Ship Tracking Under Complex Background Based On Deep Learning

With the continuous integration of computers into all aspects of people's lives,it is particularly important to record and protect people's security through video surveillance.Video object tracking is one of the most important research topics in the field of computer vision.At the same time,it is a very challenging topic,which is widely used in human-computer interaction,video surveillance,military weapons applications,etc.Object tracking is an important part of video surveillance.The process is to track the object in real time by marking the first frame of the video.By matching the similarity between the feature images of the front and back frames,and associating the information between the data,the moving trajectory of the object can be obtained continuously.A good track is usually accompanied by efficient and accurate tracking.However,in general,realistic scenes are complex and can cause deformation or occlusion.In addition,the different illumination can also interfere with the sampling,all of which bring a lot of difficulties to the research of object tracking.Rely on the rich river port resources of Hunan Province,this paper proposes an object tracking algorithm based on deep learning for the complex video surveillance of ferry,which is used to monitor the ferry ships.The main contributions of this paper are as follows:(1)Object tracking based on convolutional neural network.The traditional method has poor expressiveness in target feature extraction,such as incomplete retrieval information and insufficient feature segmentation.Firstly,the off-line training on CIFAR-10 data sets proves the superiority of the convolutional neural network in feature extraction.The feature extracted by convolutional neural network is better and more expressive,so it can be applied to the feature extraction stage of object tracking.For realistic scenes,we use the convolutional neural network to train the collected hundreds of thousands of ship offline training image data,and obtain the general structural characteristics of the target from simple to complex.Then,a deep network of ship is constructed to track the ship afterwards.(2)Proposed "offline + online + particle filter" object tracking algorithm.Through the above offline training,we obtained the deep network of the ship.In the process of tracking,we use deep model to train the object which is detected online,and implement classification.Then,the particle filter algorithm is used to track the object on-line.Because the ship's deep feature obtained by pre-training model is very rich,even when the target is blocked,the object can be tracked again by online training to fine tune the network parameters.At the same time,in the framework of particle filtering,we resample the object to ensure that the object is not lost and improve the tracking accuracy.Experiments show that the method proposed in this paper has good real-time performance and accuracy for ship tracking.