CN-121982066-A - Ship multi-source data anti-shielding fusion detection method introducing CAFM modules

CN121982066ACN 121982066 ACN121982066 ACN 121982066ACN-121982066-A

Abstract

The invention relates to a ship multisource data anti-shielding fusion detection method introducing CAFM modules, which comprises the steps of receiving AIS data, cleaning the AIS data, carrying out prediction interpolation to achieve time alignment of video data, converting longitude and latitude coordinates obtained by the AIS data into image coordinates, constructing a CAFM-YOLOv8 model, carrying out ship image target recognition by adopting the CAFM-YOLOv8 model, carrying out ship video target tracking by adopting the Bot-Sort-ReID, initially generating a detection frame with unique ID, introducing priori knowledge to carry out anti-shielding treatment, predicting a shielding or deformation detection frame by utilizing the moving characteristics of the AIS data and the video data, updating shielding area judgment, calculating the similarity between each video track and the AIS track by utilizing an E-FastDTW algorithm, finding an optimal result by adopting a Hungary algorithm, and outputting complete state information comprising a ship unique identification code MMSI, longitude, latitude and the like after matching is completed. The invention can realize effective extraction and tracking of the ship target and effectively improve the detection robustness and the target tracking accuracy in the shielding environment.

Inventors

LEI JINYU
ZHONG HONGLIN
Wan Sitong
HE WEI
LIU XINGLONG
WANG ZHIYUAN

Assignees

闽江学院

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (8)

