Human Pose Estimation In Low-light Conditions Based On Improved Feature Fusion Strategy

With the development of artificial intelligence technology,the traditional industry ushered in an information,intelligent reform.Among them,intelligent security,driverless and intelligent medical fields are particularly concerned by people.These industries are closely related to people's daily life,which can greatly improve people's quality of life,and efficient and accurate human posture estimation technology is the core of human-computer interaction in these fields.However,due to the defects of algorithms and hardware devices,the complex real scene will seriously affect the detection accuracy of human keypoints,which limits the development of the industry.Therefore,it is of great significance for the future social development to conduct in-depth research on human pose estimation technology and improve its detection ability in complex environment.In this thesis,the problem of human pose estimation in low-light environment is studied.The multi-scale feature fusion strategy is improved.Combined with low-light image processing technology,the detection accuracy of keypoints of human body in low-light image is improved,and a low-light human pose estimation system is designed.The main work of this paper is as follows:(1)This thesis propose a improved I-Fpn(Improved feature pyramid networks)structure.Aiming at the problems of large feature loss and low utilization rate in multi-scale feature fusion of human pose estimation network.In this structure,the residual enhancement method is used to supplement the top-level features extracted from the network,which improves the human key point classification ability of the network model.At the same time,the adaptively feature fusion strategy is adopted to further process the extracted feature graph,and improves the feature representation ability of the network.Experiments show that the CPN human pose estimation network model using I-FPN structrue achieves 73.3% mAP value on Mscoco2017 dataset,which is 1.2% higher than the original network.(2)This thesis proposes a new low-light image processing network.Aiming at the problems of low brightness,color distortion and excessive noise when taking human body pictures in low-light environment.The network establishes the mapping relationship between the low-light image and the enhanced image.By adding adaptive parameters,it supervises the feature fusion process of the network.This network change the strategy of feature fusion by adding adaptively parameters,and add the deformable convolution to increase the receptive field of the feature map.At the same time,it uses deformable convolution to increase the receptive field of the feature image,which enhances the feature extraction ability of the network.Experiments show that the processing ability of the network is improved obviously,and the network structure is simple and robust.(3)This thesis designs a low-light human pose estimation system based on deep learning.The system consists of image acquisition module,image preprocessing module and human posture estimation module.The system collects the image data through the image acquisition module,and uses the image preprocessing module to improve the image quality,and then uses the human body posture estimation module to detect the coordinates of the key points of the human body for the use of subsequent modules.In order to verify the effect of the system,a new low-light human dataset is established.The dataset has human images and keypoints labels under low-light conditions.In this paper,based on the previous work,the algorithm module is retrained to improve the detection ability of low-light human keypoints.Experiments show that the algorithm module achieves 68.4%mAP value on the low-light human body dataset,which is 2.3% higher than the original network.