High-precision ship multi-target tracking method, device and equipment and storage medium

Abstract

The invention discloses a high-precision ship multi-target tracking method, device and equipment and a storage medium, the ship multi-target tracking method based on YOLOv5 and an improved DeepSORT algorithm comprises the following steps: firstly, connecting an IP camera through an RTSP protocol, then collecting ship data by using the camera, screening and marking the collected ship image data, and finally, carrying out high-precision ship multi-target tracking. A detection data set in a YOLO format and a data set in an MOT format are obtained, and the detection data set and the data set are divided into three types of training, verification and testing; then, training is carried out in the YOLOv5 network by utilizing a self-built detection data set, and a ship model weight based on the YOLOv5 network is obtained; secondly, training an improved DeepSORT network by using a self-established tracking data set to obtain a tracking weight, and finally, combining a detection algorithm with a tracking algorithm, synchronizing a detection result to an improved DeepSORT multi-target tracking algorithm, and obtaining ID information and position information of a ship target; according to the invention, a DeepSORT multi-target tracking algorithm is improved, and a YOLOv5 algorithm and a heavy-load pan-tilt camera are matched to track a ship in real time; and the tracking precision is high and the real-time performance is good.

Description

Technical Field

[0001] The present invention belongs to the field of target tracking, and in particular relates to the technology of a high-precision ship multi-target tracking method, device, equipment and storage medium. Background Art

[0002] Target tracking is to detect, extract, identify and track moving targets in a time series of continuous frame images, obtain the position and motion trajectory of the tracking target, and understand the behavior of the moving target to complete a higher level of detection tasks. According to the number of tracking targets, the tracking algorithm can be divided into single target tracking and multi-target tracking. Compared with single target tracking, multi-target tracking is more complex and difficult. The main task of multi-target tracking is to locate multiple targets of interest at the same time in a given video, maintain the ID of each target of interest and record their trajectories.

[0003] In view of the specific marine environment, first of all, part of the problem of ship multi-target tracking lies in the accuracy of target detection on its basis. In the target tracking algorithm, the accuracy of target detection affects the tracking effect; secondly, the ship will be affected by the wind and waves during navigation, causing the camera to be unable to stably obtain image data with a fixed perspective, resulting in more difficult target tracking and thus causing large errors in the ship's multi-target tracking; finally, the changeable marine climate causes the camera to produce different light and shadow images when acquiring image data, resulting in the target tracking algorithm being unable to track the target ideally during implementation, and the current target tracking algorithm is difficult to overcome the impact of problems such as overlapping and occlusion of targets during the ship's navigation.

[0004] In recent years, target tracking has become a common method for multi-target tracking, mainly represented by the SORT algorithm and the DeepSORT algorithm. The core of SORT is Kalman filtering and Hungarian matching. However, the ID of the same person will change after occlusion and camera posture changes, and the navigation environment of ships at sea is complex and changeable, so it cannot guarantee detection accuracy and real-time performance. To solve this problem, the appearance branch and motion branch of the DeepSORT algorithm are improved.

[0005] The ship multi-target tracking method is the core of the marine ship intelligent monitoring system, and it is of great significance to improve the accuracy of marine target detection and tracking. Summary of the invention

[0006] In view of this, an embodiment of the present invention provides a high-precision ship multi-target tracking method to optimize the target feature extraction capability, replace the feature update strategy, and improve the accuracy of target tracking.

[0007] In a first aspect, an embodiment of the present invention provides a high-precision ship multi-target tracking method, characterized by comprising:

[0008] S1: Obtain an IP camera image video stream and collect ship image data, pre-process the collected ship image data, and construct training, verification and test data sets based on the processed images;

[0009] S2: Improve the appearance branch and motion branch of the DeepSORT algorithm to obtain the improved DeepSORT algorithm;

[0010] S3: Based on the training data set and the validation data set, the YOLOv5 algorithm and the improved DeepSORT algorithm are trained to obtain a YOLOv5 detection model and a DeepSORT tracking model;

[0011] S4: combining the YOLOv5 detection model and the DeepSORT tracking model, synchronizing the target position information detected by the YOLOv5 detection model to the improved DeepSORT algorithm, and the DeepSORT algorithm tracks the ship in real time and obtains tracking ID information;

[0012] The S1 obtains the IP camera image video stream through the RTSP protocol using the device user name, user password, device IP address and RTSP port number.

