CN118053091A - Infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm - Google Patents
Infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm Download PDFInfo
- Publication number
- CN118053091A CN118053091A CN202311753567.1A CN202311753567A CN118053091A CN 118053091 A CN118053091 A CN 118053091A CN 202311753567 A CN202311753567 A CN 202311753567A CN 118053091 A CN118053091 A CN 118053091A
- Authority
- CN
- China
- Prior art keywords
- layer
- infrared
- target
- ship
- yolov
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 66
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000000605 extraction Methods 0.000 claims abstract description 39
- 238000012549 training Methods 0.000 claims abstract description 21
- 238000001514 detection method Methods 0.000 claims description 55
- 238000004088 simulation Methods 0.000 claims description 45
- 230000004927 fusion Effects 0.000 claims description 37
- 238000011176 pooling Methods 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 4
- 101100441251 Arabidopsis thaliana CSP2 gene Proteins 0.000 claims description 3
- 102100027557 Calcipressin-1 Human genes 0.000 claims description 3
- 101100247605 Homo sapiens RCAN1 gene Proteins 0.000 claims description 3
- 101100112467 Rattus norvegicus Cblc gene Proteins 0.000 claims description 3
- 101150064416 csp1 gene Proteins 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 7
- 238000012795 verification Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- RTAQQCXQSZGOHL-UHFFFAOYSA-N Titanium Chemical compound [Ti] RTAQQCXQSZGOHL-UHFFFAOYSA-N 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 239000000779 smoke Substances 0.000 description 1
Landscapes
- Aiming, Guidance, Guns With A Light Source, Armor, Camouflage, And Targets (AREA)
Abstract
The invention discloses an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, which is used for acquiring an optimal model by using a self-built dataset for YOLOv s network training and Deepsort algorithm depth feature extraction network training on infrared ship video, so as to be used for tracking models in Deepsort algorithm and complete multi-target tracking. The invention can successfully track a plurality of ships repeatedly appearing and advancing in the video by utilizing YOLOv s in combination with Deepsort algorithm, and can timely re-track the target when the ship target is re-appearing after being shielded, thereby realizing stable and continuous multi-target tracking.
Description
Technical Field
The invention belongs to the technical field of infrared target detection tracking, and particularly relates to an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm.
Background
With the gradual development of the deep learning technology, the deep learning algorithm is widely applied to the fields of target detection, identification and tracking. In the infrared target tracking field, the difficulty of multi-target tracking is higher compared with that of a single-target tracking algorithm, and a better detector is introduced to combine for target tracking in order to improve the tracking effect of the algorithm. On the premise of good detection effect, the performance of the tracking algorithm can be effectively improved.
Disclosure of Invention
The technical solution of the invention is as follows: the infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm is provided, and aiming at infrared simulation video, a plurality of ships repeatedly appearing and advancing in the video can be successfully tracked, and the targets can be timely re-tracked when the ship targets are re-appearing after being blocked, so that stable and continuous multi-target tracking is realized
In order to solve the technical problems, the invention discloses an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, which comprises the following steps:
Constructing an infrared ship data set;
taking the infrared ship data set as input of YOLOv s detection network, training YOLOv s detection network, and obtaining a detection model;
Constructing an infrared ship cutting data set;
Taking the infrared ship cutting data set as input of a depth feature extraction network, training the depth feature extraction network, and obtaining a depth feature extraction model;
And inputting the detection model and the depth feature extraction model into a Deepsort model, outputting a tracking result, and completing tracking of the infrared multi-ship target.
In the above-mentioned infrared multi-target tracking method based on YOLOv s in combination with Deepsort algorithm, constructing an infrared ship dataset includes:
Obtaining infrared simulation images of N ship targets through laboratory simulation to form an infrared simulation image data set; wherein N is more than or equal to 100;
And (3) carrying out image horizontal and vertical overturning on each infrared simulation image in the infrared simulation image data set, rotating clockwise by 45 degrees and 90 degrees, and increasing Gaussian white noise of the image so as to realize expansion of the infrared simulation image data set and obtain the infrared ship data set.
