Ship detection and tracking method and device

Patent metadata

Field	Value
Publication number	US-2024371012-A1
Application number	US-202318393732-A
Country	US
Kind code	A1
Filing date	Dec 22, 2023
Priority date	May 5, 2023
Publication date	Nov 7, 2024
Grant date	—

Abstract

Disclosed is a ship detection and tracking method and device, the method comprises: pre-training a yolov5s network model with a first preset number of high-speed moving object datasets to obtain an initial ship detection model; adding a preset attention mechanism to the initial ship detection model to obtain a transitional ship detection model; inputting a second preset number of annotated ship datasets into the initial ship detection model with the preset attention mechanism for training to obtain a target ship detection model; inputting an original ship video frame into the target ship detection model, and outputting the position information of the ship to be tracked; performing ship video image tracking on the ship to be tracked according to the position information. This disclosure performs pre-training, which only requires a small amount of annotated ship data to determine the target ship detection model, achieving video image tracking of the target ship.

First claim

1 . A ship detection and tracking method, comprising:
pre-training a yolov5s network model with a first preset number of high-speed moving object datasets to obtain an initial ship detection model, wherein motion characteristics of the high-speed moving objects are similar to those of ships, the high-speed moving object is a high-speed moving vehicle; adding a preset attention mechanism to the initial ship detection model to obtain a transitional ship detection model; based on a transfer learning strategy, inputting a second preset number of annotated ship datasets into the initial ship detection model with the preset attention mechanism for training to obtain a target ship detection model, comprising: freezing feature extraction layer of the transition ship detection model and performing iterative trainings for a first preset number of times; thawing the feature extraction layer of the transition ship detection model, and obtaining the target ship detection model by iterative trainings for a second preset number of times; wherein in the first preset number of times of iterative trainings, weight parameters of all network layers except last Head layer are frozen, and only ownership weight parameters of Head detection head are updated to obtain an initial pre-trained weight model, in the next second preset number of times of iterative trainings, all feature extraction layers are thawed so that the yolov5s network model updates all the weight parameters; inputting an original ship video frame into the target ship detection model, and outputting the position information of the ship to be tracked; based on a preset tracking algorithm, performing ship video image tracking on the ship to be tracked according to the position information.

2 . (canceled)

3 . The ship detection and tracking method according to claim 1 , based on a preset tracking algorithm, performing ship video image tracking on the ship to be tracked based on the position information, comprising:
using the Kalman filter algorithm to predict the position of the ship to be tracked according to the position information, and determining the track set of the ship to be tracked; based on a cascade matching method, matching the ship track set to be tracked with a detection box, and returning a first matching result; based on an IOU matching strategy, matching and filtering according to the first matching result, and returning a second matching result; updating the track set of the ship to be tracked according to the second matching result, and tracking the next frame of ship video image.

4 . The ship detection and tracking method according to claim 3 , using the Kalman filter algorithm to predict the position of the ship to be tracked according to the position information, and determining the track set of the ship to be tracked, comprising:
determining whether the track of the ship to be tracked exists according to the position information through the Kalman filter algorithm; if the track of the ship to be tracked exists, using the Kalman filter algorithm to predict the position of the ship to be tracked to obtain a priori track; initializing and classifying the priori track to obtain the track set of the ship to be tracked.

5 . The ship detection and tracking method according to claim 3 , based on a cascade matching method, matching the ship track set to be tracked with a detection box, and returning a first matching result, comprising:
calculating a first cost matrix according to the track set of the ship to be tracked; based on the preset matching method, matching the track set of the ship to be tracked with the detection box according to the first cost matrix, and obtaining the first matching result.

6 . The ship detection and tracking method according to claim 5 , based on an IOU matching strategy, matching and filtering according to the first matching result, and returning a second matching result, comprising:
calculating a second cost matrix of the track to be matched and the unmatched detection box according to the first matching result; based on a preset matching method, matching the track to be matched and the unmatched detection box according to the second cost matrix; deleting the area whose IOU is less than a preset threshold after matching, and obtaining the second matching result.

7 . The ship detection and tracking method according to claim 3 ,
updating the track set of the ship to be tracked according to the second matching result, and tracking the next frame of ship video image, comprising: updating the matched track state through the Kalman filter algorithm, and continuing to match the state of a failed track; converting the unmatched detection box into an unmatched track, and adding the successfully matched track to the ship track set to be tracked; updating the ship track set to be tracked, and performing the next frame of ship video image tracking until the target ship video image tracking is completed.

8 . (canceled)

9 . An electronic device, comprising a memory and a processor, wherein:
the memory configured to store one or more programs which, when executed by the processor, cause the processor to: pre-train yolovs network model with a first preset number of high-speed moving object datasets to obtain an initial ship detection model, wherein motion characteristics of the high-speed moving objects are similar to those of ships, the high-speed moving object is a high-speed moving vehicle; add a preset attention mechanism to the initial ship detection model to obtain a transitional ship detection model; based on a transfer learning strategy, input a second preset number of annotated ship datasets into the initial ship detection model with the preset attention mechanism for training to obtain a target ship detection model, comprising: freeze feature extraction layer of the transition ship detection model and perform iterative trainings for a first preset number of times; thaw the feature extraction layer of the transition ship detection model, and obtain the target ship detection model by iterative trainings for a second preset number of times; wherein in the first preset number of times of iterative trainings, weight parameters of all network layers except last Head layer are frozen, and only ownership weight parameters of Head detection head are updated to obtain an initial pre-trained weight model, in the next second preset number of times of iterative trainings, all feature extraction layers are thawed so that the yolov5s network model updates all the weight parameters; input an original ship video frame into the target ship detection model, and output the position information of the ship to be tracked; based on a preset tracking algorithm, perform ship video image tracking on the ship to be tracked according to the position information.

10 . A non-transitory computer-readable storage medium, which is used to store programs or instructions that can be read by a computer, when the program or instruction is executed by a processor, causes the processor to perform a method, the method comprising:
pre-training a yolov5s network model with a first preset number of high-speed moving object datasets to obtain an initial ship detection model, wherein motion characteristics of the high-speed moving objects are similar to those of ships, the high-speed moving object is a high-speed moving vehicle; adding a preset attention mechanism to the initial ship detection model to obtain a transitional ship detection model; based on a transfer learning strategy, inputting a second preset number of annotated ship datasets into the initial ship detection model with the preset attention mechanism for training to obtain

Assignees

• Univ Wuhan Tech

Inventors

• Chen Linying
• Huang Yamin
• Chen Pengfei
• Mou Junmin

Classifications

• G06T7/246 Primary

  using feature-based methods, e.g. the tracking of corners or segments · CPC title

• G06T7/70

  Determining position or orientation of objects or cameras (camera calibration G06T7/80) · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/10016

  Video; Image sequence · CPC title