Research On The Technology And Application Of Navigate-route Environment Perception Based On Deep Learning

Waterway transportation is the foundation of national economy.As an important means of waterway transportation,the navigation safety of ships is particularly important.With the“Intelligent Ship Development Action Plan(2019-2021)”,“Made in China 2022” and other documents proposed,intelligent ships will be one of the important development directions in the future.Intelligent ships can help ship navigation and supervision by sensing the surrounding environment information autonomously.At present,the environmental information of ship navigation is mainly obtained through Marine radar(Marine radar),Electronic Chart Display and Information System(ECDIS),Ship Automatic Identification System(AIS),etc.With the improvement of computer operation efficiency,the perception algorithm based on computer vision and deep learning can be used as an effective technical scheme and means to obtain environmental information and provide a supplement for the above system.Therefore,this thesis uses deep learning technology to design a algorithm of the navigate-route environment perception,and develops a system of the navigate-route environment perception to serve the navigation and supervision of intelligent ships.The main research contents of this thesis are as follows:(1)Water surface object detection is the core of navigate-route environment perception.There are a variety of targets over water,including ships,navigational markers,bridges,and other obstacles.In order to ensure the safety of ships,an efficient detection algorithm is essential.The current performance Of SOTA(State-of-arts)object detection algorithm is analyzed,and the one-stage object detection algorithm YOLOv4,which takes both accuracy and speed into consideration,is selected as the baseline model of the detection algorithm in this paper.It is composed of convolution neural networks and has good performance in some public datasets.However,in order to make it more suitable for water surface object detection,it is improved.Firstly,the parameter reconstruction technology is used to optimize the backbone network of object detection,which not only improves the feature extraction ability of the model,but also speeds up the inference speed of the model.Secondly,the Feature Pyramid Networks(FPN)are improved by using the attention mechanism and the idea of cross-layer connection,which improves the multi-scale detection ability of the model and further reduces the number of parameters.Finally,the original Spatial Pyramid Pooling(SPP)layer is improved by using void convolution,which makes up for the information loss of the original Max Pooling layer and effectively expands the model’s receptive field.Finally,the detection algorithm of Ship YOLOV2 is proposed in this paper,and the experiment is performed on the open dataset WSODD(Water Surface Object Detection Dataset).Compared with the existing SOTA algorithm,Ship YOLOv2 has the best detection accuracy while guaranteeing the real-time inference speed.(2)Buoy is a special category among numerous aquatic targets.As important navigation aids,the navigation information can be obtained by object detection algorithm during the day.At night,its main manifestation is the navigation mark’s light quality,which is a video with dynamic information.Therefore,this paper proposes a novel multi-label video classification algorithm NMLNet to complete the identification of nighttime navigation mark’s light quality.First of all,for NMLNet,it adopts the double branch structure,and divides the input video frame into RGB format image and V channel(Value)format image as the input of the two branches.The RGB branch is used for color label identification of the navigation light,and the V-channel branch is used for flashing label identification of the navigation light.Secondly,we integrate the channel attention mechanism into the lightweight feature extraction algorithm Mobile Netv2 to complete the task of color label classification,and serve as the Backbone of blinking label classification.For flicker label,video classification algorithm based on CNN&LSTM is used to complete recognition.Finally,the method of binary relevance was used to fuse the multi-label recognition results according to the classification rules of navigation mark’s light quality.Through experiments on self-made nighttime navigation mark’s light quality data,it is proved that our proposed NMLNet can effectively complete the identification task of nighttime navigation mark’s light quality.(3)Based on the model of Ship YOLOv2 and NMLNet,a navigate-route environment sensing system is developed in this thesis,which can serve for intelligent ships.According to the functional requirements of the system,the back-end development of the system is completed by Flask framework based on B/S architecture,the front-end page design of the system is completed by Vue framework,and the construction of the system database is completed by My SQL.Tensor RT technology is used to accelerate the deep learning model and improve the operating efficiency of the system.The water surface object detection algorithm and night navigation mark’s light quality identification algorithm proposed in this paper realize the perception of water surface targets.The designed and developed navigate-route environment perception system can provide water surface object detection and video classification based on image and video.The research results can provide intelligent navigation of the ship with effective visual perception of the service,has application value.