In the infrared multi-target tracking method based on YOLOv s and Deepsort algorithm, the network structure of the YOLOv s detection network comprises four parts connected in sequence: the system comprises an input end, a backbone network, a branch network and an output end;
The output end comprises: slicing and splicing the layer Focus;
The backbone network comprises the following components: standard volume base cbl_1, convolution and one residual structure fusion layer csp1_1, standard volume base cbl_2, convolution and one residual structure fusion layer csp1_2, standard volume base cbl_3, convolution and one residual structure fusion layer csp1_3, standard volume base cbl_4 and spatial pyramid pooling layer SPP;
The backbone network comprises the following components: the convolution and two residual structure fusion layers csp2_1, the standard volume base layer cbl_5, the first upsampling layer, the splice layer Concat _1, the convolution and two residual structure fusion layers csp2_2, the standard volume base layer cbl_6, the second upsampling layer, the splice layer Concat _2, the convolution and two residual structure fusion layers csp2_3, the standard convolution layer cbl_7, the splice layer Concat _3, the convolution and two residual structure fusion layers csp2_4, the standard convolution layer cbl_8, the splice layer Concat _4, the convolution and two residual structure fusion layers csp2_5;
the output end comprises: three different scale convolutional layers CONV: convolutional layer conv_1, convolutional layer conv_2, and convolutional layer conv_3.
In the infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm,
The slice splicing layer Focus consists of a cutting layer, a characteristic splicing layer and a convolution layer;
the standard convolution layer CBL consists of a convolution layer, a normalization layer and an activation function layer; wherein the activation function is LeakyReLU functions;
the convolution and residual structure fusion layer CSP1 consists of a standard convolution layer and a residual structure;
The convolution sum and two residual structure fusion layers CSP2 consist of a standard convolution layer and two residual structures;
The space pyramid pooling layer SPP is formed by splicing a plurality of maximum pooling layers; the SPP layer adopts a 1*1, 5*5, 9*9 and 13 x 13 maximum pooling mode to carry out multi-scale fusion;
The dimensions of the convolution layers conv_1, conv_2 and conv_3 are respectively: 76 x 76, 38 x 38 and 19 x 19.
In the infrared multi-target tracking method based on YOLOv s and Deepsort algorithm, YOLOv s detection network cuts an input infrared simulation image into a feature map with 640 x 640 pixels through slice splicing layer Focus by adopting slice operation; the YOLOv s detection network output is a characteristic map of 76 x 76, 38 x 38 and 19 x 19 pixel sizes.
In the above-mentioned infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, constructing an infrared ship cutting dataset includes:
manually labeling ship targets contained in each infrared simulation image in the infrared ship data set to obtain an infrared simulation image carrying labels;
And cutting each infrared simulation image carrying the labels into an infrared image of a single ship target according to the coordinates of the labels, and obtaining an infrared ship cutting data set.
In the above-mentioned infrared multi-target tracking method based on YOLOv s and Deepsort algorithm, taking the infrared ship cutting dataset as the input of the depth feature extraction network, training the depth feature extraction network to obtain a depth feature extraction model, comprising:
dividing a plurality of single ship target infrared images in an infrared ship cutting data set into a training set and a testing set according to the proportion of 8:2;
and inputting the training set and the testing set into a depth feature extraction network for training to obtain a depth feature extraction model.
In the above-mentioned infrared multi-target tracking method based on YOLOv s and Deepsort algorithm, the network structure of the depth feature extraction network includes sequentially connected: the system comprises an input layer, a layer 1 convolution layer, a layer 2 convolution layer, a maximum pooling layer, a residual error module and an average pooling layer; the pixel size of the input image of the depth feature extraction network is 64×128; the output is a 128-dimensional feature vector.
