Convolutional Neural Network Based On Object Detection For Vessel Traffic

Waterway transportation is an important mode of transport,which has beenplaying an important role in the trade between countries and regions.With in-depth strategy of implement of the China's marine power and "the Belt and Road",the water traffic has become increasingly busy.The ship characteristics of large scale,high speed and intelligence are more and more obvious.On the other hand,in the new round of industrial revolution,such as intelligent manufacturing,industrial 4.0,artificial intelligence 2.0,ship intellectualization has become a general trend.As one of the key technologies of intelligent navigation of ships,ship assisted driving technology has attracted wide attention.At present,ship assisted driving technology mainly relies on the modern navigation information system such as ship automatic identification system and radar,and the information is limited in water actual object detection.With the rapid development of deep learning theory and technology in recent years,the target detection technology based on computer vision can dig out the target characteristic information deeply,which greatly compensates the shortcomings of ship automatic identification system and radar.Carrying out the research and practice of water object detection based on the theory and technology of computer vision and deep learning has become the new direction and frontier of the development of navigation technology in this century.In this paper,a method of water object detection based on convolution neural network is proposed,and two deep convolution neural network models under different strategies are designed.The networks are trained and tested by self-collected images data set,and the results of the object detection were analyzed.The main research work in this paper includes:(1)Firstly,the structure,characteristics and object detection strategies of convolution neural network are analyzed.Two kinds of object detection methods based on convolution neural network and multi-scale improved models based on regional proposal network are reviewed.(2)Two different strategies are designed for deep convolution neural network basing on the idea of convolution neural network object detection technology.The first model is based on the deepening of the network layer strategy,using the latest depth residual network as the feature extraction network,through the connection design of the region proposal network and multi-task loss network,build end-to-end object detection model.Introduce batch normalization layer,Xavier weight initialization,dropout and other strategies to optimize the network.The second model is based on the fusion of multiple strategies.According to the object characteristics of the image dataset,different strategies are introduced in the network.VGG19 network is used as the feature extraction network,and the C.ReLU structure block,Inception structure block,multi-layer feature fusion and other strategies to optimize the network,according to the object characteristics of the image to optimize the region proposal network,and finally connect the multi-task loss network,build end-to-end object detection network.(3)The image data of various related ships were collected and carried out manual processing,then were input into the design models for training and testing.The best performance model based on the regional proposal network and the recursive method network were compared.Experiments show that the accuracy of the designed model is higher than that of the contrast model in the ship data set,especially the accuracy of the model two is much higher than the other three network models,and the result is state of art.