Anti-interference tracking method for low, small and slow targets in air
By building an end-to-end target tracking algorithm and object detection algorithm, the tracking problem of low and small and slow targets in the air in the case of interference and occlusion is solved, and a more stable and reliable target tracking effect is achieved.
Patent Information
- Application Number
- CN202411943484.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-27
- Publication Date
- 2025-04-25
AI Technical Summary
Stable and reliable tracking of low and small air targets (such as drones) in the fields of security monitoring and military alert faces problems of interference and occlusion, resulting in a decline in tracking quality.
An end-to-end target tracking algorithm is used to construct a target classification model composed of feature extraction backbone networks and classification and position regression subnets, and interference detection and target repositioning are combined with the target detection algorithm.
It effectively overcomes the adverse effects of interference and occlusion on target tracking, and improves the stable and reliable tracking capabilities of low, small and slow targets in the air.
Smart Images
Figure CN119887841A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of target tracking, and in particular relates to an anti-interference tracking method for low, small and slow targets in the air. Background Art
[0002] Stable and reliable tracking of "low, small and slow" targets such as drones is being increasingly studied in the fields of security monitoring, military alert, etc.
[0003] However, in practical applications, small targets are easily disturbed or even blocked by flying birds or buildings. At the same time, the field of view that can be used for observation and tracking positioning is also limited. Therefore, stable and reliable tracking of such targets is still very challenging.
[0004] Therefore, a method is needed to solve this problem by determining whether target tracking is interfered with and by detecting and positioning the target again after it is lost. Summary of the invention
[0005] The object of the present invention is to provide a method for anti-interference tracking of low, small and slow targets in the air, so as to solve the problem of anti-interference tracking of low, small and slow targets in the air.
[0006] The technical solution adopted by the present invention to solve the technical problem is: a method for anti-interference tracking of low, small and slow targets in the air, comprising the following steps: S1, using an end-to-end target tracking algorithm, constructs a target classification model consisting of a feature extraction backbone network and a classification and position regression subnetwork, wherein the feature extraction backbone network is a classification network such as AlexNet, Vgget or Reset that removes the fully connected layer and only retains the convolution layer, pooling layer and non-linear activation layer; the classification and position regression subnetwork is connected to the feature extraction backbone network, and is composed of a classification subnetwork and a regression subnetwork with a smaller number of layers than the feature extraction backbone network. The classification subnetwork is used to distinguish each element category of the feature map output by the feature extraction backbone network, and the regression subnetwork is used to obtain the precise coordinate position of the tracked target according to the probability distribution map obtained by the classification subnetwork; S2, target tracking loss criterion: input the tracked target image into the target classification model, and judge whether the target tracking is disturbed according to the output result;
[0007] S3, if the criterion is disturbed, the target detection algorithm is used to relocate the target after it is lost: the manually annotated true value is used as a template, and compared with the features of the target area of the current frame to obtain the classification and position regression information of the current target, and the target detection algorithm based on the convolutional neural network YOLO is used to perform network training on the above-organized data to obtain a training set; the material with occlusion exceeding 50% of the target pixel is selected and marked as disturbed, and the target coordinates are estimated and marked according to the historical size and position information of the image sequence, and the rest of the material that is not disturbed is marked as normal, and the outline of the target is marked as the coordinates;
[0008] S4: If the criterion is not disturbed, the original target tracking algorithm is used to continue the target tracking until the target position is finally obtained.
[0009] Furthermore, the feature extraction backbone network includes a template branch and a detection branch that share CNN parameters: the template branch takes the target block of the first frame of the video as the image input, represented by Y; the detection branch takes the center of the target frame predicted in the previous frame as the reference point, and after cropping, it is used as the target block input of the current frame, represented by X, and the two networks use Y and X to represent the network output.
[0010] Furthermore, the classification subnetwork is composed of a convolutional layer, a nonlinear activation layer, and a softmax function; and the regression subnetwork is composed of a convolutional layer and a nonlinear activation layer.