In the above-mentioned infrared multi-target tracking method based on YOLOv s and Deepsort algorithm, the detection model and the depth feature extraction model are input into the Deepsort model, and the tracking result is output, so as to complete the tracking of the infrared multi-ship target, which comprises the following steps:
Tracking information prediction is carried out on the ship target by adopting a Kalman filtering algorithm, so as to obtain tracking information;
Acquiring target detection information output by a detection model;
Matching the target detection information with the tracking information by adopting a Hungary matching algorithm, and updating the target detection information according to a depth feature extraction model;
And for the situation that tracking information prediction of the ship target is uncertain by adopting a Kalman filtering algorithm when the target is shielded for a long time, repositioning and tracking the lost target by using cascade matching.
In the infrared multi-target tracking method based on YOLOv s and Deepsort algorithm, when the Hungary matching algorithm is adopted to match target detection information with tracking information, the adopted matching mode adopts fusion measurement of motion information association and target apparent information association.
The invention has the following advantages:
(1) The invention discloses an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, which uses YOLOv s for infrared multi-target detection and effectively improves infrared target detection accuracy.
(2) The invention discloses an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, which uses YOLOv s as a detector and combines Deepsort algorithm to perform infrared multi-target tracking, thus realizing stable infrared multi-target tracking.
Drawings
FIG. 1 is a flow chart of steps of an infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm in an embodiment of the invention;
FIG. 2 is a schematic diagram of an implementation of an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm in an embodiment of the present invention;
FIG. 3 is a network configuration diagram of a YOLOv s detection network in accordance with one embodiment of the present invention;
FIG. 4 is a network architecture diagram of a depth feature extraction network in accordance with an embodiment of the present invention;
FIG. 5 is a partial screenshot of an infrared simulation video for validating a tracking algorithm in an embodiment of the invention;
FIG. 6 is a schematic diagram of a verification result using YOLOv and Deepsort algorithms in accordance with one embodiment of the present invention;
FIG. 7 is a schematic diagram of a verification result using the algorithm of the present invention in an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the embodiments of the present invention disclosed herein will be described in further detail with reference to the accompanying drawings.
One of the core ideas of the invention is: the infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm is provided, and the method is characterized in that: constructing an infrared ship data set (warship) for training; YOLOv5 was used as detector (detection model); cutting the infrared image according to the detected marking frame to obtain an infrared ship cutting data set (crop) consisting of single ship targets; training Deepsort the depth feature extraction network by using the infrared ship cutting data set; and combining a detector (detection model) and a depth network model, and performing multi-target tracking in the infrared ship video by using Deepsort algorithm to finish stable and continuous tracking.
Referring to fig. 1 and 2, in the present embodiment, the infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm includes:
And 1, constructing an infrared ship data set.
In this embodiment, first, through laboratory simulation, an infrared simulation image of N ship targets is obtained, and an infrared simulation image dataset is formed. Then, each infrared simulation image in the infrared simulation image data set is subjected to image horizontal and vertical overturning, and is rotated clockwise by 45 degrees and 90 degrees, so that Gaussian white noise of the image is increased, expansion of the infrared simulation image data set is realized, and an infrared ship data set is obtained. Wherein, N is more than or equal to 100, and the larger the value of N is, the higher the accuracy of YOLOv s detection network is.
And 2, taking the infrared ship data set as input of YOLOv s detection network, training YOLOv s detection network, and obtaining a detection model.
In this embodiment, as shown in fig. 2, the network structure of the YOLOv s detection network includes four parts connected in sequence: input, backbone network and output. Wherein, the output includes: slice splicing layer Focus. The backbone network comprises the following components: standard volume base layer cbl_1, convolution and one residual structure fusion layer csp1_1, standard volume base layer cbl_2, convolution and one residual structure fusion layer csp1_2, standard volume base layer cbl_3, convolution and one residual structure fusion layer csp1_3, standard volume base layer cbl_4 and spatial pyramid pooling layer SPP. The backbone network comprises the following components: the convolution and two residual structure fusion layers csp2_1, the standard volume base layer cbl_5, the first upsampling layer, the splice layer Concat _1, the convolution and two residual structure fusion layers csp2_2, the standard volume base layer cbl_6, the second upsampling layer, the splice layer Concat _2, the convolution and two residual structure fusion layers csp2_3, the standard convolution layer cbl_7, the splice layer Concat _3, the convolution and two residual structure fusion layers csp2_4, the standard convolution layer cbl_8, the splice layer Concat _4, the convolution and two residual structure fusion layers csp2_5. The output end comprises: three different scale convolutional layers CONV: convolutional layer conv_1, convolutional layer conv_2, and convolutional layer conv_3.
