Ship Detection And Monitoring Method Based On Deep Learning

Ship monitoring technology is widely used in water traffic management,ship detection and ship violation identification.With the development of deep learning,ship detection algorithm based on convolutional neural network has become an important method of ship monitoring.In ship detection,the scale of ships varies widely,and the aspect ratio of the shape is large,which seriously affects the detection accuracy.In this regard,thesis has carried out the research on ship target detection algorithm based on deep learning and ship tracking application research,focusing on the generation method of multi-scale ship matching anchor box,adaptive feature extraction and prediction head decoupling detection method,and improved Deep-Sort ship tracking algorithm integrating D-IoU distance,the main research contents are as follows:（1）Due to the scale range of ships varies greatly,and the size and aspect ratio of the manual anchors are fixed,the traditional anchor strategy cannot provide effective reference boxes for ships,which easily leads to missed detections.Aiming at the problem that the traditional manual anchor cannot match the multi-scale ship target,a featureguided anchor generation algorithm is proposed.Based on the semantic features of the ship,the convolutional neural network is used to generate the optimal matching anchor box for the ship,which has the greatest overlap with the ground-truth box.Experiments show that,based on the Retina Net detector,the detection accuracy of our algorithm is improved by 4.8 percentage points compared with the traditional anchor method,and the calculation amount is reduced by 29.6%,which proves that the method in thesis can generate accurate anchor for ships,significantly improve the detection accuracy,and provide an efficient anchor generation method for the detectors.（2）Since the shape of the ship is mainly long strip and the aspect ratio is large,and most convolutional networks use square convolution to extract features.So the accuracy of coordinate regression needs to be improved.Aiming at the problem that the 3×3standard convolution of feature extraction does not fit the ship,an adaptive feature extraction algorithm based on deformable convolution is proposed.The convolutional network is used to predict the coordinate offset of the sampling point,then deformable convolution is used to adaptively extract the features of offset position,and finally constructs a decoupling prediction network to improve the feature extraction ability and prediction ability of the network.Based on the YOLOv4 network,the accuracy and average detection accuracy of this algorithm are improved by 5.6% and 5.1% respectively compared with the original algorithm,which proves that this algorithm is more efficient in feature extraction of ship targets.（3）Since the Deep-Sort algorithm is prone to tracking ID switch and tracking failures when the background changes greatly.Aiming at the fact that the cascade matching loss of the Deep-Sort algorithm cannot accurately measure the relative positions of the two boxes,an improved Deep-Sort ship tracking algorithm integrating D-IoU distance is proposed.The intersection and union ratio is used to represent the overlap degree of two frames,and the diagonal distance and center point distance of the minimum circumscribed rectangle of the matching frame are constructed to represent the non-overlapping area and relative position.In addition,a two-line statistical method is proposed for the determination of the ship’s motion direction.Experiments show that in the three test videos,the performance of the algorithm in thesis is better than the Deep-Sort algorithm,and it can achieve stable tracking for ship with large scale changes and partial occlusions.During the tracking,the two-line statistical method can effectively represent the positional relationship between the ship and the marking lines,and accurately judge the traveling direction of the ship.