Research On Behavior Recognition Of Agricultural Activities In Images With Multiple Features

With the development of information technology,agricultural information gradually turns to intelligence.In order to highlight the authenticity and timeliness of information,it is very common to record the production process of agricultural products in agricultural products traceability system.Agricultural production process record is an important part of agricultural product quality traceability,so it is particularly important to use agricultural production image to automatically identify agricultural activities.At present,there are two kinds of behavior recognition methods: traditional learning and deep learning,which have achieved good results for most simple human behaviors.Behavior recognition has a wide range of application scenarios,such as intelligent monitoring,human-computer interaction,etc.,but it has not been significantly studied and utilized in the field of agricultural applications.There will be different research focuses in different application scenarios.For the application scenarios of agricultural activities,behavior should pay more attention to the interaction information between people and things.With the increase of farming types,the differences of different behavior characteristics will be gradually obvious.Therefore,aiming at the application of behavior recognition in the field of agriculture,we propose a method of agricultural activity behavior recognition based on feature fusion,and proposes a behavior representation method of key behavior features for agricultural behavior.In order to make full use of the features in the static image and consider the hidden behavior structure in the image,we propose a method of agricultural activity recognition based on feature fusion by using the openpose model of human pose estimation and the yolov3 model of target detection.Because the key joint points in the feature fusion method are manually selected,with the increasing of agricultural behavior categories,the difference of key behavior representation will be gradually obvious.In order to improve the generalization ability of the model,a behavior representation method based on key behavior features is proposed,and the key features of each type of behavior are extracted.The main research contents and results of this article are as follows:(1)The display feature of agricultural activity image is constructed by human pose estimation and target detection technology,and the implicit feature is constructed by the association structure information implied by the explicit feature in agricultural activity image.A behavior recognition method based on image feature fusion is constructed by using the display feature and implicit feature.The results show that the recognition accuracy of multi feature is higher than that of single feature,which proves the effectiveness of the method.(2)The method of behavior recognition based on feature fusion selects the key joint points by hand.In view of the increasing agricultural activities,the fixed key joint points selected by hand can not represent the characteristics of more behaviors.Therefore,we propose a behavior representation method based on key behavior features.In this method,twin neural network is used to compare the similarity degree of each kind of bone frame graph,and the skeleton subgraph with the maximum average similarity is taken as the behavior representative of each behavior.The results show that the behavior recognition based on the characteristics of key joint points not only ensures the accuracy of behavior recognition,but also improves the generalization ability of behavior recognition method.