Further, the slice splicing layer Focus is composed of a clipping layer, a characteristic splicing layer and a convolution layer. The standard convolution layer CBL consists of a convolution layer, a normalization layer and an activation function layer, wherein the activation function is LeakyReLU functions. The convolution and one residual structure fusion layer CSP1 consists of a standard convolution layer and one residual structure. The convolution and two residual structure fusion layer CSP2 consists of a standard convolution layer and two residual structures. The SPP layer is formed by splicing a plurality of largest pooling layers, and the SPP layer adopts a largest pooling mode of 1*1, 5*5, 9*9 and 13 x 13 to carry out multi-scale fusion. The dimensions of the convolution layers conv_1, conv_2 and conv_3 are respectively: 76 x 76, 38 x 38 and 19 x 19.
Further, YOLOv s detection network cuts the input infrared simulation image into a feature map with 640 pixels by slice operation through slice splicing layer Focus; the YOLOv s detection network outputs 256, 512 and 1024 feature vectors, and the pixel sizes are 19×19, 38×38 and 76×76 respectively.
And 3, constructing an infrared ship cutting data set.
In this embodiment, first, a ship target included in each infrared simulation image in an infrared ship data set is manually marked, so as to obtain an infrared simulation image carrying the mark. Then, each infrared simulation image carrying the labels is cut into a single ship target infrared image according to the coordinates of the labels, and an infrared ship cutting data set is obtained.
For example, there are 2043 infrared simulation images in the infrared ship data set, each infrared simulation image contains one to four different ship targets (the infrared ship data set also contains part of weak ship targets, ship targets under island background and sea background, and infrared simulation images with infrared smoke curtains), and the ship targets contained in each infrared simulation image in the infrared ship data set are manually marked to obtain an infrared simulation image carrying the mark; cutting each infrared simulation image carrying the labels into single ship target infrared images according to the coordinates of the labels, cutting 5308 single ship target infrared images altogether, and obtaining an infrared ship cutting data set.
And 4, taking the infrared ship cutting data set as input of a depth feature extraction network, training the depth feature extraction network, and obtaining a depth feature extraction model.
In this embodiment, in the tracking section, deepsort algorithm adds a depth feature extraction network for extracting apparent depth features of the target. The infrared ship target infrared images of a plurality of single ships in the infrared ship cutting data set can be divided into a training set and a testing set according to the proportion of 8:2. And then, inputting the training set and the testing set into a depth feature extraction network for training to obtain a depth feature extraction model.
Further, as shown in fig. 3, the network structure of the depth feature extraction network includes sequentially connected: input layer, layer 1 convolution layer, layer 2 convolution layer, max pooling layer, residual structure, and average pooling layer. The pixel size of the input image of the depth feature extraction network is 64×128; the output is a 128-dimensional feature vector.
And 5, inputting the detection model and the depth feature extraction model into a Deepsort model, outputting a tracking result, and completing tracking of the infrared multi-ship target.
In this embodiment, the multi-target tracking employs Deepsort algorithm. Firstly, tracking information prediction is carried out on a ship target by adopting a Kalman filtering algorithm to obtain tracking information; and acquiring target detection information output by the detection model. Then, a Hungary matching algorithm is adopted to match the target detection information with the tracking information, and the target detection information is updated according to the depth feature extraction model. Finally, for the situation that tracking information prediction of ship targets is uncertain by adopting a Kalman filtering algorithm when the targets are shielded for a long time, cascade matching is used for repositioning and tracking lost targets.
