An image segmentation method based on tracking targets

Abstract

The application discloses a kind of based on tracking target's image segmentation method, belong to artificial intelligence field, comprising: obtaining first image frame sequence;First image frame sequence is sequentially input into the first feature extraction channel of target extraction model, obtains first feature extraction result;The template image of to-be-identified target is input into the second feature extraction channel of target extraction model, obtains second feature extraction result;First feature extraction result and second feature extraction result are input into the feature comparison channel of target extraction model, obtains target identification result;According to target identification result, target identification is carried out to first image frame sequence, obtains second image frame sequence;According to second image frame sequence, generate target tracking path.The present application solves the technical problems of low target recognition accuracy and poor tracking stability in the prior art, and achieves the technical effects of improving target recognition accuracy and tracking stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and particularly relates to an image segmentation method based on tracking targets. BACKGROUND

[0002] The key of target recognition and tracking is to accurately and efficiently extract the features of input images and target templates for matching. In the prior art, different feature extraction networks are usually used to extract input image features and target template features respectively. However, this method cannot achieve ideal target recognition accuracy and target tracking stability due to the mismatched expression of the two kinds of features. SUMMARY

[0003] The present application provides an image segmentation method based on tracking targets, aiming to solve the technical problems of low target recognition accuracy and poor tracking stability in the prior art.

[0004] In view of the above problems, the present application provides an image segmentation method based on tracking targets.

[0005] The first aspect of the present application provides an image segmentation method based on tracking targets, which comprises: acquiring a first image frame sequence; inputting the first image frame sequence into a first feature extraction channel of a target extraction model in turn to obtain a first feature extraction result; inputting a target template image to be recognized into a second feature extraction channel of the target extraction model to obtain a second feature extraction result, wherein the first feature extraction channel and the second feature extraction channel have the same network structure and network parameters; inputting the first feature extraction result and the second feature extraction result into a feature comparison channel of the target extraction model to obtain a target recognition result; performing target identification on the first image frame sequence according to the target recognition result to obtain a second image frame sequence; and generating a target tracking path according to the second image frame sequence.

[0006] In another aspect of the present application, a target tracking-based image segmentation system is provided, which comprises: an image sequence acquisition module configured to acquire a first image frame sequence; a first feature extraction module configured to sequentially input the first image frame sequence into a first feature extraction channel of a target extraction model to obtain a first feature extraction result; a second feature extraction module configured to input a target template image to be recognized into a second feature extraction channel of the target extraction model to obtain a second feature extraction result, wherein the first feature extraction channel and the second feature extraction channel have the same network structure and network parameters; a feature comparison channel module configured to input the first feature extraction result and the second feature extraction result into a feature comparison channel of the target extraction model to obtain a target recognition result; an image target identification module configured to identify targets in the first image frame sequence according to the target recognition result to obtain a second image frame sequence; and a tracking path generation module configured to generate a target tracking path according to the second image frame sequence.

[0007] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0008] Thanks to the technical solutions that the first image frame sequence is acquired, the first image frame sequence is input into the first feature extraction channel of the target extraction model to obtain the first feature extraction result, the target template image is input into the second feature extraction channel of the target extraction model to obtain the second feature extraction result, the first feature extraction result and the second feature extraction result are input into the feature comparison channel of the target extraction model to obtain the target recognition result, and the first image frame sequence is identified according to the target recognition result to obtain the second image frame sequence, the technical problems of low target recognition accuracy and poor tracking stability in the prior art are solved, and the technical effects of improving the target recognition accuracy and the tracking stability are achieved.

[0009] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the present application can be implemented in accordance with the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A possible flowchart of a target tracking-based image segmentation method is provided for the embodiments of the present application;

[0011] Figure 2 A possible flowchart of acquiring a feature extraction channel in a target tracking-based image segmentation method is provided for the embodiments of the present application;

[0012] Figure 3A possible flowchart of a feature comparison channel in a tracking target-based image segmentation method is provided for the embodiments of the present application.

[0013] Figure 4 A possible structure diagram of a tracking target-based image segmentation system is provided for the embodiments of the present application.

