Diver Pose Estimation Based On Deep Learning

Pose estimation is a very challenging subject in the field of computer vision.With the continuous development of related cutting-edge technologies such as deep learning,gesture recognition methods based on deep learning have greatly promoted the research in this field.Even so,there are still few mature and effective gesture recognition algorithms and data sets for underwater environments.Therefore,this article analyzes the current research status of gesture recognition technology at home and abroad,and based on the Openpose algorithm in the 2D multi-person pose estimation algorithm,combined with the characteristics of the underwater environment,the two parts of the algorithm are improved and improved.A set of diver’s posture recognition algorithms suitable for surface environment are proposed,and then experimental verification and analysis are carried out on related data sets.The specific research content is as follows:First of all,the research focus of this article is to use deep learning technology to solve the diver’s pose estimation problem.In order to facilitate subsequent research,according to the characteristics of typical irregular viewing angles and complex underwater environment occlusion in the underwater environment,this paper has produced the diver data set Under Water-2020 for the verification and verification of subsequent algorithm functions.Secondly,the freedom of movement of the human body is higher in the underwater environment,and the probability of body scale imbalance in the taken pictures is much higher than that in the terrestrial environment,which leads to the failure of conventional gesture recognition algorithms.In order to solve the problems caused by irregular viewing angles,that is,to obtain a model that has good accuracy and can also solve the viewing angle problems,this paper uses high-resolution networks as the basis to study its principles and learn from it on this basis.The feature pyramid network enables the network to handle features from different perspectives.Later,based on Openpose,the backbone network of the algorithm was replaced,and experiments were conducted on MSCOCO-2017 val and the diver data set Under Water-2020.The results show that the improved backbone network is very robust to pose recognition problems from complex perspectives.Great.Finally,in order to efficiently solve the problems of complex underwater equipment occlusion,self-occlusion,environmental occlusion,etc.,this paper improves the part of the Part Affinity Fields(PAF)in the second stage of the Openpose algorithm,combining the attention model and ergonomics The model proposes a Large Part Affinity Fields(LPAF),which replaces PAF.Later,on this basis,the hole convolution is introduced to increase the receptive field of the network,so that the network can capture more implicit body features,improve the accuracy of the network and reduce the parameters of the network.In addition,LPAF was pruned and the LPAF-B structure was proposed,which further improved the real-time performance of the network while meeting the accuracy requirements,and verified it on MSCOCO-2017 val and the diver data set Under Water-2020.The final experiment shows that the LPAF structure can effectively predict most of the occluded joint points.