Multi-person Pose Estimation Algorithm Based On SSD-Hourglass And Its Optimization

Object detection is literally a significant and challenging subject in the pattern recognition in the computer vision.The Multi-person pose estimation is an important branch in the object detection,which is,detecting a figure with multi-person object in each person pose,majoring in intelligent monitoring,behavior recognition and human-computer interaction,etc.Recently,with the mature development of deep learning technology in the computer vision area.The Multi-person pose estimation based on deep learning is applied widely nowadays.This paper proposes the background and significance of the human pose estimation,with an analysis of the research status from both domestic and overseas.We focus on the Multi-person pose estimation with the existing issues and difficulties,also illustrate the elementary definitions of Deep Learning and Reinforcement Learning with their application in Multi-person pose estimation.According to the existing problems in the project,two aspects are implemented as following:(1)The introduction of the SSD-Hourglass Multi-person pose estimationThis paper proposes a two-step estimation algorithm named SSD-Hourglass which belongs to the Multi-person pose estimation.Firstly we use SSD as the object detector to detect the target human in pictures with multiple bodies and identify them with bounding boxes,afterwards we clip and pad the identified bounding boxes to attain the single human picture and resize it as 256×256.We detect the pose of all the single human pictures through the Stacked Hourglass Network as the pose detector.Finally we reflect all the articulation points obtained in the pose estimation to the original images.This algorithm has trained and detected,which has achieved 72.1mAP on the MPII dataset.(2)Improved SSD-Hourglass Multi-person pose estimation based on the Reinforcement LearningThe accuracy of two-step detection algorithm is affected due to the inefficient fit degree of human body that caused by the regressive bounding boxes of object detector SSD.Aiming to the issue,this paper came up with an Object Refine Model based on a Reinforcement Learning.The Object Refine Model is composed of the base network consisting of VGG-16 and the full-connected layer consisting of Q network.Aiming at the inefficient fit degree of bounding boxes,we use Markov Decision Processes and Q network to adjust translationally from eight directions:up,down,left and right at two corners,upper left and upper right.The terminating action stops by the model when the bounding boxes close to the human body adequately.This paper also modifies the original object refine model via replacing the VGG-16 network by DenseNet due to its insufficient.Eventually,the improved algorithm has approved on MP? dataset,the result has risen to 73.7mAP,with an increment of 1.6mAP compared with the SSD-Hourglass Multi-person pose estimation.