Automatic ship tracking method and system based on deep learning network and mean shift

1. An automatic ship tracking method based on a deep learning network and a mean shift, comprising the following steps:

step 1: collecting surveillance video data, wherein the surveillance video data comprises coastal region surveillance video data under visible light; and extracting a video image from the surveillance video data;

step 2: performing a preprocessing based on the video image obtained in the step 1 to extract a positive sample and a negative sample of a first ship target;

step 3: inputting the positive sample and the negative sample of the first ship target in the video image into a neural network to train a model by a region-based convolutional neural network method;

step 4: extracting initial frame data of the video image, and performing a ship detection and a probability density calculation on the initial frame data according to the model obtained in the step 3, wherein the probability density calculation is realized by dividing a gray color space of a target region to obtain a gray histogram formed by multiple equal histogram intervals, and calculating the probability density according to one histogram interval of the multiple equal histogram intervals, a gray value of a pixel in the target region belongs to the one histogram interval; and

step 5: determining a ship tracking result at a current moment by a calculation result of a previous moment, wherein the step 5 further comprises the following processing:

step A: taking ξ ship positions tracked at a moment t−1 as initial positions, respectively taking a center coordinate f 0 of each of the ξ ship positions as an initial target position of a ship tracked at a moment t, taking the f 0 as a center of a search window to obtain a center position coordinate f of a corresponding candidate ship, calculating a region histogram of a candidate position, and further realizing the probability density calculation;

step B: describing a degree of similarity between a ship model and the corresponding candidate ship by a Bhattacharyya coefficient, calculating a mean shift iteration equation in a center of the target region, so that the ship model moves continuously towards a direction with a maximum color change until last two moving distances are less than a corresponding preset threshold, finding a ship position obtained by a mean shift result at the moment t, setting that multiple ship positions Boxm t are obtained, and expressing an id th ship position as Boxm id t ;

step C: by the region-based convolutional neural network method, performing the ship detection on the video image at the moment t, and obtaining the num th detection coordinate Boxd num t of multiple ships in the video image at the moment t, calculating a degree of overlap between the Boxd num t and the Boxm id t , recording a value of degree of overlap O max between each Boxm id t and the corresponding Boxd num t with a maximum degree of overlap, and if the O max is less than a corresponding threshold θ 1 , considering the ship position as false-alarm and deleting the Boxm id t ; and

step D: updating a second ship target appeared at the moment t by a neural network detection result of the step C, wherein the step D further comprises calculating a maximum value of the degrees of overlap obtained respectively between each Boxd num t and all the Boxm id t , if the O max ′ is less than a corresponding threshold θ 2 , considering the second ship target as the ship appeared at the moment t, and adding the second ship target into a tracking result of the moment t to obtain a complete tracking result set.

2. The automatic ship tracking method based on the deep learning network and the mean shift according to claim 1 , wherein a calculation formula of the degree of overlap is as follows:

O

=

S

⁡

(

Boxm

id

t

⋂

Boxd

num

t

)

S

⁡

(

Boxm

id

t

⋃

Boxd

num

t

)

wherein, S represents an area size.

3. An automatic ship tracking system based on a deep learning network and a mean shift, comprising:

a first module, wherein the first module is configured to collect surveillance video data and extract a video image from the surveillance video data, and the surveillance video data comprises coastal region surveillance video data under visible light;

a second module configured to perform preprocessing based on the video image obtained by the first module to extract a positive sample and a negative sample of a first ship target;

a third module configured to input the positive sample and the negative sample of the first ship target in the video image into a neural network to train a model by a region-based convolutional neural network method;

a fourth module configured to extract initial frame data of the video image and perform a ship detection and a probability density calculation on the initial frame data according to the model obtained by the third module, wherein the probability density calculation is realized by dividing a gray color space of a target region to obtain a gray histogram formed by multiple equal histogram intervals, and calculating the probability density according to one histogram interval of the multiple equal histogram intervals, a gray value of a pixel in the target region belongs to the one histogram interval; and

a fifth module configured to determine a ship tracking result at a current moment by a calculation result of a previous moment, comprising the following steps:

taking ξ ship positions tracked at a moment t−1 as initial positions, respectively taking a center coordinate f 0 of each of the ξ ship positions as an initial target position of a ship tracked at a moment t, taking the f 0 as a center of a search window to obtain a center position coordinate f of a corresponding candidate ship, calculating a region histogram of a candidate position, and further realizing the probability density calculation;

describing a degree of similarity between a ship model and the corresponding candidate ship by a Bhattacharyya coefficient, calculating a mean shift iteration equation in a center of the target region, so that the ship model moves continuously towards a direction with a maximum color change until last two moving distances are less than a corresponding preset threshold, finding a ship position obtained by a mean shift result at the moment t, setting that multiple ship positions Boxm t are obtained, and expressing an id th ship position as Boxm id t ;

performing the ship detection on the video image at the moment t by the region-based convolutional neural network method, setting that a num th detection coordinate of multiple ships in the video image at the moment t is Boxd num t , by the region-based convolutional neural network method, performing the ship detection on the video image at the moment t, and obtaining the num th detection coordinate Boxd num t of multiple ships in the video image at the moment t, calculating a degree of overlap between the Boxd num t and the Boxm id t , recording a value of degree of overlap O max between each Boxm id t and the corresponding Boxd num t with a maximum degree of overlap, and if the O max is less than a corresponding threshold θ 1 , considering the ship position as false-alarm and deleting the Boxm id t ; and

updating a second ship target appeared at the moment t by a neural network detection result, wherein the updating further comprises calculating a maximum value of the degrees of overlap obtained respectively between each Boxd num t and all the Boxm id t , if the O max ′ is less than a corresponding threshold θ 2 , considering the second ship target as the ship appeared at the moment t, and adding the second ship target into a tracking result of the moment t to obtain a complete tracking result set.

4. The automatic ship tracking system based on deep learning network and mean shift according to claim 3 , wherein a calculation formula of the degree of overlap is as follows:

O

=

S

⁡

(

Boxm

id

t

⋂

Boxd

num

t

)

S

⁡

(

Boxm

id

t

⋃

Boxd

num

t

)

wherein, S represents an area size.