Design And Application Of Lightweight Network For Pose Estimation

This paper mainly discusses the lightweight network design and practical application of human pose estimation.Human pose estimation is an important and difficult task in computer vision.The purpose of this task is to detect the human key points in pictures or videos and to analyze the motion of the key points.At present,people's requirements for quality of life are gradually improving,Human pose estimation can be applied in video monitoring,human-computer interaction,advanced driver assistance systems and other fields,which has important significance for intelligent life.At the same time,the method of human pose estimation is applied to the key point detection of rectangular objects in this paper,which can quickly locate the corners of rectangular objects such as business cards and id cards and improve the office efficiency of staff.Traditional human pose estimation and key points detection technology depends on more complex image processing technology and post-processing techniques,thus the inference speed and precision is low.With the rapid development of deep learning and computer vision,convolutional neural network has become the main development direction of human pose estimation and key point detection.Compared with traditional methods,it has more accurate results and more development possibilities.However,there are still some problems in pose estimation and key point detection based on deep learning.While focusing on accuracy,it is often easy to ignore the calculation amount and inference speed of the model,which makes the final network model run only on GPU and difficult to be applied to lightweight devices.Therefore,aiming at the above problems,this paper discusses the lightweight network design and practical application of pose estimation from the following two aspects:(1)This paper proposes a lightweight human pose estimation network,the network first selected the suitable lightweight backbone for feature extraction;then reasonably reduce the redundant feature processing module when processing feature maps,only an initial layer and a refinement layer are used to generate the final heatmaps of the key points;at the same time,the combination of several small convolutional layers is used to replace the large convolutional layer,and the operation of dilated convolution is added to improve the receptive field of the network,so as to make the human pose estimation network more lightweight and ensure the accuracy of key point detection.(2)In this paper,the human pose estimation algorithm in computer vision is applied to the key point detection of rectangular objects.In this network,the heatmaps of key points is obtained only through the lightweight feature extraction network and the feature pyramid layer which integrates multiple layers of semantic information.The design of the network not only reduces the computation and parameters,but also makes the final detection result more accurate due to the fusion of multi-layer information.In this paper,the human pose estimation experiments are conducted mainly on the MPII dataset,and compared with other pose estimation algorithm.The key point detection experiment of rectangular objects is mainly carried out on the mixed datasets of program synthesis data and real data.The experimental results show that the proposed method is feasible and effective.