| The goal of human pose estimation is to determine the positions of body keypoints in images or videos that contain human instances,and then connect them together according to certain rules to form a skeletal structure.The resulting skeletal structure can partly represent the posture of the human body,providing a basis for further human motion recognition and behavior analysis,and has a high research value.In reality,the scenes in which human bodies are located are often complex,with limbs and keypoints frequently being occluded by objects or by the human body itself,which greatly affects the effectiveness of human pose estimation in practice,as well as the final accuracy.This article mainly studies 2D human pose estimation methods,using a top-down approach to estimate human pose from input images.The original backbone network has poor detection performance for some occluded keypoints and small target human keypoints.We retain the advantage of the original network’s high-resolution feature map output and add an attention mechanism module to the backbone network to reweight the output feature map,enhancing the network’s ability to recognize and extract keypoints of different scales.We also use larger convolution kernels in the module to increase the model’s effective receptive field,and strengthen the network’s ability to detect occluded keypoints,enabling the model to detect more human keypoints.On the dataset,we increase the diversity of the dataset through image augmentation methods,enhance the ability of the model to extract spatial and semantic information from images,and improve the final accuracy.The model was trained and tested on both the COCO dataset and the MPII dataset,proving that it can detect keypoints that the original network did not detect or detected incorrectly,and the final accuracy was improved by 2.6%. |
