Satellite Video Moving Vehicle Tracking Method Based on Feature Enhancement and Location Prediction

Abstract

The present invention relates to the technical field of moving target tracking, and specifically discloses a satellite video moving vehicle tracking method based on feature enhancement and position prediction. By optimizing the bounding box constraint size of SiamRPN, it is made more suitable for the satellite video small target tracking task. By introducing the scSE attention mechanism, the key feature information of the target is fully mined, and the drift probability of the model under similar interference and motion blur interference is reduced; then, the classification score map of SiamRPN after enhancing the features by the scSE attention mechanism is used to construct an F max -APCE-CSL tracking status monitoring module, and the position of the occluded target is predicted and corrected by using an inertial mechanism, realizing the continuous tracking of the moving vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of moving target tracking, and particularly relates to a satellite video moving vehicle tracking method based on feature enhancement and position prediction. Background Art

[0002] Video target tracking technology plays an important role in urban public security and can be applied to scenarios such as crime fighting, police law enforcement, and public security prevention. Target tracking refers to the process of automatically obtaining the state of a target in subsequent video frames given the initial state (such as position and size) of a specific target in a certain frame of a video. Currently, the perspective of urban monitoring systems mainly based on ground video systems is limited, which is not conducive to continuous target tracking. In recent years, video satellite imaging has a large swath width, and the data obtained has a time resolution of seconds and a spatial resolution of sub-meters, providing the possibility for continuous vehicle target tracking at the urban scale.

[0003] However, compared with ordinary video target tracking, vehicle target tracking based on satellite video faces more prominent challenges: (1) Large image swath width and small target size: The video satellite has a large field of view, and the video swath width can reach 4 km × 2 km. Targets such as vehicles only occupy a few to dozens of pixels in satellite video, and have few color and texture features. (2) Severe interference from similar objects: There are a large number of targets, and the distinguishability between the target and the interference object is poor, and the similarity is high. (3) Motion blur: The situation where the target is blurred due to the rapid movement of the target or the jitter of the satellite platform. (4) Target occlusion: The target is easily occluded by houses, trees, overpasses, etc.

[0004] Current satellite video moving vehicle tracking methods are mainly divided into two categories: ① Traditional target tracking methods: Traditional generative target tracking methods and target tracking methods based on correlation filtering usually use a single or a combination of several features such as gray-scale features and histogram of oriented gradients to characterize the target, and such features have weak expression ability and a small applicable range, and are prone to tracking failure in complex scenarios such as similar interference objects and motion blur. ② Deep learning-based target tracking methods: Deep learning-based tracking methods can learn shallow features containing detailed information and deep features with rich semantic information and strong anti-interference ability. The tracking adaptability is strong and can perform robust tracking, providing a new way for target tracking. How to use deep learning to carry out satellite video vehicle target tracking work is a problem worthy of in-depth study.

[0005] To improve the accuracy of satellite video target tracking, in the prior art, in the patent with the application number: 201910813189.9 and the invention title "Satellite Video Target Tracking Method Based on Fully Convolutional Siamese Network and Trajectory Prediction", the invention first utilizes the shallow-layer information of the fully convolutional siamese network to obtain fine-grained representations of small targets, performs similarity measurement and tracking. To address the situations of occlusion and motion blur of small targets during tracking, a Gaussian mixture model is used to monitor the target tracking state, and a Kalman filtering mechanism is introduced to adaptively predict the motion trajectory, ultimately achieving satellite video target tracking. When solving the occlusion problem, the tracking accuracy of this invention depends not only on the accuracy of the target tracking method based on the siamese network, but also on the accuracy of the Gaussian mixture model. In the patent with the application number: 202010060627.1 and the invention title "A Satellite Video Target Tracking Method Based on High-Resolution Siamese Network", the invention uses a lightweight parallel network to obtain high-resolution representations of small targets, achieving robust and real-time tracking. And a pixel-level fine-tuning model is proposed based on the tracking. It utilizes inter-frame motion information, establishes a Gaussian mixture model, detects changing pixels, and obtains a binary mask image. The mean shift algorithm is repeatedly iterated on the mask image to achieve fine-tuning at the pixel level. At the same time, considering the complementarity of surface features and motion features, the tracking position and the fine-tuning position are further adaptively fused, ultimately obtaining more accurate tracking.

