Research On Detection,Recognition And Tracking Of Channel Ships Under Dynamic Background Based On Deep Learning

Water transport is an important part of China's comprehensive transport system,and is increasingly showing its great role.In China's inland rivers,there are a large number of ships navigating daily,but some channels sometimes need to prohibit or restrict the passage of ships.For example the construction of bridges,the construction of cables across the channel and other time periods that need to control the navigation channels,so it is necessary to warn and even drive the passing ships as soon as possible to avoid accidents.In order to identify ships as early as possible to send out signals in advance,network cameras with long focal lenses are used to periodically scan and monitor the designated waters of the channel.However,due to the influence of surface ripple,wave caused by ship running,camera rotation and tremor,the traditional target detection algorithm has the problems of low accuracy and large amount of calculation in the process of ship detection.Aiming at the above problems,this paper develops a method based on deep learning to detect and recognize ship targets in dynamic background,which improves the detection result remarkably.The main work in this paper is summarized as follows:1)The structure,principle and efficiency of three typical deep learning-based target detection networks,Faster R-CNN,SSD and YOLOv3,are compared and optimized for YOLOv3 network.In the Ubuntu system,the Python programming language and the Tensorflow deep learning framework are used to build the network model,and then the ship dataset collected and labeled in the channel is used for training analysis.The results show that YOLOv3 has 10% lower average false detection rate than Faster R-CNN,faster reasoning speed and similar mean average precision,and 10% higher mean average precision than SSD,and 5% lower average false detection rate.A series of optimization training for YOLOv3 detection network were further carried out,which made the mean average precision reach 86.53%,the detection accuracy rate of small fishing vessels increased by 21%,and the average detection accuracy rate increased by 4.23%.2)An anti-jamming post-processing algorithm for the output of deep learning network is developed to detect and track ship targets accurately.Aiming at the prediction results of YOLOv3 network in continuous video frames,ship targets are matched by combining the centroid distance,IOU coincidence and color histogram correlation information of two adjacent frames.On this basis,a set of target tracking algorithm is designed.The test results show that the developed algorithm can accurately recognize and continuously track ship targets.3)Software development of ship detection system based on deep learning.Using the Hikvision spherical network camera(DS-2DC4420IW-D)with PTZ platform,the real-time video information flow is obtained through the network and the rotation of the PTZ platform is controlled.Under the C++ programming language and framework,the channel ship detection software under the dynamic background is developed by combining the Qt interface development library,OPENCV image processing library,Hikvision SDK library and Tensorflow deep learning library.The requirements of ship monitoring are also met.