[0013] Preferably, in S2, the step of improving the DeepSORT algorithm is specifically:

[0014] S021: performing standardization preprocessing on the images in the training data set, and inputting the preprocessed images into the ResNet50 network for feature extraction;

[0015] S022: based on the loss function and back propagation, the gradient processing feature data is updated in real time and verified using the verification data set;

[0016] S023: Obtain the improved DeepSORT algorithm.

[0017] Preferably, the YOLOv5 network includes a backbone part, a neck part and a prediction part;

[0018] In S3, the method for obtaining the YOLOV5 detection model includes:

[0019] S31: performing standardization preprocessing on the images in the training data set, and inputting the preprocessed images into the backbone part to obtain feature maps of different scales;

[0020] S32: Inputting the feature maps of different scales into the neck part, performing upsampling and feature fusion to obtain tensor data of different scales;

[0021] S33: Input the tensor data of different scales into the prediction part, calculate the gradient based on the loss function and back propagation, update the gradient in real time and verify it using the verification set to obtain the YOLOv5 detection model.

[0022] Preferably, the improved DeepSORT algorithm construction process in S3 is to replace the original simple CNN network with the ResNet50 network.

[0023] Preferably, the ship image data in S1 includes an open source dataset of ship-related image data and image data obtained by photographing the ship with a camera;

[0024] The open source datasets include but are not limited to: COCO dataset, VOC dataset and SeaShip dataset.

[0025] Preferably, in any of the above methods, the specific steps of performing the standardization preprocessing on the image include:

[0026] S011: Filtering the collected ship image data to obtain images containing ships in the COCO dataset and the VOC dataset;

[0027] S012: Annotate and merge the images containing ships, the SeaShip ship dataset, and the images obtained through the camera;

[0028] S013: Create a measurement dataset in YOLO format and a tracking dataset in MOT format.

[0029] Preferably, S4 includes:

[0030] S041: performing target detection on the ship image data frame by frame using the YOLOv5 algorithm and the improved DeepSORT algorithm to obtain a detection frame of the target ship in the i-th frame of the ship image data;

[0031] S042: Based on the detection frame of the current i-th frame, a Kalman filter algorithm with global motion compensation technology is used to perform state prediction to obtain a corresponding target ship tracking frame;

[0032] S043: Calculating the motion matching degree and appearance matching degree between all the target ship detection frames and the target ship tracking frames;

[0033] Comprehensively matching the motion matching degree and the appearance matching degree to obtain an associated cost, and constructing an associated cost matrix;

[0034] S044: performing association matching on the association cost matrix by using the Hungarian algorithm;

[0035] S045: updating the parameters of the target ship tracking frame that has been matched by the Kalman filter update formula, predicting the target ship at the next moment, allocating a new tracking frame to the target ship that has not been matched to predict the motion state, and updating the information obtained from all tracking frames for calculation of the next frame;

[0036] Among them, i=1,2,3...N.

[0037] In a second aspect, a high-precision ship multi-target tracking device is provided, characterized in that the device comprises:

[0038] Ship image data acquisition module: The image sources mainly include open source datasets and ship image data acquired through cameras. The open source datasets mainly use COCO datasets and VOC datasets containing ship-related image data and SeaShip ship datasets;

[0039] Detection dataset processing module: Filter out images with the "boat" label in the COCO dataset and the VOC dataset; annotate the filtered images and the images of ships acquired by the camera using image annotation software in the format of category and location information; then organize the annotated images and other image data into a new self-built dataset, which is integrated into the six categories in the SeaShip dataset, namely: ore carrier, general cargo ship, bulk cargo carrier, container ship, fishing boat, and passenger ship;

[0040] Tracking data set processing module: The ship video captured by the camera is split into continuous frame images and annotated using annotation software in the format of frame number, ID information, location, whether the target trajectory is within the consideration range, category, and the impact between targets.

[0041] According to a third aspect, a high-precision ship multi-target tracking device is provided, characterized in that it includes: at least one heavy-duty camera of a pan / tilt head, a processor, at least one memory, and computer program instructions stored in the memory, and when the computer program instructions are executed by the processor, the method described in any one of the above items is implemented.

[0042] According to a fourth aspect, a storage medium is provided, on which computer program instructions are stored, characterized in that when the computer program instructions are executed by a processor, a method as described in any one of the above items is implemented.