[0006] With the emergence of tracking technologies based on siamese region proposal networks in recent years, which can more accurately predict the target bounding box, related research has gradually increased, and certain progress has been made in the field of satellite video target tracking. However, there are still the following challenges: 1) Current satellite video target tracking technologies based on siamese region proposal networks mostly directly use shallow fine-grained depth features to represent small targets in satellite videos. Although the tracking performance has been improved to a certain extent, the effective feature information has not been fully exploited, and model drift is still likely to occur under complex conditions such as similar interference objects and motion blur, and the tracking accuracy needs to be further improved; 2) Most satellite video target tracking technologies based on siamese region proposal networks do not fully utilize the generated classification score maps and do not have the ability to monitor the tracking state and predict the position, resulting in limited anti-occlusion ability, and the tracking method cannot continuously track the target when the target is occluded.

[0007] Therefore, there is an urgent need for a tracking method for moving targets that can solve the above problems and can achieve accurate tracking even when the target is occluded. Summary of the Invention

[0008] To solve the problems existing in the prior art, the present invention provides a satellite video moving vehicle tracking method based on feature enhancement and position prediction, introducing a concurrent spatial and channelsqueeze&excitation (scSE) hybrid attention mechanism to improve the ability of the target tracking model to extract effective features of satellite video moving vehicle targets, reducing the probability of the tracking model drifting under similar interferences and motion blur interferences, constructing a tracking state monitoring module and combining it with an inertial mechanism to solve the problem that the target tracking model after introducing the attention mechanism to enhance features still cannot cope with occlusion interference, so as to realize the continuous tracking of moving vehicle targets based on satellite video and solve the problems mentioned in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions: A satellite video moving vehicle tracking method based on feature enhancement and position prediction, comprising the following steps:

[0010] S1. Construct a satellite video target tracking data set and divide the data set into a training set and a test set;

[0011] S2. Adopt the Siamese Region Proposal Network (SiamRPN), and then optimize the SiamRPN bounding box constraint size;

[0012] S3. Add the scSE hybrid attention mechanism to construct a SiamRPN satellite video moving vehicle tracking model based on the scSE attention mechanism, and then input the training set to train the model to obtain an optimal accuracy training model;

[0013] S4. Input the test set into the optimal accuracy training model, and construct a tracking state monitoring module F max -APCE-CSL through the classification score map of SiamRPN enhanced by the scSE attention mechanism, and then calculate the tracking state monitoring index using the classification score map to judge whether the target is occluded;

[0014] S5. When it is judged that the target is not in an occluded state, use the tracking result as the final target tracking result;

[0015] S6. When it is judged that the target is in an occluded state, use the inertial mechanism to predict its position and the target bounding box size to achieve continuous tracking.

[0016] Preferably, in step S2, the optimization of the SiamRPN bounding box constraint size is specifically: reducing the target bounding box constraint of SiamRPN from 10×10 pixels to 5×5 pixels.

[0017] Preferably, the scSE hybrid attention mechanism includes an sSE spatial attention mechanism and a cSE channel attention mechanism; the original feature map passes through the sSE and cSE respectively to obtain features enhanced in the spatial and channel dimensions, and these two feature maps are added together to output the final classification score map.

[0018] Preferably, the sSE spatial attention mechanism is used to process the two-dimensional spatial information of the feature map: first, the input feature map of H×W×C is convolved through a 1×1×C×1 convolutional layer to obtain a feature map of H×W×1, and then this feature map is mapped to the range [0, 1] through the Sigmoid activation function to represent the importance of different spatial position information. Finally, it is multiplied element-wise with the original feature map to obtain the feature after spatial calibration;

[0019] The cSE channel attention mechanism is used to autonomously learn the weights of each channel: first, the feature map of size H×W×C is input into a global average pooling layer to generate 1×1×C feature values corresponding to the weights of C channels, and then a convolutional operation of 1×1×C×C / 2 is performed on it. The convolutional result passes through the non-linear activation function ReLU, and the obtained output passes through a 1×1×C / 2×C convolutional operation to obtain a tensor of 1×1×C. This tensor is normalized to the range [0, 1] through the Sigmoid layer, that is, the weights of each channel, and finally multiplied with the original feature map to obtain the feature after channel calibration.