[0014] Legend of reference signs: image sequence acquisition module 11, first feature extraction module 12, second feature extraction module 13, feature comparison channel module 14, image target identification module 15, and tracking path generation module 16. DETAILED DESCRIPTION

[0015] The general idea of the technical solutions provided by the present application is as follows:

[0016] The embodiments of the present application provide a tracking target-based image segmentation method. First, a first image frame sequence is acquired. Then, the first image frame sequence is input into a first feature extraction channel of a target extraction model to obtain a first feature extraction result, i.e., an input image feature. Meanwhile, a target template image is input into a second feature extraction channel of the target extraction model to obtain a second feature extraction result, i.e., a target template feature. Since the first feature extraction channel and the second feature extraction channel have the same network structure and parameters, the two feature expression methods are matched. Next, the first feature extraction result and the second feature extraction result are input into a feature comparison channel of the target extraction model to obtain a target identification result, i.e., a matching result of the input image and the target template. Then, the first image frame sequence is identified according to the target identification result to obtain a second image frame sequence, i.e., an image sequence with a target label. Finally, a target tracking path is generated according to the second image frame sequence, i.e., a motion trajectory of the target in the image sequence.

[0017] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be specifically introduced in combination with the drawings of the specification.

[0018] Embodiment One

[0019] As shown in Figure 1 The embodiments of the present application provide a tracking target-based image segmentation method, which includes the following steps.

[0020] Step S10000: acquiring a first image frame sequence;

[0021] The first image frame sequence refers to a target image sequence to be processed arranged in time. The first image frame sequence is composed of multiple continuous frame image data of a tracking target acquired by image acquisition, and is used as the input of a subsequent target extraction model.

[0022] The image data collected by the image collection device includes a plurality of continuous frames, each frame corresponding to a collected image, and the image data of the plurality of continuous frames constitutes a first image frame sequence as an input of image segmentation to provide input image data for subsequent image processing.

[0023] Step S20000: sequentially input the first image frame sequence into a first feature extraction channel of a target extraction model to obtain a first feature extraction result.

[0024] The target extraction model refers to a convolutional neural network model used for target recognition and tracking. The target extraction model includes a first feature extraction channel, a second feature extraction channel, and a feature comparison channel. The first feature extraction channel is used for feature extraction of sequentially input images, the second feature extraction channel is used for feature extraction of target template images, and the feature comparison channel is used for feature comparison to achieve target recognition.

[0025] The obtained first image frame sequence is input as an input, and each image frame in the first image frame sequence is input into the first feature extraction channel in the time order. The first feature extraction channel extracts features from each input image frame to obtain the corresponding first feature extraction result. The first feature extraction channel is composed of a convolutional neural network. When the first image frame is input into the first feature extraction channel, the convolutional neural network automatically learns the features such as contours and edges in the image frame to obtain the first feature extraction result corresponding to the first image frame. Then, the second image frame is input into the first feature extraction channel, and the convolutional neural network continues to learn to obtain the first feature extraction result corresponding to the second image frame. Similarly, until all image frames in the first image frame sequence are input into the first feature extraction channel and the corresponding first feature extraction result is obtained, the feature information is provided for subsequent feature comparison and target recognition.

[0026] Step S30000: input the target template image to be recognized into the second feature extraction channel of the target extraction model to obtain a second feature extraction result, wherein the first feature extraction channel and the second feature extraction channel have the same network structure and network parameters.

[0027] The to-be-identified target template image refers to a target reference image used for target identification. The second feature extraction channel of the target extraction model has the same network structure and network parameters as the first feature extraction channel, so that the second feature extraction channel has the same feature extraction capability as the first feature extraction channel. The structure ensures that the same feature extraction results can be obtained by the two feature extraction channels when the same image is input. The to-be-identified target template image is input into the second feature extraction channel of the target extraction model. Like the first feature extraction channel, the second feature extraction channel is also composed of a convolutional neural network. When the to-be-identified target template image is input into the second feature extraction channel, the convolutional neural network automatically learns the same features in the to-be-identified target template image because the network structure and network parameters are the same as those of the first feature extraction channel, and obtains the second feature extraction result corresponding to the to-be-identified target template image, thereby providing feature reference information for subsequent feature comparison and target identification. The same feature extraction channel structure can ensure that the feature information in the two feature extraction results is comparable, thereby facilitating feature comparison and target identification.

