Research And Application Of Human Behavior Detection In River And Lake Area Based On YoloX

China’s rivers and lakes have a large number of resources and rich water resources,which have important economic and social value.However,in some reservoirs and river and lake waters located in remote areas,where it is impossible to arrange supervisors to be on duty 24 hours a day,a large number of people lose their lives every year because of swimming and fishing drowning.Deep learning-based target detection models in the field of river and lake area detection with high accuracy,high performance,wide range of applications and other advantages,gradually become one of the important means of water resources supervision industry.However,the existing vision-based target detection methods also cannot achieve real-time detection of illegal wild swimming and wild fishing behavior of people in river and lake areas due to the difficulty of both speed and accuracy,or the large size and difficulty of deployment.For the detection problem of wild swimming and fishing in river and lake areas,this paper intends to apply the target detection model based on deep learning in the field of detecting people’s behavior in river and lake areas,to provide a feasible solution for establishing an intelligent supervision system in river and lake waters,and to explore the accurate and scientific supervision means of prohibited swimming and fishing waters.Specifically,the YoloX-ECA model with attention mechanism is proposed for the detection of small and medium-sized target objects in river and lake areas,and the ECA module is added to the residual blocks and feature pyramid network in the YoloX backbone network to improve the detection of wild swimming and wild fishing in river and lake areas while maintaining the faster detection speed of the original YoloX model.The ECA method has a high accuracy and reliability for the detection of wild swimming and fishing in river and lake areas,and can provide a reference for research and practice in related fields.The experiments based on the homemade dataset prove that the improved YoloX-ECA model has a detection performance(AP)of over 90% for wild swimming and wild fishing behaviors in river and lake waters.The overall performance of the model(mAP)is improved by 1.21%compared with the original YoloX model,and the performance is also superior compared with other target detection algorithms such as Faster-RCNN.The improved YoloX-ECA model achieves the expected design goals of real-time and accuracy,and has a greater prospect of application in areas such as intelligent supervision of river and lake basins.Based on this,the YoloX-BiFPN algorithm is proposed to solve the problems of insufficient feature extraction capability and insufficient information interaction in the feature fusion network of YoloX network.The algorithm is improved in terms of backbone feature extraction network and feature fusion network.BiFPN is a bi-directional feature pyramid network,which can handle both contextual and detail information,thus better fusing semantic and location information.Therefore,applying the BiFPN structure to the feature fusion network can improve the adequacy of its location and semantic information interaction.The improved YoloX-BiFPN model significantly improves the accuracy of detecting and localizing the behavior of people in river and lake areas,with 8.72% improvement in the detection of people in proximity warning compared to the FPN feature fusion network,2.03%improvement compared to the PANet feature fusion network,and 1.31% improvement in the comprehensive performance compared to the YoloX model,which will help improve the safety of river and lake areas and management efficiency.In conclusion,the YoloX-BiFPN algorithm is an effective improvement scheme to improve the feature extraction and fusion capability of the YoloX network.