[0020] Preferably, the tracking state monitoring module F max -APCE-CSL includes three tracking state monitoring indicators: the peak of the score map F max , APCE, and CSL;

[0021] Among them, APCE reflects the fluctuation degree of the score map and the confidence level of the tracking result, and the expression is as follows:

[0022]

[0023] CSL quantitatively describes the steepness of the peak of the score map, and the expression is as follows:

[0024]

[0025] In the formula, F represents the score map of size N×N, F i,j represents the value at the i-th row and j-th column in the score map, F max , F min represent the maximum and minimum values of the score map respectively, and μ and v represent the row and column corresponding to the peak of the score map.

[0026] Preferably, the specific determination of whether the target is occluded includes: when Fmax Less than the threshold T1 and APCE less than the threshold T2, or when F max Less than the threshold T1 and CSL greater than the threshold T3, it is determined that the target is in an occluded state; where T1, T2, and T3 are respectively the average values of F max corresponding to the previous n frames of the current frame, APCE, and CSL.

[0027] Preferably, in step S6, the inertial mechanism is used to predict its position and the target bounding box size. Specifically: First, calculate the average displacements Δx and Δy of the target in the x and y directions within n frames. To prevent the introduction of noise, update Δx and Δy every m frames; then, based on the position and bounding box size of the target in the (t - 1) frame, predict the target position and bounding box size x t and y t and w t and h t in the t-th frame, as shown in the following formula:

[0028]

[0029]

[0030]

[0031] In the formula, x t and y t respectively represent the pixel coordinates of the center point of the target bounding box in the x and y directions in the t-th frame, w t and h t respectively represent the width and height of the target bounding box in the t-th frame, and w t-1 and h t-1 represent the width and height of the target bounding box in the (t - 1)-th frame.

[0032] The beneficial effects of the present invention are:

[0033] 1) The present invention takes into account the characteristics of small vehicle targets in satellite videos, optimizes the bounding box constraints of SiamRPN, makes the prediction results of SiamRPN more fitting with the target size, makes it more suitable for the small target tracking task of satellite videos, and improves the tracking accuracy of the model.

[0034] 2) The present invention introduces the scSE attention mechanism to construct a feature enhancement module, fully excavates the key feature information of the target, better highlights the feature elements beneficial to the tracking task, improves the problem of weak feature expression ability of small targets, and reduces the drift probability of the model under similar interference and motion blur interference.

[0035] 3) The present invention constructs a combination of F max-APCE-CSL and the anti-occlusion module of the inertial mechanism. Aiming at the problem that the model cannot continuously track the target after the target is occluded, the classification score map of SiamRPN after enhancing the features by the scSE attention mechanism is used to construct F max -APCE-CSL tracking status monitoring module. When it is monitored that the target is occluded, the inertial mechanism is used to predict and correct the position of the occluded target, realizing the continuous tracking of moving vehicles. Description of the Drawings

[0036] Figure 1 It is a schematic diagram of the technical idea steps of the present invention;

[0037] Figure 2 It is a schematic diagram of the overall model structure of the method of the present invention;

[0038] Figure 3 It is a schematic diagram of the scSE attention mechanism structure;

[0039] Figure 4 It is a schematic diagram of the tracking status monitoring module;

[0040] Figure 5 It is the 10-layer classification score map of SiamRPN;

[0041] Figure 6 It is a schematic diagram of the anti-occlusion process. Detailed Implementation Manner

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0043] Please refer to Figures 1-6 , the present invention provides a technical solution: a satellite video moving vehicle tracking method based on feature enhancement and position prediction, including the following steps:

[0044] S1. Construct a satellite video target tracking data set and divide the data set into a training set and a test set;

[0045] S2. Adopt the twin region candidate network SiamRPN, and then optimize the bounding box constraint size of SiamRPN;

[0046] S3. Add the scSE hybrid attention mechanism to construct a SiamRPN satellite video moving vehicle tracking model based on the scSE attention mechanism, and then input the training set to train the model to obtain the training model with the optimal accuracy;

[0047] S4. Input the test set into the training model with the optimal accuracy. Use the classification score map of SiamRPN after enhancing features through the scSE attention mechanism to construct the tracking state monitoring module F max -APCE-CSL, and then calculate the tracking state monitoring metrics using the classification score map to determine whether the target is occluded;

[0048] S5. When it is determined that the target is not in an occluded state, use the tracking result as the final target tracking result;

[0049] S6. When it is determined that the target is in an occluded state, use the inertial mechanism to predict its position and the size of the target bounding box to achieve continuous tracking.

[0050] The technical idea step flow of the technical solution of the present invention is as Figure 1 shown, mainly divided into two stages:

[0051] I. Model improvement and training: First, screen the basic model, then select a suitable training method, then optimize the boundary box constraint size of the basic model according to the characteristics of small vehicle targets in satellite videos, and finally add the scSE attention mechanism to fully mine effective feature information.

[0052] II. Target tracking and accuracy evaluation: Screen out the training model with the optimal precision, and construct an anti-occlusion module, which is combined with this model to form a satellite video moving vehicle tracking method based on feature enhancement and position prediction. Determine the hyperparameter combination of the anti-occlusion module through grid search, then track the satellite video vehicle target, obtain the tracking result, and conduct quantitative and qualitative evaluations to determine a better hyperparameter combination of the anti-occlusion module to achieve target tracking.

[0053] For the satellite video moving vehicle tracking method based on feature enhancement and position prediction proposed by the present invention, its model structure is as Figure 2 shown. Aiming at the problem that the prior art does not fully mine the effective feature information of the target, the first stage of the present invention is specifically:

[0054] (1) Optimization of boundary box constraints

[0055] To prevent the target bounding box sizes predicted by the Siamese region proposal network (SiamRPN) from being too small and not conforming to the actual situation, considering the characteristic that the target sizes in ordinary videos are mostly larger than 10×10 pixels, the original SiamRPN imposed a constraint of 10×10 pixels on the predicted target bounding box sizes. However, the present invention is for tracking moving vehicle targets in satellite videos. The vehicle targets in satellite videos are small, and overly large target bounding boxes may contain too much background information and similar targets, resulting in low model tracking accuracy. To address the above problems, the present invention reduces the target bounding box constraint of SiamRPN from 10×10 pixels to 5×5 pixels, so that the SiamRPN prediction results are more in line with the target size and the model tracking accuracy is improved.

[0056] (2) Feature enhancement based on the scSE attention mechanism

[0057] In order to enable the tracking model to pay more attention to the features beneficial to target tracking and suppress useless feature information, considering that the scSE attention mechanism includes a spatial attention mechanism and a channel attention mechanism, which can meet the need to enhance features from spatial and channel positions, and has a simple structure and does not require a large amount of data for training, the scSE is combined with SiamRPN to construct a SiamRPN satellite video moving vehicle tracking model based on the scSE attention mechanism, alleviating the model drift problem under similar interference and motion blur interference.

[0058] sSE attention mechanism

[0059] sSE is the spatial attention mechanism, which processes the two-dimensional spatial information of the feature map, as Figure 3 shown. First, the input feature map of H×W×C is convolved through a 1×1×C×1 convolutional layer to obtain a feature map of H×W×1. Then, the values of this feature map are mapped to [0, 1] through the Sigmoid activation function, representing the importance of different spatial position information. Finally, it is multiplied element-wise with the original feature map to obtain the feature after spatial calibration. sSE is the spatial attention mechanism, which processes the two-dimensional spatial information of the feature map (see Figure 3 ).