[0028] Step S40000: inputting the first feature extraction result and the second feature extraction result into a feature comparison channel of the target extraction model to obtain a target identification result;

[0029] The feature comparison channel refers to a channel used for feature comparison and target identification. The feature comparison channel receives the first feature extraction result and the second feature extraction result as input, and judges whether the two feature extraction results are similar through feature comparison to achieve target identification.

[0030] The obtained first feature extraction result and the obtained second feature extraction result are input into the feature comparison channel of the target extraction model. The feature comparison channel performs feature comparison on the first feature extraction result and the second feature extraction result, and judges whether the two feature extraction results are similar. If they are similar, it means that the image frames in the first image frame sequence contain the to-be-identified target, and the target identification result obtained is similar. Otherwise, it means that the image frames in the first image frame sequence do not contain the to-be-identified target, and the target identification result obtained is not similar.

[0031] The target identification result is obtained by judging whether the image frames in the first image frame sequence contain the to-be-identified target through feature comparison, thereby achieving target identification and providing a judgment basis for subsequent image segmentation and target tracking.

[0032] Step S50000: performing target identification on the first image frame sequence according to the target identification result to obtain a second image frame sequence;

[0033] Target identification refers to marking the position of a target in an image frame and framing the target. The second image frame sequence refers to the image frame sequence after target identification.

[0034] If the target recognition result is similar, it indicates that the image frames in the first image frame sequence contain the target to be recognized. Target identification is performed on the image frames, which is realized by marking the position of the target and framing the target in the image frames. The position of the target is marked by arrows, dots, etc. to indicate the specific position of the target. The target is framed by a rectangular frame to frame the target area. The image frame sequence after target identification, i.e. the second image frame sequence, is obtained. On the other hand, if the target recognition result is not similar, it indicates that the image frames in the first image frame sequence do not contain the target to be recognized, and target identification is not needed. At this time, the second image frame sequence is the same as the first image frame sequence. Target identification is realized on the input image frame sequence according to the target recognition result, and target position information is provided for subsequent target tracking.

[0035] Step S60000: generating a target tracking path according to the second image frame sequence.

[0036] If the second image frame sequence contains target identification information, the motion of the target is tracked according to the target identification information, and the target tracking path is obtained. In two consecutive image frames, the position of the target changes, and this change corresponds to the motion direction and motion trajectory of the target. According to the target identification information in multiple consecutive image frames, the motion direction and motion trajectory of the target can be determined, target tracking is realized, and the target tracking path is obtained. On the other hand, if the second image frame sequence is the same as the first image frame sequence, i.e. there is no target identification, target tracking and target tracking path cannot be realized.

[0037] Target tracking is realized according to the target identification information, and the motion trajectory of the target is obtained, which provides target motion information for image segmentation. The target tracking path is path data representing the motion trajectory of the target, which directly shows the motion direction and motion range of the target.

[0038] Further, as shown in Figure 2 The embodiments of the present application further include:

[0039] Step S71000: collecting a training image set;

[0040] Step S72000: performing shape convolution feature identification on the training image set to obtain shape convolution identification features;

[0041] Step S73000: performing color convolution feature identification on the training image set to obtain color convolution identification features;

[0042] Step S74000: training a convolutional neural network according to the training image set and the shape convolution identification features to obtain a first feature extraction sub-channel;

[0043] Step S75000: training a convolutional neural network according to the training image set and the color convolution identification feature, to obtain a second feature extraction sub-channel;

[0044] Step S76000: connecting the first feature extraction sub-channel and the second feature extraction sub-channel as parallel nodes, to obtain the first feature extraction channel and the second feature extraction channel.

[0045] The training image set refers to a large amount of image data used for training the target extraction model. The shape convolution identification feature and the color convolution identification feature refer to shape information and color information extracted from the training image set through convolution operation. The first feature extraction sub-channel and the second feature extraction sub-channel refer to convolution sub-channels obtained by training according to the shape convolution identification feature and the color convolution identification feature.