Furthermore, when the Hungary matching algorithm is adopted to match the target detection information with the tracking information, the adopted matching mode adopts the fusion measurement of the motion information association and the target apparent information association.
Preferably, the motion information association uses a horse-type distance:
Wherein d (1) (i, j) represents a mahalanobis distance between the i-th tracking information and the j-th target detection information; d j denotes the jth target detection information, y i denotes the ith tracking information, and S i denotes the covariance matrix of the observation space.
Preferably, the target apparent information association uses a minimum cosine distance:
Wherein d (2) (i, j) represents the minimum cosine distance between the ith tracking information and the jth target detection information; r j denotes the j-th detection apparent information, The i-th tracking appearance information representing the kth track, R i represents the tracking appearance information set.
The fusion metric c i,j is formulated as follows:
ci,j＝λd(1)(i,j)+(1-λ)d(2)(i,j)
Where λ represents a metric factor. If c i,j is within the intersection of the two associated distance thresholds, it indicates that the tracking frame and the detection frame are correctly matched.
Based on the above embodiments, the effects of the present invention will be further described in conjunction with simulation experiments.
Simulation conditions: the simulation experiment condition of the invention is that a software environment of python3.7, pytorch1.6.0 and torchvision0.7.0 is constructed on a single display card of TITAN V12 GB model.
Simulation content and result analysis
The simulation results of the invention are shown in fig. 5-7, fig. 5 is a partial screenshot of an infrared simulation video for verifying a tracking algorithm, fig. 6 is a verification result adopting YOLOv and Deepsort algorithms, and fig. 7 is a verification result adopting the algorithm of the invention. Comparing fig. 6 and fig. 7, when the same infrared data set is used for verification, when different detectors are used for detecting infrared ship targets, the detection precision of two algorithms can be compared through the mAP value, and the higher the mAP value is, the higher the precision is. Compared with the result of the same frame in the infrared simulation video, the method has better tracking effect on the infrared multiple targets. Therefore, the method of the invention overcomes the problems in the prior art, can successfully track a plurality of ships repeatedly appearing and advancing in video, and can timely re-track the target when the ship target is re-appearing after being blocked, thereby realizing stable and continuous multi-target tracking.
In summary, the invention discloses an infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, which adopts a classical YOLO algorithm in a deep learning algorithm for target detection, has the advantages of higher detection speed and better guaranteeing precision and recognition effect, while YOLOv5 is the latest algorithm in the YOLO series, and the verification effect on a public data set is superior to other YOLO algorithms, so YOLOv is selected as a detector used in multi-target tracking. The DeepSort algorithm is used for target tracking, the apparent information of the target is extracted to carry out nearest neighbor matching while matching calculation is carried out, and a deep learning model which is trained offline on a multi-target tracking dataset is introduced, so that the target tracking effect under the condition of shielding can be improved and the problem of target ID jump is also reduced in the real-time target tracking process.
Although the present invention has been described in terms of the preferred embodiments, it is not intended to be limited to the embodiments, and any person skilled in the art can make any possible variations and modifications to the technical solution of the present invention by using the methods and technical matters disclosed above without departing from the spirit and scope of the present invention, so any simple modifications, equivalent variations and modifications to the embodiments described above according to the technical matters of the present invention are within the scope of the technical matters of the present invention.
What is not described in detail in the present specification belongs to the known technology of those skilled in the art.
Claims (10)
1. An infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm, which is characterized by comprising the following steps:
Constructing an infrared ship data set;
taking the infrared ship data set as input of YOLOv s detection network, training YOLOv s detection network, and obtaining a detection model;
Constructing an infrared ship cutting data set;
Taking the infrared ship cutting data set as input of a depth feature extraction network, training the depth feature extraction network, and obtaining a depth feature extraction model;
And inputting the detection model and the depth feature extraction model into a Deepsort model, outputting a tracking result, and completing tracking of the infrared multi-ship target.
