Research On Multi-person Pose Estimation Based On Lightweight Multi-scale Neural Network

Human pose estimation refers to detecting human joint points in images or videos,and connecting the joint points to form a skeleton.This task is the basis for tasks such as skeleton-based behavior recognition and pedestrian re-recognition,and is widely used in security monitoring,movie production,interactive games,and other fields.The development of convolutional neural networks has basically solved the problem of human pose estimation in single-person scenes,but there is still much room for development in complex scenes with multiple people.In multi-person scenes,there are often problems such as person-to-person overlap and joint points being blocked,which causes the network to misidentify the positions of some joint points.Therefore,multi-person pose estimation is still the focus of research in the field of computer vision.In view of the problems existing in multi-person pose estimation,this paper focuses on two key issues: image feature extraction and neural network lightweight research.Details as follows:(1)In view of the problem of low accuracy of joint point recognition in multi-person scenes,this paper adopts a design idea from coarse to fine,connects an improved cascade pyramid network after a high-resolution network,and uses a multi-stage approach to further improve recognition accuracy,thus proposes a new network structure multi-stage improvement cascade pyramid network,also called MSIPN.MSIPN makes full use of image information of different scales,which can effectively improve the accuracy of pose estimation.The improvement to the cascade pyramid network is to replace the interpolation upsampling with transposed convolution upsampling so that the network upsampling operation can be learned.In order to make more effective use of feature information than previous multi-stage networks,this paper uses a cross-stage feature fusion mechanism between stages to combine the features of adjacent stages and generate different sizes ground truth heatmaps to achieve coarse to fine intermediate supervision.The experimental results on the complex multi-person pose estimation dataset COCO show that the sub-modules of the multi-stage improved cascade pyramid network proposed in this paper help to improve the detection accuracy in multi-person scene,and have achieved on the COCO dataset competitive results,while bring the cost of slightly increased complexity.(2)In view of the above-mentioned multi-stage improved cascade pyramid network calculation and parameter problems,this paper further conducts lightweight research on the network,and proposes to use an improved dense connection module and an improved Inception-Res Net module to replace the basic unit residual module of this network.Among them,the improved dense connection module directly connects all layers,strengthens feature transfer,reduces network complexity,and improves parameter utilization;The improved Inception-Res Net module utilizes the technique of decomposing the convolution kernel and increasing the parallel convolution layer to reduce the network complexity to a certain extent.The experimental results on the multi-person pose estimation dataset MPII show that the improved dense connection module can significantly reduce the model parameters and computational complexity,and also benefit the detection accuracy,while the improved Inception-Res Net module does not significantly improve the efficiency of the model,and it is equivalent to the improved dense connection module in terms of improving accuracy.Under comprehensive consideration,this paper selects the improved dense connection module as the basic unit of the lightweight multi-stage improved cascade pyramid network,also called LightMSIPN.