[0046] First, a large number of training images are collected to form a training image set, which contains images of different targets and backgrounds and can be used to train the target extraction model. Then, shape convolution operation is performed on the training image set to extract shape information such as contours and edges in the images, to obtain shape convolution identification features. At the same time, color convolution operation is performed on the training image set to extract main colors and color distribution information in the images, to obtain color convolution identification features. Then, the convolutional neural network is trained according to the training image set and the shape convolution identification features, to obtain a first feature extraction sub-channel for shape feature extraction in the images. Subsequently, the convolutional neural network is trained according to the training image set and the color convolution identification features, to obtain a second feature extraction sub-channel for color feature extraction in the images.

[0047] Then, the first feature extraction sub-channel and the second feature extraction sub-channel are fully connected as parallel nodes to obtain the first feature extraction channel and the second feature extraction channel. The first feature extraction channel and the second feature extraction channel are used as part of the target extraction model for feature extraction of input images and template images, to provide model support for the target-based image segmentation method.

[0048] Further, the embodiments of the present application also include:

[0049] Step S41000: iteratively training the feature comparison channel based on N to-be-identified targets, and obtaining the feature comparison channel when the comparison accuracy of a continuous preset number of times meets a preset accuracy.

[0050] Wherein, any one training includes:

[0051] Step S41100: obtaining first image convolution features and second image convolution features of a first to-be-identified target.

[0052] Step S41200: Obtain third image convolution features and fourth image convolution features of a second to-be-identified target;

[0053] Step S41300: Train the feature comparison channel based on the first image convolution features, the second image convolution features, the third image convolution features, and the fourth image convolution features.

[0054] The N to-be-identified targets refer to the number of reference targets for training the feature comparison channel. The feature comparison channel refers to a network channel for realizing feature comparison and target identification. The first image convolution features and the second image convolution features refer to two features of a certain to-be-identified target, and the third image convolution features and the fourth image convolution features refer to two features of another to-be-identified target.

[0055] The feature comparison channel is iteratively trained based on the N to-be-identified targets to obtain the feature comparison channel. In any one training, two to-be-identified targets are involved. When the comparison accuracy of a continuous preset number of times of training meets a preset accuracy, the iterative training ends, and the feature comparison channel is obtained. First, the first image convolution features and the second image convolution features of a first to-be-identified target are obtained by performing convolution operation on the image of the first to-be-identified target. At the same time, the third image convolution features and the fourth image convolution features of a second to-be-identified target are obtained by performing convolution operation on the image of the second to-be-identified target. Subsequently, the feature comparison channel is trained based on the first image convolution features, the second image convolution features, the third image convolution features, and the fourth image convolution features. The feature comparison channel gradually masters the judgment criterion for judging whether two features are similar during the training process, and finally has strong feature comparison capability, realizing the training of the feature comparison channel. After the training of the feature comparison channel ends, the feature comparison channel can effectively judge and compare any two features, realizing the feature comparison function.

[0056] Further, as shown in Figure 3 the embodiment of the present application also includes:

[0057] Step S41310: input the first image convolution features and the second image convolution features into the feature comparison channel to obtain a first output result;

[0058] Step S41320: input the third image convolution features and the fourth image convolution features into the feature comparison channel to obtain a second output result;

[0059] Step S41330: input the first image convolution features and the third image convolution features into the feature comparison channel to obtain a third output result;

[0060] Step S41340: input the first image convolution feature and the fourth image convolution feature into the feature comparison channel, and obtain a fourth output result;

[0061] Step S41350: input the second image convolution feature and the third image convolution feature into the feature comparison channel, and obtain a fifth output result;

[0062] Step S41360: input the second image convolution feature and the fourth image convolution feature into the feature comparison channel, and obtain a sixth output result;

[0063] Step S41370: determine an output accuracy rate according to the first output result, the second output result, the third output result, the fourth output result, the fifth output result and the sixth output result.