2. The method for infrared multi-target tracking based on YOLOv s combined with Deepsort algorithm according to claim 1, wherein constructing an infrared ship dataset comprises:
Obtaining infrared simulation images of N ship targets through laboratory simulation to form an infrared simulation image data set; wherein N is more than or equal to 100;
And (3) carrying out image horizontal and vertical overturning on each infrared simulation image in the infrared simulation image data set, rotating clockwise by 45 degrees and 90 degrees, and increasing Gaussian white noise of the image so as to realize expansion of the infrared simulation image data set and obtain the infrared ship data set.
3. The infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm according to claim 1, wherein the network structure of the YOLOv s detection network comprises four parts connected in sequence: the system comprises an input end, a backbone network, a branch network and an output end;
The output end comprises: slicing and splicing the layer Focus;
The backbone network comprises the following components: standard volume base cbl_1, convolution and one residual structure fusion layer csp1_1, standard volume base cbl_2, convolution and one residual structure fusion layer csp1_2, standard volume base cbl_3, convolution and one residual structure fusion layer csp1_3, standard volume base cbl_4 and spatial pyramid pooling layer SPP;
The backbone network comprises the following components: the convolution and two residual structure fusion layers csp2_1, the standard volume base layer cbl_5, the first upsampling layer, the splice layer Concat _1, the convolution and two residual structure fusion layers csp2_2, the standard volume base layer cbl_6, the second upsampling layer, the splice layer Concat _2, the convolution and two residual structure fusion layers csp2_3, the standard convolution layer cbl_7, the splice layer Concat _3, the convolution and two residual structure fusion layers csp2_4, the standard convolution layer cbl_8, the splice layer Concat _4, the convolution and two residual structure fusion layers csp2_5;
the output end comprises: three different scale convolutional layers CONV: convolutional layer conv_1, convolutional layer conv_2, and convolutional layer conv_3.
4. The method for infrared multi-target tracking based on YOLOv s combined with Deepsort algorithm according to claim 3,
The slice splicing layer Focus consists of a cutting layer, a characteristic splicing layer and a convolution layer;
the standard convolution layer CBL consists of a convolution layer, a normalization layer and an activation function layer; wherein the activation function is LeakyReLU functions;
the convolution and residual structure fusion layer CSP1 consists of a standard convolution layer and a residual structure;
The convolution sum and two residual structure fusion layers CSP2 consist of a standard convolution layer and two residual structures;
The space pyramid pooling layer SPP is formed by splicing a plurality of maximum pooling layers; the SPP layer adopts a 1*1, 5*5, 9*9 and 13 x 13 maximum pooling mode to carry out multi-scale fusion;
The dimensions of the convolution layers conv_1, conv_2 and conv_3 are respectively: 76 x 76, 38 x 38 and 19 x 19.
5. The infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm according to claim 3, wherein YOLOv s detection network uses slice operation to cut the input infrared simulation image into a feature map with 640 x 640 pixels size through slice splicing layer Focus; the YOLOv s detection network output is a characteristic map of 76 x 76, 38 x 38 and 19 x 19 pixel sizes.
6. The infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm according to claim 2, wherein constructing the infrared ship clipping dataset comprises:
manually labeling ship targets contained in each infrared simulation image in the infrared ship data set to obtain an infrared simulation image carrying labels;
And cutting each infrared simulation image carrying the labels into an infrared image of a single ship target according to the coordinates of the labels, and obtaining an infrared ship cutting data set.
7. The method of claim 6, wherein taking the infrared ship cut dataset as input to a depth feature extraction network, training the depth feature extraction network to obtain a depth feature extraction model comprises:
dividing a plurality of single ship target infrared images in an infrared ship cutting data set into a training set and a testing set according to the proportion of 8:2;
and inputting the training set and the testing set into a depth feature extraction network for training to obtain a depth feature extraction model.
