Research On Body Pose Estimation Of White-spotted Bamboo Shark Based On Semantic Component Model

Video analysis technology is an important technical means to study the pose,action and behavior of human and animal bodies.The pose estimation is the key link of action representation and recognition,which has important research significance.This paper uses video analysis technology to study the body pose estimation algorithm of White-spotted Bamboo Shark under the aquaculture monitoring video,and realizes the migration from the body pose estimation to the fish body pose estimation based on the component model method in the study of human body pose estimation.Pose estimation is the basis of video behavior recognition and analysis.Based on the Flexible mixtures-of-parts Model(FMP),this paper proposes a semantic component model to model the White-spotted Bamboo Shark body and solve its pose to capture accurate keypoints information of the fish body.This method focuses on semantic components instead of traditional joints.A hybrid filter template is used for each part of the fish body to capture the HOG features of the local appearance,and the part relationship is spatially coded to determine the co-occurrence relationship between each part.In view of the high flexibility of the fish tail and the subtle contour characteristics,various types of mixed templates are discriminatively trained,and a unique bark shark tree structure is designed.This tree structure indicates the information of the score function.Delivery direction.In view of the increased complexity and increased inference time caused by the increase of hybrid components,a pose estimation algorithm using image segmentation to reduce the search area is proposed.This method first obtains the foreground area of the fish body and performs subsequent pose estimation in this area.Finally,the trained hybrid part represents the pose of the White-spotted Bamboo Shark body.Experiments have proved that the average PCK of each key point detection of the fish body posture based on the semantic component model is 87.2%(scale factora=1.0),which has improved accuracy compared with the FMP model,and the fish body posture can be inferred in a shorter time.This method can further improve the accuracy and greatly improve the algorithm speed.And using the detected keypoints of the pose can effectively express and describe the target action.Finally,the posture parameters of the fish body are analyzed,including type,position,and angle parameters.The pose and action are parameterized using the obtained fish body skeleton information,and combined with the key point information of the pose,a pose classification method based on SVM is proposed.Using the distance between pectoral fins and pelvic fins to jointly construct a two-dimensional feature vector,the classification accuracy of left and right turns reaches 98.86%,which proves the effectiveness and accuracy of the fish body pose estimation and recognition method proposed in this paper.