, high-resolution to low-resolution subnets are connected in parallel, and high-resolution representation is improved by utilizing low-resolution representation with the same depth and similar level, so that the high-resolution representation is also sufficient for attitude estimation, a predicted heat map is ensured to be more accurate in space, and the accuracy, robustness and adaptability of the human body key point detection algorithm are improved.

Description

technical field [0001] The invention relates to the technical field of human body posture and behavior analysis, in particular to a posture and behavior analysis module and analysis method based on HRNet deep learning. Background technique [0002] With the development and application of artificial intelligence, the analysis of human body posture and behavior has received widespread attention. It has broad application prospects in medical, education, interrogation and other fields. The accuracy and speed of recognition directly affect the follow-up video analysis system. , the current behavior analysis is mainly to detect the position information of the human body's skeletal joint points by analyzing the target image or video, and identify the human body's actions through logical judgment. Therefore, how to efficiently and accurately detect the key points of the human body has become the top priority in the field of behavior analysis. heavy; [0003] The current human body ...

AI Technical Summary

Problems solved by technology

[0003] The current human body key point detection method is mainly based on the classification convolutional neural network structure, by introducing upsampling operations and / or combining dilated convolutions to reduce the number of downsamplings to improve the resolution of the representation, and then to compensate for the loss of spatial accuracy. , but the existing technology has the following problems: In this type of network structure, the final high-resolution representation mainly comes from two parts: the first is the original high-resolution representation, but due to only a small amount of convolution operations, It can only provide low-level semantic expression; the second is the high-resolution representation obtained through upsampling of low-resolution representations. Although it has a good semantic expression ability, upsampling itself cannot completely compensate for spatial resolution. loss of rate

Therefore, the spatial sensitivity of the final output high-resolution representation is not high, which is largely limited by the resolution corresponding to the representation with strong semantic expressive power.

Images