[0064] First, input the first image convolution feature and the second image convolution feature into the feature comparison channel, and obtain a first output result. If the feature comparison channel judges that the two features are similar, the first output result is 1, otherwise the first output result is 0. Then, input the third image convolution feature and the fourth image convolution feature into the feature comparison channel, and obtain a second output result. If the feature comparison channel judges that the two features are similar, the second output result is 1, otherwise the second output result is 0. Subsequently, input the first image convolution feature and the third image convolution feature into the feature comparison channel, and obtain a third output result. If the feature comparison channel judges that the two features are similar, the third output result is 1, otherwise the third output result is 0. Then, input the first image convolution feature and the fourth image convolution feature into the feature comparison channel, and obtain a fourth output result. If the feature comparison channel judges that the two features are similar, the fourth output result is 1, otherwise the fourth output result is 0. Next, input the second image convolution feature and the third image convolution feature into the feature comparison channel, and obtain a fifth output result. If the feature comparison channel judges that the two features are similar, the fifth output result is 1, otherwise the fifth output result is 0. Then, input the second image convolution feature and the fourth image convolution feature into the feature comparison channel, and obtain a sixth output result. If the feature comparison channel judges that the two features are similar, the sixth output result is 1, otherwise the sixth output result is 0. Finally, calculate the output accuracy rate according to the first output result, the second output result, the third output result, the fourth output result, the fifth output result and the sixth output result, and evaluate the training effect of the feature comparison channel.

[0065] The output result of the feature comparison channel is obtained according to the features of the two to-be-recognized targets, and the output accuracy is calculated according to the output result, so as to evaluate the training effect of the feature comparison channel. If the output accuracy reaches a predetermined threshold, it means that the feature comparison channel is trained; otherwise, the feature comparison channel needs to be continuously trained. A supervision mechanism is provided for the training of the feature comparison channel, so that the trained feature comparison channel reaches the expected judgment accuracy.

[0066] Further, the embodiments of the application also include:

[0067] Step S41371: Identifying the first image convolution feature and the second image convolution feature, obtaining a first similar identification, wherein the first similar identification corresponds to the first to-be-recognized target;

[0068] Step S41372: Identifying the third image convolution feature and the fourth image convolution feature, obtaining a second similar identification, wherein the second similar identification corresponds to the second to-be-recognized target;

[0069] Step S41373: Identifying the first image convolution feature and the third image convolution feature, obtaining a first non-similar identification;

[0070] Step S41374: Identifying the first image convolution feature and the fourth image convolution feature, obtaining a second non-similar identification;

[0071] Step S41375: Identifying the second image convolution feature and the third image convolution feature, obtaining a third non-similar identification;

[0072] Step S41376: Identifying the second image convolution feature and the fourth image convolution feature, obtaining a fourth non-similar identification;

[0073] Step S41377: Comparing the first output result, the second output result, the third output result, the fourth output result, the fifth output result and the sixth output result based on the first similar identification, the second similar identification, the first non-similar identification to the fourth non-similar identification, and obtaining a comparison accuracy.

[0074] First, the first image convolution feature and the second image convolution feature are identified to obtain a first similarity identifier, indicating that the two features are similar, corresponding to a first target to be identified. Meanwhile, the third image convolution feature and the fourth image convolution feature are identified to obtain a second similarity identifier, indicating that the two features are similar, corresponding to a second target to be identified. Then, the first image convolution feature and the third image convolution feature are identified to obtain a first non-similarity identifier, indicating that the two features are not similar; the first image convolution feature and the fourth image convolution feature are identified to obtain a second non-similarity identifier, indicating that the two features are not similar. Meanwhile, the second image convolution feature and the third image convolution feature are identified to obtain a third non-similarity identifier, indicating that the two features are not similar; the second image convolution feature and the fourth image convolution feature are identified to obtain a fourth non-similarity identifier, indicating that the two features are not similar. Finally, based on the first similarity identifier, the second similarity identifier, the first non-similarity identifier to the fourth non-similarity identifier, the output results (first output result to sixth output result) of the feature comparison channel are judged to determine whether the output results are consistent with the similarity identifier and the non-similarity identifier. The number of accurate comparisons is divided by the total number of outputs to obtain the comparison accuracy.