[0011] Furthermore, in step S2, the convolutional neural network YOLO is first used to detect the target image to be tracked, and the material that occludes more than 50% of the target pixels is selected and marked as disturbed, and the remaining materials that are not disturbed are marked as normal to obtain a training set; then a typical target classification network is selected, such as Alexnet, Vgget, Reset, etc., and the training set is trained to obtain a new classification network for determining whether the target tracking is disturbed.
[0012] Furthermore, the step S3 also includes post-processing of the output results: abandoning the prediction box generated by the anchor point far away from the center point; reordering the candidate areas using the cosine window and scaling transformation penalty items; and selecting the final prediction frame result using the non-maximum suppression method.
[0013] The beneficial effect of the present invention is that a classification network is used in the tracking method of the present invention to evaluate the tracking quality. When the tracking process is affected by interference, occlusion, etc., a target detection algorithm can be used to re-detect the target, thereby effectively overcoming the adverse effects of interference on stable tracking. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 The following is a flow chart of the tracking method of the present invention. DETAILED DESCRIPTION
[0015] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0016] Reference Figure 1 As shown, the present invention discloses an anti-interference tracking method for low, small and slow targets in the air, which includes the following steps.
[0017] S1, build the target classification model. The end-to-end target classification model mainly consists of a feature extraction backbone network and a classification and position regression sub-network.
[0018] The feature extraction backbone network removes the fully connected layer and only retains the convolutional layer, pooling layer, and non-linear activation layer, and uses classification networks such as AlexNet, Vgget, and Reset.
[0019] The feature extraction backbone network consists of a template branch and a detection branch: the template branch takes the target block of the first frame of the video as the image input, represented by Y; the detection branch uses the center of the target box predicted in the previous frame as the reference point, and after cropping, it is used as the target block input of the current frame, represented by X. The two networks share CNN parameters, and Y and X represent the network outputs.
[0020] The classification and position regression subnetwork is connected after the feature extraction backbone network, and consists of a classification subnetwork and a regression subnetwork with fewer layers than the feature extraction backbone network. The classification subnetwork consists of a convolutional layer, a nonlinear activation layer, and a softmax function. The regression subnetwork consists of a convolutional layer and a nonlinear activation layer. Generally, the number of layers of these two subnetworks is smaller than that of the feature extraction backbone network.
[0021] The classification subnetwork is used to distinguish the category of each element in the output feature map. In tracking, it is a binary classification, that is, target or background. Since the output result of the softmax function is the probability value of the target classification between 0 and 1, the feature map can be considered as a probability distribution map. The regression subnetwork obtains the precise coordinate position of the tracked target based on the probability distribution map obtained by the classification subnetwork.
[0022] S2, target tracking loss criterion: input the tracked target image into the target classification model, and judge whether the target tracking is disturbed according to the output result.
[0023] Network training: First use the convolutional neural network YOLO to detect and track the target image, select the materials that occlude more than 50% of the target pixels and mark them as disturbed, and mark the rest of the materials that are not disturbed as normal to obtain the training set.
[0024] Determine whether it is interfered with: Then select a typical target classification network, such as Alexnet, Vgget, Reset, etc., train the training set according to the labels of the above materials, and obtain a new classification network to determine whether the target tracking is interfered with.
[0025] S3, based on the target tracking loss criterion, once the criterion is disturbed, that is, it is judged that the current target tracking reliability is not high, the target detection algorithm is used to quickly achieve repositioning after the target is lost.
[0026] Network training: The manually annotated true value is used as a template and compared with the features of the target area in the current frame. The region generation network is further used to obtain the classification and position regression information of the current target. Considering the real-time performance, the target detection algorithm based on the convolutional neural network YOLO is used to perform network training on the above-organized data to obtain the training set.