[0060] cSE attention mechanism

[0061] cSE is the channel attention mechanism, which is used to autonomously learn the weights of each channel, as Figure 3As shown in the figure, a feature map of size H×W×C is input into a global average pooling layer to generate C feature values of size 1×1×C, corresponding to the weights of C channels. Then, a convolution operation of 1×1×C×C / 2 is performed on it, and the convolution result is passed through the non-linear activation function ReLU. The obtained output is then passed through a convolution operation of 1×1×C / 2×C to obtain a tensor of size 1×1×C. This tensor is passed through a Sigmoid layer to normalize each value to the value range [0, 1], which is the weight of each channel. Finally, it is multiplied by the original feature map to enhance key features while suppressing redundant features.

[0062] scSE attention mechanism

[0063] scSE is a hybrid attention mechanism, as Figure 3 shown in the figure. The original feature map passes through sSE and cSE respectively to obtain features enhanced in the spatial and channel dimensions. These two feature maps are added together, and then the final feature map is output through an activation function. Therefore, scSE can re-calibrate the feature map along the spatial and channel dimensions respectively and then merge the output, effectively enhancing the model's attention to effective information in both the spatial and channel aspects, which is beneficial to improving the model's performance.

[0064] Aiming at the problem that the model cannot continuously track the target after the target is occluded in the prior art, the second stage of the present invention specifically is:

[0065] (3) Construct an anti-occlusion module (Anti-occlusion module, AOM) that combines tracking state monitoring and position prediction

[0066] To solve the problem of constructing a tracking state monitoring module F for the classification score map of SiamRPN after enhancing features using the scSE attention mechanism max -APCE-CSL, as Figure 4 shown in the figure. When it is detected that the target is occluded, the inertial mechanism is used to predict and correct the position of the occluded target, realizing continuous tracking of the moving vehicle.

[0067] Tracking state monitoring

[0068] SiamRPN will generate 10 layers of classification score maps, as Figure 5 shown in the figure, which respectively represent the probabilities of 5 kinds of anchor boxes with aspect ratios of 0.33, 0.5, 1, 2, and 3 being the background (layer1-layer5) or the foreground (layer6-layer10). SiamRPN only uses the maximum value F in layer6-layer10 max corresponding position to determine the target position. When the target is occluded, similar targets within the search area will obtain the maximum score. At this time, the tracker still relies on F maxTo determine the position of the target, resulting in the loss of the true target. The present invention first uses the classification score map to calculate multiple tracking status monitoring indicators to determine whether the target is occluded, so as to decide whether to predict the position of the occluded target. Specifically, the maximum value F of the classification score map is used max 、Average Peak-to-Correlation Energy (APCE) and Centralized Sidelobe Leakage (CSL) to construct a tracking status monitoring module F max -APCE-CSL.

[0069] APCE can reflect the fluctuation degree of the score map and the confidence level of the tracking result. When encountering an occluder, the score map fluctuates violently. The calculation method is shown in Equation (1).

[0070]

[0071] In the formula, F represents the score map of size N×N, F i,j represents the value at the i-th row and j-th column in the score map, F max 、F min represent the maximum and minimum values of the score map respectively.

[0072] CSL can quantitatively describe the steepness of the peak of the score map. The lower the CSL value, the more reliable the tracking result and the better the tracking status. The calculation method is shown in Equation (2).

[0073]

[0074] In the formula, F represents the score map of size N×N, F i,j represents the value at the i-th row and j-th column in the score map, μ and v respectively represent the row and column corresponding to the peak of the score map.

[0075] Using only a single tracking status monitoring indicator to determine whether the target is occluded is prone to misjudgment and has low robustness. The present invention comprehensively uses the peak value F of the score map max 、APCE and CSL to construct a tracking status monitoring module F max -APCE-CSL, and the structure is as Figure 4 shown.

[0076] Specifically, when F max is less than the threshold T1 and APCE is less than the threshold T2, or when F max is less than the threshold T1 and CSL is greater than the threshold T3, it is determined that the target is in an occluded state. Among them, T1, T2, and T3 respectively take the F corresponding to the previous n frames of the current frame max, the average values of APCE and CSL.