[0043] In summary, the beneficial effects of the present invention are as follows:

[0044] Through the above method, the target feature extraction capability is optimized, the feature update strategy is replaced, and the tracking accuracy is improved; it can also effectively solve the ID transformation of camera posture changes; Gaussian process regression is used to propose a lightweight interpolation method, which can solve the impact of the emergence of new heavy targets on tracking after the detection is lost due to occlusion; ResNet50 is used to replace the original CNN as the backbone network to improve the tracking ability of the tracking model for ships; the EMA feature update strategy is used to replace the feature library update strategy to improve the algorithm running speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide further understanding of the present invention and constitute a part of this application. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0046] In the attached picture:

[0047] Figure 1 The system workflow diagram provided by the present invention;

[0048] Figure 2 This is a flow chart of multi-target tracking provided by the present invention. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0050] Example 1

[0051] See also Figure 1 and Figure 2 The embodiment of the present invention provides a high-precision ship multi-target tracking method, comprising:

[0052] In a first aspect, an embodiment of the present invention provides a high-precision ship multi-target tracking method, characterized by comprising:

[0053] S1: Obtain an IP camera image video stream and collect ship image data, pre-process the collected ship image data, and construct training, verification and test data sets based on the processed images;

[0054] S2: Improve the appearance branch and motion branch of the DeepSORT algorithm to obtain the improved DeepSORT algorithm;

[0055] S3: Based on the training data set and the validation data set, the YOLOv5 algorithm and the improved DeepSORT algorithm are trained to obtain a YOLOv5 detection model and a DeepSORT tracking model;

[0056] The method of synchronizing the target position information detected by the YOLOv5 model to the improved DeepSORT target tracking algorithm is as follows: the target detected by YOLOv5 is used as the output, replacing the original detection algorithm in the DeepSORT target tracking algorithm;

[0057] S4: combining the YOLOv5 detection model and the DeepSORT tracking model, synchronizing the target position information detected by the YOLOv5 detection model to the improved DeepSORT algorithm, and the DeepSORT algorithm tracks the ship in real time and obtains tracking ID information;

[0058] The S1 obtains the IP camera image video stream through the RTSP protocol using the device user name, user password, device IP address and RTSP port number.

[0059] Preferably, in S2, the step of improving the DeepSORT algorithm is specifically:

[0060] S021: performing standardization preprocessing on the images in the training data set, and inputting the preprocessed images into the ResNet50 network for feature extraction;

[0061] S022: based on the loss function and back propagation, the gradient processing feature data is updated in real time and verified using the verification data set;

[0062] S023: Obtain the improved DeepSORT algorithm.

[0063] Preferably, the YOLOv5 network includes a backbone part, a neck part and a prediction part;

[0064] In S3, the method for obtaining the YOLOV5 detection model includes:

[0065] S31: performing standardization preprocessing on the images in the training data set, and inputting the preprocessed images into the backbone part to obtain feature maps of different scales;

[0066] S32: Inputting the feature maps of different scales into the neck part, performing upsampling and feature fusion to obtain tensor data of different scales;

[0067] S33: Input the tensor data of different scales into the prediction part, calculate the gradient based on the loss function and back propagation, update the gradient in real time and verify it using the verification set to obtain the YOLOv5 detection model.

[0068] Preferably, the improved DeepSORT algorithm construction process in S3 is to replace the original simple CNN network with the ResNet50 network.

[0069] Preferably, the ship image data in S1 includes an open source dataset of ship-related image data and image data obtained by photographing the ship with a camera;

[0070] The open source datasets include but are not limited to: COCO dataset, VOC dataset and SeaShip dataset.

[0071] Preferably, in any of the above methods, the specific steps of performing the standardization preprocessing on the image include:

[0072] S011: Filtering the collected ship image data to obtain images containing ships in the COCO dataset and the VOC dataset;

[0073] S012: Annotate and merge the images containing ships, the SeaShip ship dataset, and the images obtained through the camera;

[0074] S013: Create a measurement dataset in YOLO format and a tracking dataset in MOT format.