[0075] By identifying the similarity and non-similarity of the four features of the two targets to be identified, and judging the accuracy of the output results of the feature comparison channel based on the identification, the judgment ability of the feature comparison channel is evaluated. If the comparison accuracy reaches a predetermined threshold, it means that the feature comparison channel is trained; otherwise, the feature comparison channel needs to be further trained.

[0076] Further, the embodiments of the present application also include:

[0077] Step S51000: When the first image frame sequence includes a first target tracking identifier and a second target tracking identifier, the first target tracking identifier and the second target tracking identifier are different.

[0078] The first target tracking identifier and the second target tracking identifier refer to the identification of different targets in the first image frame sequence. If an image frame in the first image frame sequence contains two or more targets, each target is identified to track them. At this time, each target is given a target tracking identifier, the first target tracking identifier corresponds to one target, and the second target tracking identifier corresponds to another target. In order to accurately judge and distinguish different targets by the image segmentation method, the first target tracking identifier and the second target tracking identifier should be different, such as using different colors, identifiers, etc. for distinction.

[0079] By using different target tracking identifiers for each target when the input first image frame sequence contains multiple targets, accurate target distinction and tracking are achieved, and the complete motion trajectory of each target is obtained, providing accurate target motion information for image segmentation and improving the accuracy of image segmentation.

[0080] In summary, the image segmentation method based on tracking target provided by the embodiments of the present application has the following technical effects:

[0081] The first image frame sequence is obtained to provide image information basis for subsequent steps; the first image frame sequence is sequentially input into the first feature extraction channel of the target extraction model to obtain the first feature extraction result; the features of the input image sequence are extracted by using the first feature extraction channel to obtain the input image features; the target template image to be recognized is input into the second feature extraction channel of the target extraction model to obtain the second feature extraction result; the first feature extraction channel and the second feature extraction channel have the same network structure and network parameters; the features of the target template are extracted by using the second feature extraction channel to obtain the target template features; the first feature extraction result and the second feature extraction result are input into the feature comparison channel of the target extraction model to obtain the target recognition result, which provides data support for subsequent target identification; the target identification is performed on the first image frame sequence according to the target recognition result to obtain the second image frame sequence; and the target tracking path is generated according to the second image frame sequence, thereby achieving the technical effects of improving the target recognition accuracy and tracking stability.

[0082] Embodiment Two

[0083] Based on the same inventive concept as the image segmentation method based on tracking target in the foregoing embodiments, as shown in Figure 4 The embodiments of the present application provide an image segmentation system based on tracking target, which comprises:

[0084] An image sequence acquisition module 11 is configured to acquire a first image frame sequence.

[0085] A first feature extraction module 12 is configured to sequentially input the first image frame sequence into a first feature extraction channel of a target extraction model to obtain a first feature extraction result.

[0086] A second feature extraction module 13 is configured to input a target template image to be recognized into a second feature extraction channel of the target extraction model to obtain a second feature extraction result, wherein the first feature extraction channel and the second feature extraction channel have the same network structure and network parameters.

[0087] A feature comparison channel module 14 is configured to input the first feature extraction result and the second feature extraction result into a feature comparison channel of the target extraction model to obtain a target recognition result.

[0088] An image target identification module 15 is configured to perform target identification on the first image frame sequence according to the target recognition result to obtain a second image frame sequence.

[0089] The tracking path generation module 16 is configured to generate a target tracking path according to the second image frame sequence.

[0090] Further, the embodiment of the present application further comprises an extraction channel training module, which comprises the following execution steps:

[0091] collecting a training image set;

[0092] performing shape convolution feature identification on the training image set to obtain shape convolution identification features;

[0093] performing color convolution feature identification on the training image set to obtain color convolution identification features;

[0094] training a convolutional neural network according to the training image set and the shape convolution identification features to obtain a first feature extraction sub-channel;

[0095] training a convolutional neural network according to the training image set and the color convolution identification features to obtain a second feature extraction sub-channel;

[0096] connecting the first feature extraction sub-channel and the second feature extraction sub-channel as parallel nodes full connection to obtain the first feature extraction channel and the second feature extraction channel.