[0077] Position prediction

[0078] The vehicle targets in satellite videos move relatively slowly between adjacent frames, and their trajectories are continuous and nearly linear. Kalman filtering and inertial mechanisms are often used for target trajectory prediction. Some studies have pointed out that for moving targets in satellite videos, Kalman filtering can converge only after 30 - 50 frames. The inertial mechanism directly predicts the target position in the next frame based on the historical position information of the target in the previous few frames, and the movement of the target between adjacent frames is close to uniform linear motion. Therefore, the present invention uses the inertial mechanism to predict the position and the size of the target bounding box when the target is occluded. The calculation methods are shown in Equations (3), (4), and (5).

[0079]

[0080]

[0081]

[0082] In the formula, x t , y t respectively represent the pixel coordinates of the center point of the target bounding box in the x and y directions in the t-th frame, w t , h t respectively represent the width and height of the target bounding box in the t-th frame, w t-1 , h t-1 represent the width and height of the target bounding box in the (t - 1)-th frame.

[0083] The anti-occlusion module flow constructed by the present invention is as Figure 6 shown below:

[0084] (1.1) Input

[0085] The classification score map of the t-th frame; the target historical bounding box information (x, y, w, h); three lists F, A, C, which are respectively used to store the historical F max , APCE, and CSL values; hyperparameters n1, n2, n3, where n1 > n2.

[0086] (1.2) Tracking state monitoring

[0087] ① Calculate the F max , APCE, and CSL values of the score map of the t-th frame, and then store them into the corresponding F, A, C lists respectively for mean value calculation;

[0088] ② Determine whether t is less than n1: If t < n1, it indicates that the tracking of the target has just started. By default, the target is not occluded at this time, so there is no need to call the anti-occlusion module. Directly use the tracking result of the target tracking method as the final target bounding box. If t >= n1, perform the subsequent operations of the anti-occlusion module;

[0089] ③ Calculate the average values of F max , APCE, and CSL within the historical n2 frames, and set them as thresholds T1, T2, and T3 respectively;

[0090] ④ Use the constructed F max -APCE-CSL index to monitor the target tracking status. Determine whether the target is occluded. If so, perform the subsequent position prediction operation. Otherwise, exit the anti-occlusion module and use the tracking result of the target tracking method as the final target tracking result.

[0091] (1.3) Position prediction

[0092] First, calculate the average displacements Δx and Δy of the target in the x and y directions within n1 frames. To prevent the introduction of noise, update Δx and Δy every n3 frames. Then, based on the position and bounding box size of the target in the (t - 1) frame, predict the position and bounding box size x t 、y t 、w t 、h t .

[0093] The present invention optimizes the bounding box constraint size of SiamRPN to make it more suitable for the small target tracking task of satellite videos. By introducing the scSE attention mechanism, it fully excavates the key feature information of the target and reduces the drift probability of the model under similar interferences and motion blur interferences. Then, a F max -APCE-CSL tracking status monitoring module is constructed using the classification score map of SiamRPN after enhancing the features with the scSE attention mechanism, and the position of the occluded target is predicted and corrected using the inertial mechanism, realizing the continuous tracking of moving vehicles.