[0075] Preferably, S4 includes:

[0076] S041: performing target detection on the ship image data frame by frame using the YOLOv5 algorithm and the improved DeepSORT algorithm to obtain a detection frame of the target ship in the i-th frame of the ship image data;

[0077] S042: Based on the detection frame of the current i-th frame, a Kalman filter algorithm with global motion compensation technology is used to perform state prediction to obtain a corresponding target ship tracking frame;

[0078] S043: Calculating the motion matching degree and appearance matching degree between all the target ship detection frames and the target ship tracking frames;

[0079] Comprehensively matching the motion matching degree and the appearance matching degree to obtain an associated cost, and constructing an associated cost matrix;

[0080] The appearance matching is improved by replacing the ResNet50 network and the EMA feature update strategy to improve accuracy and speed;

[0081] S044: performing association matching on the association cost matrix by using the Hungarian algorithm;

[0082] The tracking frames are divided according to the length of time the target corresponding to the tracking frame is occluded. The tracking frames of the targets with shorter occlusion time and more recent appearance are given priority, while the tracking frames of the targets with longer occlusion time are given lower priority. The detection frames may be lost due to occlusion, and the lightweight interpolation method is used to smooth out the situation of short-term target loss.

[0083] S045: updating the parameters of the target ship tracking frame that has been matched by the Kalman filter update formula, predicting the target ship at the next moment, allocating a new tracking frame to the target ship that has not been matched to predict the motion state, and updating the information obtained from all tracking frames for calculation of the next frame;

[0084] When the predicted result of the updated parameters cannot match, it means that the current target ship may have been lost, and the tracking frame is deleted;

[0085] The tracking data set includes but is not limited to a data set annotated with images acquired through a camera;

[0086] Among them, i=1,2,3...N.

[0087] In a second aspect, a high-precision ship multi-target tracking device is provided, characterized in that the device comprises:

[0088] Ship image data acquisition module: The image sources mainly include open source datasets and ship image data acquired through cameras. The open source datasets mainly use COCO datasets and VOC datasets containing ship-related image data and SeaShip ship datasets;

[0089] Detection dataset processing module: Filter out images with the "boat" label in the COCO dataset and the VOC dataset; annotate the filtered images and the images of ships acquired by the camera using image annotation software in the format of category and location information; then organize the annotated images and other image data into a new self-built dataset, which is integrated into the six categories in the SeaShip dataset, namely: ore carrier, general cargo ship, bulk cargo carrier, container ship, fishing boat, and passenger ship;

[0090] Tracking data set processing module: The ship video captured by the camera is split into continuous frame images and annotated using annotation software in the format of frame number, ID information, location, whether the target trajectory is within the consideration range, category, and the impact between targets.

[0091] According to a third aspect, a high-precision ship multi-target tracking device is provided, characterized in that it includes: at least one heavy-duty camera of a pan / tilt head, a processor, at least one memory, and computer program instructions stored in the memory, and when the computer program instructions are executed by the processor, the method described in any one of the above items is implemented.

[0092] According to a fourth aspect, a storage medium is provided, on which computer program instructions are stored, characterized in that when the computer program instructions are executed by a processor, a method as described in any one of the above items is implemented.

[0093] Example 2

[0094] See also Figure 1 and Figure 2 The embodiment of the present invention provides a high-precision ship multi-target tracking method. In S3, the improved target tracking algorithm includes the following steps:

[0095] 1. A GMC technique for the camera stability module is proposed. This image registration method is suitable for background motion. First, the image key points are extracted, and then sparse optical flow is used for feature tracking and translation-based local outlier suppression. The sparse registration technique allows the motion in the scene to be ignored based on the detection results, thus having the ability to estimate the background motion more accurately. Through the affine matrix Convert the prediction box from frame k-1 to frame k. The translation part of the transformation matrix only affects the center position of the bounding box. The camera motion correction step is calculated as follows:

[0096]

[0097]

[0098]

[0099]

[0100] When the matrix M∈R 2×2 When A contains the scale and rotation parts of the affine matrix A and T contains the translation part, we can define it with mathematical tricks. and

[0101] also, and are the predicted state vectors before and after camera motion compensation, respectively. k|k-1 and P k|k-1 are the covariance matrices before and after correction, respectively;

[0102] Finally, in the Kalman filter we update and

[0103] 2. Use Gaussian process regression to propose a lightweight interpolation method to model nonlinear motion.

