Traditional algorithms have advantages such as interpretability and portability in pose estimation task. However, in complex background environments, traditional algorithms suffer from poor adaptability and detection errors. When dealing with complex scenes or small targets, CNN-based algorithms exhibit superior accuracy compared with traditional algorithms. However, CNN-based algorithms of pose estimation cannot be further developed on mobile terminals due to the large number of model parameters. To address this problem, this paper proposes the DBayC algorithm. First, the LBN (Limb Behavior Network) module is designed based on the CNN (convolutional neural network) algorithm to achieve the semantic segmentation effect on the human body. Then, the node annotation of human body is performed on the semantic segmentation results from LBN module to form graph-structured data. Finally, Bayesian formula is used to perform conditional probability analysis on the nodes in the graph, and the motion trajectories between nodes are analyzed, thereby achieving pose estimation and behavior analysis. Through the training of two data sets Hi-Eve and PoseTrack2017, and comparison with some SOTAs (state of the art) models. The experimental results show that under Hi-Eve data, DBayC achieved an accuracy of 79.2%, which is 3.8% higher than HRNetV2. Under the PoseTrack2017 data set, the DBayC algorithm achieved an accuracy of 78.6%, 6.9% higher than HRNetV2. It can be concluded that not only the accuracy of the DBayC algorithm has been improved, but the portability of the algorithm has also been improved, so the DBayC algorithm has certain use value.
Due to the mechanism of pooling and convolutional layers, many important features and the correlation between the features are lost in the forward propagation process in the pixel-level semantic segmentation tasks. Therefore, here we analyze the edge features of the image by means of second-order difference, propose gradient features and design the corresponding gradient convolution layer. Based on the gradient convolution layer, we use the residual structure to achieve the fusion of high-resolution gradient features and low-resolution gradient features. Finally, we designed the GraDNet. In the tests on the Cityscapes and ADE20K datasets, GraDNet achieves the best results in both accuracy and speed compared to some SOTA algorithms.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.