Research On Pose Detection In Teaching Scenes Based On Meta-learning Methods

In teaching scenes,the students’ poses are of great significance to evaluate students’ learning state and the quality of course.With the development of intelligent information technology,more and more researchers begin to study how to use deep neural network to detect poses automatically.However,the current relevant researches ignore the problem of domain shift in different classroom scenes,and the detected human pose categories are relatively few and fixed.1.This thesis proposes a cross-domain pose detection method based on double-layer gradient optimization.In this method,a pose detection meta-model and a domain adaptive optimizer with learnable parameters are designed.Besides,the offline learning mode and online learning mode are combined to realize the fast domain adaptation of the detection model in a specific classroom scene.In the offline learning stage,the proposed external optimizer trains the parameters of the pose detection meta-model and the adaptive domain optimizer in double-layer training.In the online learning stage,guided by the adaptive domain optimizer,the meta-model can quickly adapt to the data distribution of the scene with a few labeled images.Experiments show that the detection accuracy of this method on basic scenes of cross domain classroom pose detection dataset is 90.91%,and it also has good domain adaptation effect for new scenes with a few labeled images.2.This thesis proposes a multi-category pose detection method based on new class metrics.This method adopts class metrics to distinguish the new poses.Firstly,the new pose metric module is trained based on the multi-category pose detection data set and the regional recommendations of the basic pose detection network,so that it has the ability to distinguish the new poses from the three basic poses.Then,the images that may contain human objects extracted from the basic pose detection network are input into the new pose metric module to determine whether they belong to the new pose categories.Finally,the new pose discrimination results and the basic pose detection results are combined to realize the comprehensive detection of the new poses and the basic poses.Experiments show that the comprehensive detection results of the data with one or two new poses on the multicategory pose detection data set are 81.71% and 79.82% respectively,and also demonstrate that proposed method have a certain detection ability for the new pose that have not appeared in the training data.