[0104] Build a model for the i-th trajectory:

[0105] p t =f (i) (t)+γ;

[0106] t is the frame, p t is the position coordinate of frame t, γ is Gaussian noise, and the nonlinear motion model is obtained by fitting the function f (i) accomplish;

[0107] Assume f (i) This is a Gaussian process:

[0108] f (i) ∈GP(0,k(x,x′));

[0109] Among them, k(x, x') is the Gaussian function kernel. Given a new frame set F * , the smoothed position set P * The prediction is:

[0110] P * =K(F * ,F)(K(F,F)+σ 2 I) -1 P;

[0111] This model is used to solve the impact of re-detecting the target on tracking after detection is lost;

[0112] 3. Use ResNet50 to replace the original CNN as the backbone network;

[0113] 4. Using the EMA feature update strategy to replace the update strategy of the feature library not only enhances the matching quality, but also saves the matching time. It is no longer necessary to retain 100 feature vectors for a trajectory like the feature library. The performance characteristics of trajectory i in the tth frame are as follows:

[0114]

[0115] Where α is the weight, f i t To detect appearance features, so as to improve the accuracy of target tracking;

[0116] Based on a self-built data set, the present invention trains a detection model in the YOLOv5 network, which can realize the target detection function, and trains a tracking model in the DeepSORT algorithm based on the ResNet50 network, which can realize the target tracking function. The combined detection and tracking algorithm can track multiple targets of the ship in real time.

[0117] Compared with the existing multi-target tracking technology, the GMC technology of the camera stability module proposed can effectively solve the ID transformation of the camera posture change; the lightweight interpolation method proposed by using Gaussian process regression can solve the impact of the new appearance of heavy targets on tracking after the detection is lost due to occlusion; ResNet50 is used to replace the original CNN as the backbone network to improve the tracking ability of the tracking model for ships; the EMA feature update strategy is used to replace the feature library update strategy to improve the algorithm running speed.

[0118] Example 3

[0119] See also Figure 1 and Figure 2 The embodiment of the present invention provides a high-precision ship multi-target tracking method. In order to improve the multi-target tracking accuracy, the present invention combines the target detection network (YOLOv5) and the Deepsort target tracking algorithm, and combines a heavy-duty camera with a pan-tilt head to propose a ship multi-target tracking method based on YOLOv5 and an improved DeepSORT algorithm;

[0120] It mainly solves the problem of motion compensation during camera rotation and improves the tracking accuracy of the algorithm;

[0121] DeepSORT consists of appearance branch and motion branch;

[0122] In the appearance branch, CNN is used to extract the appearance features of the detected target in each frame. The appearance features of each track in the past 100 frames are stored in the feature library mechanism. When there is a new detection, the minimum cosine distance between the track and the detected features in the feature library is used as the matching cost of the association step.

[0123] In the motion branch, Kalman is used to predict the trajectory position in the current frame, and then the Mahalanobis distance is used to calculate the temporal and spatial gap between the trajectory and the detection to eliminate targets that cannot be associated.

[0124] The present invention improves the two branch structures of the DeepSORT algorithm;

[0125] In the motion branch, Kalman filtering is used in the multi-target tracking deepsort algorithm to predict the target position trajectory of the next frame for association. Kalman filtering will make the trajectory stable as a whole and can maintain the position of the frame under some occlusions. However, in the case of non-uniform posture motion of the camera, Kalman filtering will result in suboptimal frames.

[0126] In this case, a camera motion compensation method is introduced for camera motion to reduce the impact of camera posture changes by correcting the image plane coordinates. The multi-target tracking algorithm of joint detection relies heavily on the overlap of the trajectory prediction frame and the detection frame. When the camera posture changes, the position of the image plane frame changes dramatically, resulting in an increase in ID conversion or false negatives. In the absence of additional information about camera motion (navigation, IMU, etc.) or the camera's intrinsic matrix, image matching based on adjacent frames is a good approach.

[0127] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements that are inherent to such process, method, article or device.

[0128] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0129] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A high-precision ship multi-target tracking method, characterized in that: include: S1: Obtain an IP camera image video stream and collect ship image data, pre-process the collected ship image data, and construct training, verification and test data sets based on the processed images; S2: Improve the appearance branch and motion branch of the DeepSORT algorithm to obtain the improved DeepSORT algorithm; S3: Based on the training data set and the validation data set, the YOLOv5 algorithm and the improved DeepSORT algorithm are trained to obtain a YOLOv5 detection model and a DeepSORT tracking model; S4: combining the YOLOv5 detection model and the DeepSORT tracking model, synchronizing the target position information detected by the YOLOv5 detection model to the improved DeepSORT algorithm, and the DeepSORT algorithm tracks the ship in real time and obtains tracking ID information; The S1 obtains the IP camera image video stream through the RTSP protocol using the device user name, user password, device IP address and RTSP port number.