[0097] Further, the feature comparison channel module 14 comprises the following execution steps:

[0098] iteratively training the feature comparison channel based on N to-be-identified targets, and obtaining the feature comparison channel when the comparison accuracy rate of a continuous preset number of times meets a preset accuracy rate;

[0099] wherein, any one training comprises:

[0100] obtaining a first image convolution feature and a second image convolution feature of a first to-be-identified target;

[0101] obtaining a third image convolution feature and a fourth image convolution feature of a second to-be-identified target;

[0102] training the feature comparison channel based on the first image convolution feature, the second image convolution feature, the third image convolution feature, and the fourth image convolution feature.

[0103] Further, the feature comparison channel module 14 further comprises the following execution steps:

[0104] inputting the first image convolution feature and the second image convolution feature into the feature comparison channel to obtain a first output result;

[0105] input the third image convolution feature and the fourth image convolution feature into the feature comparison channel to obtain a second output result;

[0106] input the first image convolution feature and the third image convolution feature into the feature comparison channel to obtain a third output result;

[0107] input the first image convolution feature and the fourth image convolution feature into the feature comparison channel to obtain a fourth output result;

[0108] input the second image convolution feature and the third image convolution feature into the feature comparison channel to obtain a fifth output result;

[0109] input the second image convolution feature and the fourth image convolution feature into the feature comparison channel to obtain a sixth output result;

[0110] determine an output accuracy rate according to the first output result, the second output result, the third output result, the fourth output result, the fifth output result and the sixth output result.

[0111] Further, the feature comparison channel module 14 further includes the following execution steps:

[0112] identify the first image convolution feature and the second image convolution feature to obtain a first similar identifier, wherein the first similar identifier corresponds to the first to-be-identified target;

[0113] identify the third image convolution feature and the fourth image convolution feature to obtain a second similar identifier, wherein the second similar identifier corresponds to the second to-be-identified target;

[0114] identify the first image convolution feature and the third image convolution feature to obtain a first non-similar identifier;

[0115] identify the first image convolution feature and the fourth image convolution feature to obtain a second non-similar identifier;

[0116] identify the second image convolution feature and the third image convolution feature to obtain a third non-similar identifier;

[0117] identify the second image convolution feature and the fourth image convolution feature to obtain a fourth non-similar identifier;

[0118] compare the first output result, the second output result, the third output result, the fourth output result, the fifth output result and the sixth output result based on the first similar identifier, the second similar identifier, the first non-similar identifier to the fourth non-similar identifier to obtain a comparison accuracy rate.

[0119] Further, the image target identification module 15 comprises the following execution steps:

[0120] When the first image frame sequence comprises a first target tracking identification and a second target tracking identification, the first target tracking identification and the second target tracking identification are different.

[0121] Any step of the above method can be stored as computer instructions or programs in an unrestricted computer memory and can be invoked by an unrestricted computer processor to recognize to implement any of the methods in the embodiments of the present application, without any redundant limitations.

[0122] Further, the above first or second may not only represent a sequence relationship, but also may represent a specific concept, and / or refer to the selection of a plurality of elements individually or collectively. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technologies, the present application is intended to include these modifications and variations.

Claims

1. An image segmentation method based on tracking targets, characterized in that, include: Obtain the first image frame sequence; The first image frame sequence is sequentially input into the first feature extraction channel of the target extraction model to obtain the first feature extraction result; The target template image to be identified is input into the second feature extraction channel of the target extraction model to obtain the second feature extraction result, wherein the first feature extraction channel and the second feature extraction channel have the same network structure and network parameters; The first feature extraction result and the second feature extraction result are input into the feature comparison channel of the target extraction model. Obtain target recognition results; Based on the target recognition result, the first image frame sequence is identified as a target, and a second image frame sequence is obtained; Generate a target tracking path based on the second image frame sequence; The image segmentation method based on tracking targets further includes: Collect a training image set; Shape convolution feature labeling is performed on the training image set to obtain shape convolution label features; Color convolution feature labeling is performed on the training image set to obtain color convolution label features; Based on the training image set and the shape convolutional identifier features, the convolutional neural network is trained to obtain the first feature extraction sub-channel; Based on the training image set and the color convolutional identifier features, the convolutional neural network is trained to obtain the second feature extraction sub-channel; The first feature extraction sub-channel and the second feature extraction sub-channel are fully connected as parallel nodes to obtain the first feature extraction channel and the second feature extraction channel; The image segmentation method based on tracking targets further includes: Based on N targets to be identified, the feature comparison channel is iteratively trained. When the comparison accuracy of a preset number of consecutive preset numbers meets the preset accuracy, the feature comparison channel is obtained. Any training session includes: Obtain the first image convolutional features and the second image convolutional features of the first target to be identified; Obtain the third and fourth image convolutional features of the second target to be identified; The feature comparison channel is trained based on the first image convolutional features, the second image convolutional features, the third image convolutional features, and the fourth image convolutional features.