1. A ship multi-source data anti-shielding fusion detection method introduced with CAFM modules is characterized by comprising the following steps: S1, receiving AIS data, cleaning the AIS data, predicting interpolation to realize time alignment of video data, and converting longitude and latitude coordinates obtained by the AIS data into image coordinates through a pinhole imaging model; S2, introducing a convolution-attention fusion CAFM module, embedding the convolution-attention fusion CAFM module into a network structure of YOLOv, and constructing a CAFM-YOLOv8 model; S3, carrying out ship image target identification by adopting a CAFM-YOLOv model, carrying out target tracking of ship video by adopting a multi-target tracking method Bot-Sort-ReID, and preliminarily generating a detection frame with a unique ID; s4, introducing priori knowledge to perform anti-shielding treatment, predicting a shielding or deformation detection frame by utilizing the motion characteristics of AIS data and video data, and updating shielding region judgment; And S5, calculating the similarity between each video track and the AIS track by using an E-FastDTW algorithm, finding out an optimal result by using a Hungary algorithm, and outputting complete state information including a unique ship identification code MMSI, longitude, latitude, navigation speed, heading angle and the like after matching is completed.
2. The method for detecting the anti-occlusion fusion of the marine multisource data introduced into the CAFM module according to claim 1, wherein the received AIS data is passively received by a very high frequency VHF communication link.
3. The method for detecting the anti-occlusion fusion of the ship multi-source data with the CAFM module, according to claim 1, is characterized in that the method for converting longitude and latitude coordinates obtained by AIS into image coordinates through a pinhole imaging model is further as follows: The three-dimensional data acquired by a camera are mapped to a two-dimensional image plane by a pinhole imaging model, the longitude and latitude of the camera when the camera acquires the data are recorded, the shooting level, the pitching angle and the height of the camera are determined, a camera coordinate system is established by taking the optical center of the camera as an origin to obtain the camera coordinates of a specific ship, perspective projection is formed by utilizing the similarity relation of triangles according to the pinhole imaging model, and the three-dimensional world coordinates are converted to the two-dimensional image plane of the camera, wherein the formula is as follows: (1); (2); (3); wherein, the method comprises the following steps of , ) The normalized imaging plane refers to the coordinate with the distance from the origin O as the unit 1, the longitude and latitude coordinates of the camera position are determined according to the GPS positioning point of the camera, the camera coordinate system is established by taking the camera as the center to obtain the camera coordinates (XC, YC, ZC) of the specific ship, Representing the left-right direction position of the object point relative to the camera; Indicating the up-down direction position of the object point relative to the camera, Representing the forward-backward direction distance or depth of the object from the camera; A depth scaling factor representing the perspective projection such that the three-dimensional point can be correctly mapped into the two-dimensional image; , represented as the focal length of the camera, A focal length in the horizontal direction, represents how many pixels of change will occur in the image with a1 unit movement in the x direction, A focal length in the vertical direction, representing the scaling when the change in the y direction is converted to a pixel; , Is the principal point coordinates; represented as the pixel location of the optical center projection x in the image plane, Pixel position denoted as optical center projection y in the image plane; x the left-to-right horizontal pixel position of the projection point in the image; and y is the vertical pixel position of the projection point from top to bottom in the image.
4. The method for detecting multi-source data anti-occlusion fusion of a ship incorporating a CAFM module as set forth in claim 1, wherein the convolution-attention fusion CAFM module includes a local branch and a global branch, The local branches adjust the channel dimension through 1X 1 convolution, the channel shuffling operation is used for further mixing characteristic information, and then the local characteristics are extracted through 3X 3 convolution; the calculation formula of the local branch is as follows: (4); Where W 1x1 represents a 1x1 convolution, CS is a channel reordering operation, W 3x3x3 is a 3 x 3 convolution, and Y is an input feature; the global branch first generates a query Q, key K, and value V tensor by a 1X 1 convolution, and then calculates an attention map. The calculation formula of the attention mechanism is as follows: (5); wherein α is a learnable scaling factor that controls the strength of interaction between Q and K; The output of the global branch is: (6); Finally, the output features of the local branch and the global branch are fused through addition operation, and enhanced feature representation is obtained: (7)。
5. The method for detecting the anti-occlusion fusion of the ship multi-source data introduced into the CAFM module according to claim 1, wherein the obtaining of the complete ship detection frame set at the current moment is further: Comparing the current frame detection results Target ID in the frame and previous frame shielding area If the ID appears in the target ID of (1) But is not present in If the ID is in the range of ID, the detection frame of the target needs to be predicted and corrected Meanwhile, the reliability of the detection result needs to be judged through the following formula: (8); (9); Wherein the method comprises the steps of In order to detect the area ratio of the frame, In order to detect the aspect ratio of the frame, In order to detect the degree of similarity between the frames, And Is a weight coefficient; If it is If the detection result is higher than the set threshold, the detection effect of the detection frame is still good, and the detection result is reserved; After the target to be predicted is determined, the position of the ship in the image at the current moment is predicted by utilizing the motion characteristics of AIS and image data in priori knowledge. Judging associated priori knowledge The method comprises the following steps of predicting through video ship motion characteristics if the detection frame ID is not matched with corresponding MMSI data, wherein the corresponding relation between AIS and MMSI and the detection frame ID at the last moment is included, and the method specifically comprises the following steps: (10); (11); (12); (13); Wherein, the The motion characteristic matrix of the detection frame is calculated by calculating the average value according to the first 10 motion characteristic data of the detection frame corresponding to the ID in the video, And To detect frame in horizontal pixel shift and vertical pixel shift. [ [ Solution to the problem ] Respectively, the upper left corner coordinate and the lower right corner coordinate of the last detection frame [ [ Solution ] and its preparation method The upper left corner coordinate and the lower right corner coordinate of the detection frame coordinate to be predicted; box t-1 is the last time the coordinates of the detection frame are determined by the upper left corner coordinates of the detection frame [ , And lower right corner coordinates , The composition of the composition, Then the frame coordinates to be predicted are also the upper left corner coordinates , And lower right corner coordinates , Composition ]; If the detection frame ID has the data matched with the corresponding MMSI, the AIS motion characteristics and the video motion characteristics are combined for prediction, and the specific formula is as follows: (14); (15); Is AIS motion characteristic matrix corresponding to the detection frame and comprises vertical change rate And a horizontal rate of change 。
6. The method for detecting multi-source data anti-occlusion fusion of a ship by introducing CAFM modules according to claim 1, wherein the priori knowledge includes image occlusion information of a previous moment in a real-time correlation process Correlation information of AIS and image And AIS and track information of previous image And . The information is used for identifying the target which is possibly subjected to false detection or loss, and the detection frame is predicted by combining the motion characteristics so as to obtain the complete ship image target Then it is combined with the track information of the previous image Combining to obtain the image track of the ship at the moment 。
7. The method for detecting the anti-occlusion fusion of the marine multisource data with the CAFM module according to claim 1, wherein the detection frames with the unique IDs are specifically a marine target set ID-boxes= { ID 1 ,ID 2 ,......ID i }, each detection frame follows a format ID-box= [ ID, x 1 ,y 1 ,x 2 ,y 2 ], wherein the ID is a tracking target code, and x 1 ,y 1 and x 2 ,y 2 are respectively designated as coordinates of the upper left corner and the lower right corner of the detection frame.
8. The ship multi-source data anti-occlusion fusion detection method with CAFM introduced as set forth in claim 1, wherein the calculation formula for calculating the similarity between each video track and the AIS track by utilizing FastDTW algorithm is as follows: Two tracks (sequences) x= (X 1 , x 2 , ..., x n ) and y= (Y 1 , y 2 , ..., y m ), The local distance between points is typically defined as the euclidean distance cost (i, j) = ||x i - y j || (or square distance), The DTW distance calculates an accumulated distance matrix D: D (0, 0) =0D (i, 0) = infinity, D (0, j) = infinity (or accumulated boundary) D (i, j) =cost (i, j) +min { D (i-1, j), D (i, j-1), D (i-1, j-1) }, Finally FastDTW (X, Y) =d (n, m).