[0094] Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A satellite video moving vehicle tracking method based on feature enhancement and position prediction, characterized in that, It includes the following steps: S1. Construct a satellite video target tracking dataset and divide the dataset into a training set and a test set; S2. Adopt the Siamese Region Proposal Network (SiamRPN), and then optimize the constrained size of the SiamRPN bounding box; S3. Add the scSE hybrid attention mechanism to construct a SiamRPN satellite video moving vehicle tracking model based on the scSE attention mechanism, and then input the training set to train the model to obtain the training model with the optimal accuracy; S4. Input the test set into the training model with the optimal accuracy. Based on the classification score map of SiamRPN after enhancing features through the scSE attention mechanism, construct the tracking state monitoring module F max -APCE-CSL, and then calculate the tracking state monitoring metrics using the classification score map to determine whether the target is occluded; The tracking status monitoring module F max -APCE-CSL includes the peak value of the score map F max , the fluctuation degree of the score map, and three tracking status monitoring indicators of the confidence level APCE of the tracking result and the steepness CSL of the peak value of the score map; The specific determination of whether the target is occluded includes: When F max is less than the threshold T1 and APCE is less than the threshold T2, or when F max is less than the threshold T1 and CSL is greater than the threshold T3, it is determined that the target is in an occluded state; where T1, T2, and T3 are respectively the average values of F max , APCE, and CSL corresponding to the previous n frames of the current frame; S5. When it is judged that the target is not in an occluded state, use the tracking result as the final target tracking result; S6. When it is judged that the target is in an occluded state, use the inertial mechanism to predict its position and the size of the target bounding box to achieve continuous tracking.

2. The satellite video moving vehicle tracking method based on feature enhancement and position prediction according to claim 1, wherein: In step S2, the specific optimization of the constrained size of the SiamRPN bounding box is: reducing the target bounding box constraint of SiamRPN from 10×10 pixels to 5×5 pixels.

3. The satellite video moving vehicle tracking method based on feature enhancement and position prediction according to claim 1, wherein: The scSE hybrid attention mechanism includes the sSE spatial attention mechanism and the cSE channel attention mechanism; the original feature map passes through sSE and cSE respectively to obtain the features enhanced in the spatial and channel dimensions, and these two feature maps are added together to output the final classification score map.

4. The satellite video moving vehicle tracking method based on feature enhancement and position prediction according to claim 3, wherein: The sSE spatial attention mechanism is used to process the two-dimensional spatial information of the feature map: first, the input feature map of H×W×C is convolved through a 1×1×C×1 convolutional layer to obtain a feature map of H×W×1, and then this feature map is mapped to the range [0, 1] through the Sigmoid activation function to represent the importance degree of different spatial position information. Finally, it is multiplied element-wise with the original feature map to obtain the feature after spatial calibration; The cSE channel attention mechanism is used to autonomously learn the weights of each channel: first, the feature map of size H×W×C is input into a global average pooling layer to generate 1×1×C feature values corresponding to the weights of C channels, and then a convolutional operation of 1×1×C×C / 2 is performed on it. The output of the convolutional result passes through the non-linear activation function ReLU, and the obtained output passes through another convolutional operation of 1×1×C / 2×C to obtain a tensor of 1×1×C. This tensor is normalized to the value range [0, 1] through the Sigmoid layer, that is, the weights of each channel. Finally, it is multiplied with the original feature map to obtain the feature after channel calibration.

5. The satellite video moving vehicle tracking method based on feature enhancement and position prediction according to claim 1, characterized in that: APCE reflects the fluctuation degree of the score map and the confidence level of the tracking result, and the expression is as follows: CSL quantitatively describes the steepness of the peak of the score map, and the expression is as follows: where F represents a score map of size N×N, and F i,j represents the value at the i-th row and j-th column in the score map, and F max , F min represent the maximum and minimum values of the score map respectively, and μ and v represent the row and column corresponding to the peak of the score map.

6. The satellite video moving vehicle tracking method based on feature enhancement and position prediction according to claim 1, characterized in that: In step S6, the specific method of using the inertial mechanism to predict its position and the size of the target bounding box is as follows: First, calculate the average displacements Δx and Δy of the target in the x and y directions within n frames. To prevent the introduction of noise, update Δx and Δy every m frames. Then, based on the position and the size of the bounding box of the target in the (t - 1)th frame, predict the position and the size of the bounding box of the target in the tth frame, x t 、y t 、w t 、h t , as shown in the following formula: where x t and y t represent the pixel coordinates of the center point of the target bounding box in the x and y directions at the t-th frame, w t and h t represent the width and height of the target bounding box at the t-th frame, respectively, and w t-1 and h t-1 represent the width and height of the target bounding box at the (t - 1)-th frame.

Images