Research On Human Pose Detection Methods Under Teaching Environment

Human pose information under the teaching environment plays an important role in teaching management and teaching evaluation tasks.Compared with the current routine,using manual inquiry,computer vision technology can obtain more comprehensive human pose information at a lower cost.However,human pose detection is still a difficult task due to heavy occlusion,large-scale variation of pose objects,and category imbalance under the educational environment.Considering the balance of speed and precision,this thesis conducts a study on pose detection under the teaching environment by using deep learning technology.The main works are as follows:1.A fast human pose detection method based on the adaptive receptive field is proposed in this thesis.To improve the detection performance under heavy occlusion and large-scale variation of pose objects,this method designs a convolution module with an adaptive receptive field.This module fuses the features of different adaptive receptive fields nonlinearly and learns the receptive fields adaptively by backpropagation.By integrating the convolution module with the adaptive receptive field into the SSD network further,the receptive field of shallow predicted layers could be adjusted.The receptive field of the whole network could be adapted to the scale variations of pose objects.Experimental results show that this method can effectively improve the pose detection performance under the condition of fast detection speed.2.A one-stage pose detection method based on multi-scale features and adaptive positive sample learning is proposed in this thesis.This method is further studied on multiscale learning under the teaching environment.Firstly,a multi-scale feature enrichment branch is proposed based on the RFBNet network.This branch learns enriched features by a multi-scale contextual enrichment module and instills those features into the predicted layers of different scales.Then,an adaptive fusion mechanism is adopted to fuse complementary low-level and high-level layers.The adaptive fusion mechanism makes the network more discriminative while facing category quantity imbalance.Besides,this method studies the importance of positive samples in pose detection under teaching environment and then proposes an adaptive positive sample training strategy.This strategy can select more robust positive samples and can make full use of high-quality predicted positive samples during the gradient descent process.Experimental results show that the proposed method has advantages over the existing methods in both computation speed and detection accuracy.