Multiple Object Detection And Tracking Method On Water Surface Based On Adaptive Spatial Feature Fusion

China is a big country in the world and has a vast maritime territory.It is the common goal of our researchers in the field of ships and oceans to continuously improve the guarantee ability of maritime territory.Ship intelligence is an inevitable trend and an interdisciplinary and interdisciplinary research with large investment and long cycle.Therefore,the research on ship intelligence has far-reaching significance.This paper takes the environmental information required by intelligent ships as the starting point,and studies the water boundary detection method based on vision,the water surface object detection method based on vision,the water surface multiple object tracking method based on vision and the application of deep neural network in this field.The main research contents of this paper are as follows:Firstly,the research status of water surface multiple object detection and tracking at home and abroad is investigated,the advantages and disadvantages of mainstream water boundary detection methods,water surface object detection methods and water surface multiple object tracking methods are analyzed,and the semantic segmentation network based on encoderdecoder,object detection network based on key point estimation and multiple object tracking network based on pyramid aggregation are built respectively.Secondly,aiming at the difficulty of semantic segmentation in water boundary detection,this paper introduces channel attention to deal with the loss of feature map caused by channels of different importance in the process of convolution pooling;In order to solve the problem of detection failure caused by the change of object shape and scale in the process of water surface object detection,deformable convolution is introduced to improve the feature extraction ability of water surface objects with large change of shape and scale;In order to solve the problem of tracking failure caused by mutual occlusion of objects in the process of water surface multiple object tracking,this paper introduces adaptive spatial feature fusion to improve the ability of spatial feature expression between multiple mutually occluded water surface objects.Thirdly,collect and label the water surface semantic segmentation data set,water surface object detection data set and water surface multiple object tracking data set required by the depth neural network,and demonstrate the network performance.The results show that the improved network can more effectively segment the water surface area,detect the water boundary line,and detect all kinds of objects in the water surface environment,It can more stably track multiple moving objects in water environment.Fourthly,a simulation environment of water surface visual perception method based on virtual engine is built,and several groups of water surface scenes with different weather,different time and different wind power levels are arranged.The simulation results show that the improved network has better robustness,has higher water surface area segmentation in complex water surface environment,and the water boundary is closer to the real position,It can more accurately identify the category and location information of water surface objects,reduce the probability of false detection and missed detection,reduce the number of loss of water surface object ID in the tracking process,improve the re identification ability,and meet the design requirements of this supporting project for the perception of water surface environment.