Research On The Detection Method Of Pontoon Overboard Personnel

In recent years,barges have been widely used as a kind of vessel for ships to dock and get on and off tourists.There are frequent incidents of people falling into the water on barges every year.Therefore,it is necessary to quickly and accurately detect Overboard personnel.This thesis focuses on the problem of pontoon drowning personnel detection,that is,by arranging cameras around the pontoon to collect real-time video of the surrounding waters,according to target detection and image segmentation algorithms to detect and analyze whether people have fallen into the water in real time,and design the drowning personnel.The detection water surface monitoring system is used to realize the task of detecting and alarming people falling into the water.The main work of this thesis is as follows:First,in order to solve the problem of inaccurate human detection and low detection speed caused by complex model calculation in the task of drowning person detection,three improvements have been made to the YOLOv4 algorithm: 1)The FMish activation function is used to replace the Leaky-Re LU activation function and The Mish activation function avoids the problem of oversaturation,ensures the stability during the training process,and improves the overall accuracy of the model.2)Adding the SE attention mechanism to the middle part of the YOLOv4 backbone network can enhance the network’s ability to learn important channels and improve the recognition rate of small targets.3)Using the method of image similarity detection,the images with high similarity in the video stream are ignored,and the speed and effect of video target detection are improved.Finally,an experimental platform is built,all the improved parts of the improved YOLOv4 are used for ablation experiments,and the improved YOLOv4 model is compared with other models.The results verify that the detection effect of the model is relatively good.Secondly,in order to ensure that the video images captured by the cameras arranged around the barge can be accurately and quickly segmented,certain improvements have been made on the basis of the DeepLabV3+ model.The specific improvements are: 1)In this thesis,the Si LU activation function is used instead of the Re LU activation function as the The activation function of the shallow network of Mobile V3 avoids the gradient explosion as much as possible during the training process,making the training more stable.2)Use the optimized Mobile V3 network to replace the Xception network as the backbone network,which makes the calculation of the segmentation model much smaller than the original Deeplab V3+ model while slightly reducing the segmentation accuracy.An experimental platform is built,and the improved DeepLabV3+ model is compared with other models.The results verify that the model has a good segmentation effect while greatly reducing the amount of calculation.Finally,based on Qt,a drowning person detection water surface monitoring system is designed,which mainly realizes the following functions: 1)Recognize the people appearing in the waters around the barge;2)Perform image segmentation on the images of the recognized people,which can determine whether the person falls into the water;3)The main control room transmits the personnel falling into the water signal to the alarm module for alarming.The system is tested through multiple videos,and the test results are all good.