8. The infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm according to claim 7, wherein the network structure of the depth feature extraction network includes sequentially connected: the system comprises an input layer, a layer 1 convolution layer, a layer 2 convolution layer, a maximum pooling layer, a residual error module and an average pooling layer; the pixel size of the input image of the depth feature extraction network is 64×128; the output is a 128-dimensional feature vector.
9. The infrared multi-target tracking method based on YOLOv s combined with Deepsort algorithm according to claim 1, wherein the steps of inputting a detection model and a depth feature extraction model into a Deepsort model, outputting a tracking result, and completing tracking of an infrared multi-ship target include:
Tracking information prediction is carried out on the ship target by adopting a Kalman filtering algorithm, so as to obtain tracking information;
Acquiring target detection information output by a detection model;
Matching the target detection information with the tracking information by adopting a Hungary matching algorithm, and updating the target detection information according to a depth feature extraction model;
And for the situation that tracking information prediction of the ship target is uncertain by adopting a Kalman filtering algorithm when the target is shielded for a long time, repositioning and tracking the lost target by using cascade matching.
10. The method for infrared multi-target tracking based on YOLOv s combined with Deepsort algorithm according to claim 9, wherein when the hungarian matching algorithm is used to match the target detection information with the tracking information, the matching method is used to use a fusion metric of both the motion information association and the target apparent information association.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311753567.1A CN118053091A (en) | 2023-12-18 | 2023-12-18 | Infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311753567.1A CN118053091A (en) | 2023-12-18 | 2023-12-18 | Infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN118053091A true CN118053091A (en) | 2024-05-17 |
Family
ID=91049110
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311753567.1A Pending CN118053091A (en) | 2023-12-18 | 2023-12-18 | Infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118053091A (en) |
-
2023
- 2023-12-18 CN CN202311753567.1A patent/CN118053091A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20230045519A1 (en) | Target Detection Method and Apparatus | |
| EP3633615A1 (en) | Deep learning network and average drift-based automatic vessel tracking method and system | |
| CN110659582A (en) | Image conversion model training method, heterogeneous face recognition method, device and equipment | |
| CN110647802A (en) | Remote sensing image ship target detection method based on deep learning | |
| CN113361636B (en) | Image classification method, system, medium and electronic device | |
| Sun et al. | Global Mask R-CNN for marine ship instance segmentation | |
| Zhang et al. | A object detection and tracking method for security in intelligence of unmanned surface vehicles | |
| CN111640138A (en) | Target tracking method, device, equipment and storage medium | |
| CN111429481B (en) | Target tracking method, device and terminal based on adaptive expression | |
| CN111598155A (en) | Fine-grained image weak supervision target positioning method based on deep learning | |
| CN111415370A (en) | Embedded infrared complex scene target real-time tracking method and system | |
| CN114066937B (en) | Multi-target tracking method for large-scale remote sensing image | |
| CN116416503A (en) | Small sample target detection method, system and medium based on multi-mode fusion | |
| CN115730263A (en) | Ship behavior pattern detection method and device | |
| Dai et al. | GCD-YOLOv5: An armored target recognition algorithm in complex environments based on array lidar | |
| CN114038011A (en) | Method for detecting abnormal behaviors of human body in indoor scene | |
| Yang et al. | Foreground enhancement network for object detection in sonar images | |
| CN116844055A (en) | Lightweight SAR ship detection method and system | |
| Yao et al. | An accurate box localization method based on rotated-RPN with weighted edge attention for bin picking | |
| CN118053091A (en) | Infrared multi-target tracking method based on YOLOv s combined Deepsort algorithm | |
| Devyatkin et al. | Neural network traffic signs detection system development | |
| CN115359091A (en) | Armor plate detection tracking method for mobile robot | |
| CN114743257A (en) | Method for detecting and identifying image target behaviors | |
| CN116152334A (en) | Image processing method and related equipment | |
| CN112529072A (en) | Underwater buried object identification and positioning method based on sonar image processing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination |