Studies And Implemention Of Key Technology On Inland River Ship Detection Based On Video Sequence

The inland navigation has advantage of large capacity and low cost,especially for the mass goods.To locate the targets accurately is the precondition of video ship surveillance.However,the water causes interference with its ripples and reflection,so there are many false-positive samples in the ship’s targeting using the background modeling algorithm,which affects following steps such as calibration,statistics.Besides,the illegal berthing affects normal navigation,disturbs traffic order,and brings about latent dangers.Based on the project"Research and Application of Key Technologies in Real-Time Video Monitoring and Recognizing System",we research on ships’ video surveillance,which has important theoretical significance and economic value.This paper describes current research on worldwide ship detection at recent years,and compares several common detection algorithms.Two post-processing methods are proposed:one is based on mathematical morphology,while the other is based on S-SIFT feature.Using the convolutional neural network,we give an optimization algorithm to solve the problem brought by the background method.Besides,we conduct an in-depth study on Faster R-CNN,and compare its ship detection effect with the traditional background modeling algorithm on the same datasets.Aiming at targeting effect of background modeling,we propose two post-processing approaches,mathematical morphology method and S-SIFT feature detection.Eroding and dilating the binary images with morphological structure elements to eliminate some noise and fill the void area,thus we could obtain a more complete outline of the ship.With training the SIFT features of image keypoints,we get the coding dictionary and S-SIFT features.Finally the features are classified by.a linear SVM to remove misjudgements,and therefore both accuracy and efficiency are guaranteed.Aiming at illegal berthing,this paper gives a detection algorithm based on convolutional neural network.A CNN model is learned by training ship images and forbidden berthing area images.The model contains five convolutional layers and two full-connection layers,and it classifies the features exported by the last full-connection layer.The suspected illegal berthing target is kept monitoring by the model,thus we solve the problem of the integration of long-time berthing ships into background and of misjudgment upon leaving ships.Aiming at target locating,this paper gives a detection method using Faster R-CNN.A region proposed network is constructed to extract the ROI regions on feature maps with sliding windows and anchor boxes.The classify layer uses these regions and feature maps to get ship targets.Compared with the background modeling algorithm,this meghod could detect the local targets and the adjacent targets better.Besides,it avoids getting multi-boxes on a single ship target.The main innovations of this paper are as follows:● It proposes the S-SIFT feature model,which contributes to eliminate the misjudgements,thus improving the accuracy of ship target detection.● It optimizes the illegal berthing detection algorithm by distinguishing ship and no-ship with a trained convolutional neural network,thus solving the problem of the integration of long-time berthing ships into background and of misjudgment upon leaving ships.● It gives a video ship target detection method based on Faster R-CNN,which solves the problem of target split,targets fusion.