2. The method according to claim 1, characterized in that In S2, the steps of improving the DeepSORT algorithm are specifically as follows: S021: performing standardization preprocessing on the images in the training data set, and inputting the preprocessed images into the ResNet50 network for feature extraction; S022: based on the loss function and back propagation, the gradient processing feature data is updated in real time and verified using the verification data set; S023: Obtain the improved DeepSORT algorithm.

3. The method according to claim 1, characterized in that The YOLOv5 network includes a backbone part, a neck part and a prediction part; In S3, the method for obtaining the YOLOV5 detection model includes: S31: performing standardization preprocessing on the images in the training data set, and inputting the preprocessed images into the backbone part to obtain feature maps of different scales; S32: Inputting the feature maps of different scales into the neck part, performing upsampling and feature fusion to obtain tensor data of different scales; S33: Input the tensor data of different scales into the prediction part, calculate the gradient based on the loss function and back propagation, update the gradient in real time and verify it using the verification set to obtain the YOLOv5 detection model.

4. The method according to claim 1, characterized in that The improved DeepSORT algorithm construction process in S3 is to replace the original CNN network with the ResNet50 network.

5. The method according to claim 1, wherein the ship image data in S1 includes an open source dataset of ship-related image data and image data obtained by photographing the ship with a camera; The open source datasets include but are not limited to: COCO dataset, VOC dataset and SeaShip dataset.

6. The method according to claims 1 to 5, characterized in that the specific steps of performing the standardization preprocessing on the image include: S011: Filtering the collected ship image data to obtain images containing ships in the COCO dataset and the VOC dataset; S012: Annotate and merge the images containing ships, the SeaShip ship dataset, and the images obtained through the camera; S013: Create a measurement dataset in YOLO format and a tracking dataset in MOT format.

7. The method according to claim 1, characterized in that The S4 includes: S041: performing target detection on the ship image data frame by frame using the YOLOv5 algorithm and the improved DeepSORT algorithm to obtain a detection frame of the target ship in the i-th frame of the ship image data; S042: Based on the detection frame of the current i-th frame, a Kalman filter algorithm with global motion compensation technology is used to perform state prediction to obtain a corresponding target ship tracking frame; S043: Calculating the motion matching degree and appearance matching degree between all the target ship detection frames and the target ship tracking frames; Comprehensively matching the motion matching degree and the appearance matching degree to obtain an associated cost, and constructing an associated cost matrix; S044: performing association matching on the association cost matrix by using the Hungarian algorithm; S045: updating the parameters of the target ship tracking frame that has been matched by the Kalman filter update formula, predicting the target ship at the next moment, allocating a new tracking frame to the target ship that has not been matched to predict the motion state, and updating the information obtained from all tracking frames for calculation of the next frame; Among them, i=1,2,3...N.

8. A high-precision ship multi-target tracking device, characterized in that: The device comprises: Ship image data acquisition module: The image sources mainly include open source datasets and ship image data acquired through cameras. The open source datasets mainly use COCO datasets and VOC datasets containing ship-related image data and SeaShip ship datasets; Detection dataset processing module: Filter out images with the "boat" label in the COCO dataset and the VOC dataset; annotate the filtered images and the images of ships acquired by the camera using image annotation software in the format of category and location information; then organize the annotated images and other image data into a new self-built dataset, which is integrated into the six categories in the SeaShip dataset, namely: ore carrier, general cargo ship, bulk cargo carrier, container ship, fishing boat, and passenger ship; Tracking data set processing module: The ship video captured by the camera is split into continuous frame images and annotated using annotation software in the format of frame number, ID information, location, whether the target trajectory is within the consideration range, category, and the impact between targets.

9. A high-precision ship multi-target tracking device, characterized in that , including: at least one gimbal heavy-duty camera, a processor, at least one memory and computer program instructions stored in the memory. When the computer program instructions are executed by the processor, the method described in any one of 1-7 is implemented.

10. A storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images