2. The method as described in claim 1, characterized in that, Based on the first image convolutional features, the second image convolutional features, the third image convolutional features, and the fourth image convolutional features, the feature comparison channel is trained, including: The first image convolutional features and the second image convolutional features are input into the feature comparison channel to obtain the first output result; The third image convolutional feature and the fourth image convolutional feature are input into the feature comparison channel to obtain the second output result; The first image convolutional feature and the third image convolutional feature are input into the feature comparison channel to obtain the third output result; The first image convolutional feature and the fourth image convolutional feature are input into the feature comparison channel to obtain the fourth output result; The second image convolutional feature and the third image convolutional feature are input into the feature comparison channel to obtain the fifth output result; The second image convolutional feature and the fourth image convolutional feature are input into the feature comparison channel to obtain the sixth output result; The output accuracy is determined based on the first output result, the second output result, the third output result, the fourth output result, the fifth output result, and the sixth output result.

3. The method as described in claim 2, characterized in that, Based on the first output result, the second output result, the third output result, the fourth output result, the fifth output result, and the sixth output result, the output accuracy is determined, including: For the first image convolutional feature and the second image convolutional feature identifier, obtain a first similarity identifier, wherein the first similarity identifier corresponds to the first target to be identified; For the third image convolutional feature and the fourth image convolutional feature identifier, a second similarity identifier is obtained, wherein the second similarity identifier corresponds to the second target to be identified; The first dissimilarity identifier is obtained by identifying the convolutional features of the first image and the convolutional features of the third image; Based on the convolutional features of the first image and the convolutional features of the fourth image, a second dissimilarity identifier is obtained; Based on the second image convolutional feature and the third image convolutional feature identifier, obtain the third dissimilarity identifier; Based on the convolutional features of the second image and the convolutional features of the fourth image, a fourth dissimilarity identifier is obtained; Based on the first similarity identifier, the second similarity identifier, the first dissimilarity identifier up to the fourth dissimilarity identifier, the first output result, the second output result, the third output result, the fourth output result, the fifth output result and the sixth output result are compared to obtain the comparison accuracy.

4. The method as described in claim 1, characterized in that, Also includes: When the first image frame sequence includes a first target tracking identifier and a second target tracking identifier, the first target tracking identifier and the second target tracking identifier are different.

5. An image segmentation system based on target tracking, characterized in that, The system for implementing the image segmentation method based on tracking a target as described in any one of claims 1-4, the system comprising: An image sequence acquisition module, wherein the image sequence acquisition module is used to acquire a first image frame sequence; The first feature extraction module is used to sequentially input the first image frame sequence into the first feature extraction channel of the target extraction model to obtain the first feature extraction result. The second feature extraction module is used to input the target template image to be identified into the second feature extraction channel of the target extraction model to obtain the second feature extraction result. The first feature extraction channel and the second feature extraction channel have the same network structure and network parameters. The feature comparison channel module is used to input the first feature extraction result and the second feature extraction result into the feature comparison channel of the target extraction model to obtain the target recognition result; An image target identification module is used to identify targets in the first image frame sequence based on the target recognition result, and to obtain a second image frame sequence. A tracking path generation module is used to generate a target tracking path based on the second image frame sequence.

Images