Human Pose Estimation Method Based On Multi-task Learning

With the development of deep learning and convolutional neural networks,the network model structure,the semantic information of feature extraction is stronger,the network is more robust,and it can perform end-to-end feature extraction and prediction,in the context of complex poses,occlusion scenes,and big data,pose estimation based on deep learning The method has become the mainstream method in academia as well as industry.The existing pose estimation task still has a single network prediction function,and the model is time-consuming,which is not conducive to product deployment and multi-tasking;therefore,based on the pose task,it is more important to perform multi-task learning based on joint detection and segmentation tasks,and how to optimize the pose at the same time.It is of great significance to estimate the computational load of the model and realize the fast and good learning of the pose estimation task.The main research contents and achievements of this paper are as follows:A pose estimation method for high-resolution representation is designed,which is based on the model architecture of Mask-RCNN,obtains each target instance,and then maps the detection results of the instance to the high-resolution features of the feature pyramid,and predicts the module through key points.Upsampling improves the feature resolution,and finally the predicted features are encoded into the spatial position of the classification task to obtain the two-dimensional spatial position of the key points for each instance.Comparing the designed high-resolution representation model with the existing multi-layer feature extraction algorithms,it can be seen from the experimental results that this method has higher accuracy than other methods.A pose estimation method based on multi-task learning is designed,which combines target detection,segmentation,and pose estimation into one model for learning;the method is based on the above-mentioned high-resolution representation pose prediction model,and then establishes instance segmentation prediction.Module,instance segmentation module adopts two types of prediction tasks of foreground and background,and obtains the segmentation information of each instance.Experiments show that the multi-task learning pose estimation method designed in this paper has higher performance than the single-task model and is more practical for functional tasks.Finally,based on the above multi-task pose estimation method,a lightweight and brief multi-task pose estimation model is designed.The main purpose of this method is to optimize the multi-task pose estimation model,and consider that the above multi task models are independent and have no information interaction,which is not conducive to promoting information learning among multi tasks.Therefore,this method first fuses the segmentation and attitude prediction module to extract the mutual information between them and reduce the amount of computation.Then,considering the contradiction between the dense foreground of segmentation and the sparse foreground of pose estimation,the segmentation and attitude are output from different stages.Finally,the channel number of pose estimation module is reduced,and the receptive field of attitude estimation is improved by convolution with atrous-convolution.The performance is guaranteed and the task of 2D pose estimation is faster.