[0027] The training set is processed: the materials with occlusion exceeding 50% of the target pixels are selected and marked as disturbed. According to the historical size and position information of the image sequence, the target coordinates are estimated and marked. The remaining materials that are not disturbed are marked as normal, and the outline of the target is marked as the coordinates.
[0028] S4: If the criterion is not disturbed, the original target tracking algorithm is used to continue the target tracking until the target position is finally obtained.
[0029] The target anti-interference tracking method based on the convolutional neural network of the present invention can abstract the linear and nonlinear characteristics of the target in multiple dimensions due to its massive data support and nonlinear learning ability. Therefore, it has more powerful generalization ability and target recognition ability than traditional feature extraction operators.
[0030] The above embodiments are only illustrative of the principles and effects of the present invention, as well as some embodiments of its application. For those skilled in the art, several modifications and improvements may be made without departing from the creative concept of the present invention, and all of these belong to the protection scope of the present invention.
Claims
1. A method for anti-interference tracking of low, small and slow targets in the air, characterized by: The following steps are included S1, construct a target classification model based on a feature extraction backbone network and a classification and position regression subnetwork, wherein the feature extraction backbone network is an AlexNet, Vgget or Reset classification network that only retains convolutional layers, pooling layers and nonlinear activation layers; the classification and position regression subnetwork is connected after the feature extraction backbone network, and is composed of a classification subnetwork and a regression subnetwork whose layers are smaller than those of the feature extraction backbone network. The classification subnetwork is used to distinguish each element category of the feature map output by the feature extraction backbone network, and the regression subnetwork obtains the precise coordinate position of the tracked target based on the probability distribution map obtained by the classification subnetwork; S2, input the tracked target image into the target classification model, and determine whether the target tracking is disturbed according to the output result; S3, based on the target tracking loss criterion, once it is determined that the current target tracking reliability is not high, the target detection algorithm is used to quickly achieve target relocation: the manually annotated real value is used as a template, and compared with the characteristics of the target area of the current frame to obtain the classification and position regression information of the current target, and the target detection algorithm based on the convolutional neural network YOLO is used to perform network training on the above data to obtain a training set; the material with occlusion exceeding 50% of the target pixel is selected and marked as disturbed, and the target coordinates are estimated and marked according to the historical size and position information of the image sequence, and the rest of the material that is not disturbed is marked as normal, and the outline of the target is marked as the coordinates; S4: If the criterion is not disturbed, the target tracking is continued until the target position is obtained.
2. The anti-interference tracking method for low, small and slow targets in the air according to claim 1 is characterized in that: The feature extraction backbone network includes a template branch and a detection branch that share CNN parameters: the template branch takes the target block of the first frame of the video as an image input; the detection branch uses the center of the target box predicted in the previous frame as a reference point, and after cropping, it is used as the target block input of the current frame.
3. The anti-interference tracking method for low, small and slow targets in the air according to claim 2 is characterized in that: The classification subnetwork is composed of a convolutional layer, a nonlinear activation layer, and a softmax function; the regression subnetwork is composed of a convolutional layer and a nonlinear activation layer.
4. A method for anti-interference tracking of low, small and slow targets in the air according to claim 1, 2 or 3, characterized in that: In step S2, the convolutional neural network YOLO is first used to detect the target image to be tracked, and the material that occludes more than 50% of the target pixels is selected and marked as disturbed, and the remaining materials that are not disturbed are marked as normal to obtain a training set; then Alexnet, Vgget or Reset is selected to train the training set to obtain a new classification network to determine whether the target tracking is disturbed.
5. The anti-interference tracking method for low, small and slow targets in the air according to claim 4 is characterized in that: The step S3 also includes post-processing of the output result: abandoning the prediction box generated by the anchor point far away from the center point; reordering the candidate areas using the cosine window and the scaling transformation penalty item; and selecting the final prediction frame result using the non-maximum suppression method.