Description

Ship multi-source data anti-shielding fusion detection method introducing CAFM modules Technical Field The invention relates to the technical field of offshore target identification, in particular to a ship multisource data anti-shielding fusion detection method with CAFM modules. Background With the vigorous development of global offshore trade, the running condition of the offshore ships is gradually complicated, and a series of researches are carried out on various aspects by domestic and foreign scholars in order to improve the identification precision of the offshore commercial ships and the fishing ships under the shielding condition. For example, wang et al propose an extended target measurement cluster identification method based on target Doppler speed information assistance for shielding target detection, chen Faling et al propose a shielding target detection and tracking method based on space-time environment information assistance for target shielding problem in visual images, while aiming at complex water environment, aiming at environmental shielding problem, xian allow-ing develop a research on ship tracking and identification in complex environment based on a deep learning technology, zheng Shaojie designs a set of ship remote sensing image target identification system which effectively considers cloud and fog shielding influence and network model weight reduction, improves accuracy and reliability of marine ship identification, dan Yifang et al are mountain-hugged, mountain-hugged and sea-state object identification, The utility model provides an improved comprehensive track splitting method based on multi-frame data association, which can realize automatic prediction of tracks in the process of target shielding. In the aspect of target detection and tracking of the traditional navigation radar, the method has certain effect on ship shielding research, sun Shuai and the like propose a comprehensive probability data interconnection O-IPDA algorithm under a shielding environment, and the environment shielding condition is pre-judged in real time. Aiming at the problem of shake shielding of the marine vessel, zhang Haiyong and the like, an offshore satellite communication shielding prediction model is researched based on the characteristics of the Fresnel zone and the offshore mobile communication. The automatic ship identification system (AIS) based on the current mainstream is one of hot ships shielding, and AIS data contains key information such as ship position, speed and heading, so that powerful identification support is provided for ship shielding. Based on AIS data, wu Yong and the like, an AIS and video image fusion method based on a comprehensive factor fuzzy evaluation algorithm is provided, the recognition rate of a ship intersection shielding scene is effectively improved through a Darknet network model and a YOLOv algorithm deep learning frame, video ship detection is realized by Wang Jiang based on a deep learning YOLOv s model by using AIS data, the defect of ship information in a video under a shielding condition is overcome, and the augmented reality function of a channel video picture is realized. Due to the development of YOLO series, YOLOv is gradually introduced into the line of sight of marine vessel research, and Zhang Bing provides a light-weight and accurate vessel detection method suitable for traffic intensive scenes based on YOLOv s network under the circumstance that the vessel intensive ports are subject to target detection. In the context of vessel trajectory prediction, chen Xinjiang et al propose a deep-learning prediction model ChebNet-LSTM consisting of chebyshev networks (chebyshev network, chebNet) and LSTM for vessel intersection trajectory prediction. In the field of safety, ship shielding often means occurrence of ship intersection, wang Zhi et al design a CNN+LSTM combined neural network, extract ship navigation environment and ship navigation space-time characteristics, form a ship track and navigation intention recognition model, and predict possibility of ship intersection shielding. Despite some progress, some research on marine vessel shielding has focused mainly on shielding of natural environments, and in vessel target detection and tracking algorithms, recognition failure is highly likely to occur for complex environments, especially large-area shielding between vessels. Disclosure of Invention The invention aims to provide a ship multi-source data anti-occlusion fusion detection method with CAFM modules, which can realize effective extraction and tracking of ship targets and effectively improve detection robustness and target tracking accuracy in an occlusion environment. In order to achieve the purpose, the invention provides the technical scheme that the ship multi-source data anti-shielding fusion detection method with the CAFM module comprises the following steps of: S1, receiving AIS data, cleaning the AIS data, predicting inter