A small target intelligent recognition and tracking method based on UAV video data

Abstract

The present invention relates to the field of image processing technology, and in particular to a small target intelligent recognition and tracking method based on unmanned aerial vehicle video data, comprising the following steps: step one, data preprocessing; step two, small target detection: based on the improved small target detection algorithm and multi-target tracking algorithm of YOLOv5, an improved YOLOv5 model, i.e., the YOLOv5-improved model, is used to perform small target detection on the preprocessed image; step three, multi-target tracking: the result of small target detection is input into the multi-target tracking algorithm, and the Kalman filter and the Hungarian algorithm are used as tracking components; step four, space and time verification; step five, output recognition and tracking results. The present invention is used to solve the problem of poor performance of small target objects due to factors such as small target size, few features, and interference from similar objects by optimizing the target detection algorithm, and improve the accuracy and robustness of military target detection from the perspective of unmanned aerial vehicles.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a small target intelligent recognition and tracking method based on unmanned aerial vehicle video data. Background Art

[0002] The application of drone technology in the military field is increasingly highlighting its strategic value, especially in the detection of military targets. With the rapid progress and widespread application of drone technology, military target detection from the perspective of drones has become a key area of ​​military science and technology research. This technology uses the image and video data captured by the high-resolution camera on the drone to achieve automatic identification, positioning and tracking of key targets, providing strong support for the collection of military intelligence, target monitoring, tactical planning and precision strikes. Military target detection from the perspective of drones can not only provide comprehensive target information in a vast battlefield space, but also significantly reduce the risk of personnel directly participating in reconnaissance and target search missions, and improve combat effectiveness and safety.

[0003] Although current target detection algorithms, such as YOLOv5, have made significant progress in speed and accuracy, there are still some urgent problems to be solved in the task of military target detection under the perspective of drones. The YOLOv5 algorithm achieves efficient target detection through its optimized network structure, such as Mosaic data enhancement, adaptive anchor box calculation, Focus structure and CSP structure, FPN feature pyramid network and GIOU Loss loss function. However, the algorithm still faces the following challenges in the task of small target detection under the perspective of drones: First, the YOLOv5 model uses a coupling head to simultaneously generate the rectangular box size, confidence and category information of the target. However, the classification task and the localization task have different focuses on the feature map. This coupling design may lead to performance loss, especially in small target detection. Due to limited feature information, the design of the coupling head may not fully utilize the key information in the feature map; second, in the small target detection under the perspective of drones, the positive and negative sample matching strategy of YOLOv5 may lead to insufficient quality of positive samples. Since small targets account for a small proportion of the image, the number and quality of positive samples are limited, which will affect the training effect and detection performance of the model; finally, the change in the distance between the drone and the target during flight causes a large number of scale changes in the collected images, and motion blur and occlusion problems also occur from time to time. These problems make the detection of small targets more difficult, especially under complex backgrounds and lighting conditions, where the feature information of small targets is weaker and is easily misdetected or missed.

[0004] In view of the performance loss caused by the above-mentioned coupling head, the positive and negative sample matching strategy problem, and the target scale change and occlusion problem, it is particularly important to propose a small target intelligent recognition and tracking method based on UAV video data. Summary of the invention

[0005] To solve the above problems, the present invention provides a small target intelligent recognition and tracking method based on UAV video data, which optimizes the target detection algorithm to solve the poor performance problem of small target objects due to factors such as small target size, few features and interference from similar objects, and improves the accuracy and robustness of military target detection from the perspective of the UAV.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows: a small target intelligent recognition and tracking method based on drone video data, comprising the following steps:

[0007] Step 1: Data preprocessing: normalize the pixel matrix of the images in the video data captured by the drone, normalize the pixel value range, and convert the image data type from uint8 to float32 as model input;

[0008] Step 2: Small target detection: Based on the improved small target detection algorithm and multi-target tracking algorithm of YOLOv5, the improved YOL0v5 model, namely the YOLOv5-improved model, is used to detect small targets in the preprocessed images. The YOLOv5-improved model includes a small target detection head, a lightweight decoupling detection head, a C3-SepVIT module, and a SimOTA sample matching strategy.

[0009] Step 3: Multi-target tracking: Input the results of small target detection into the multi-target tracking algorithm, using the Kalman filter and Hungarian algorithm as tracking components;

[0010] Step 4: Space and time verification: Use the height estimation model based on deep learning for space verification, and the target tracking model based on Kalman filter and Hungarian algorithm for time verification. Perform a secondary verification on the recognition result of the YOL0v5-improved model, merge the results of the space verification and time verification, and output a comprehensive verification result.

[0011] Step 5: Output recognition and tracking results: Output the verified recognition and tracking results for subsequent processing and analysis.

[0012] Furthermore, in step 1, when the pixel matrix is ​​normalized, the pixel values ​​are normalized from the range [0, 255] to the range [0.0, 1.0].

[0013] Furthermore, in step 2, the small target detection head is set on the low-level feature map of the YOL0v5 model, and the precise position information of the low-level feature map and the preset small-size anchor frame design are used to improve the positioning and classification accuracy of small target objects.

[0014] Furthermore, in step 2, the lightweight decoupled detection head is to separate the positioning task and the classification task by decoupling the coupled detection head of YOL0v5, thereby reducing the impact of bounding box disturbance on small target detection performance.

[0015] Furthermore, in step 2, the C3-SepVIT module combines the separable self-attention calculation method in the SepVIT model, and by mixing the CNN and Transformer models, enables the model to have the ability to model global features without increasing the amount of calculation too much, so as to improve the problem of mutual interference between similar targets.

[0016] Furthermore, in step 2, the SimOTA sample matching strategy is used to improve the positive and negative sample matching strategy of YOLOv5 by assigning high-quality positive samples to each GT box and determining the remaining samples as negative samples, thereby performing loss calculation to improve the model's detection performance for small targets.

[0017] Furthermore, in step three, the Kalman filter is used to predict the current state of the target and update the state estimate of the target by fusing the dynamic model and the observation data;

[0018] The Hungarian algorithm is used to solve the data association problem by minimizing the association cost between target detection and target tracking, and matching the target of the current frame with the tracked target of the previous frame.

[0019] Further, in step three, the multi-target tracking algorithm process is:

[0020] S31, predicting the state of the next frame through the Kalman filter;

[0021] S32, uses the Hungarian algorithm to solve the data association problem, that is, how to match the detected target in the current frame with the tracked target in the previous frame.

[0022] Furthermore, in step 4, the spatial verification is to estimate the actual distance and size of the target according to its height information in the image through the height estimation model, and compare it with the recognition result of the YOL0v5-improved model to preliminarily judge the recognition result;

[0023] The time verification is to track the motion trajectory and morphological changes of the target in the frames before and after the current frame through the target tracking model, and compare them with the recognition results of the YOL0v5-improved model to further verify the recognition results.

[0024] Further, in step 4, the steps to obtain the comprehensive verification results are:

[0025] S41, the spatial verification result and the temporal verification result are represented by two binary variables respectively: 0 indicates that the verification is passed, and 1 indicates that the verification is not passed;

[0026] S42, performing a logical AND operation on the two binary variables to obtain a final comprehensive verification result.

[0027] The above scheme has the following beneficial effects:

[0028] 1. This solution adopts the YOL0v5-improved model, which includes a small target detection head, a lightweight decoupling detection head, a C3-SepVIT module, and a SimOTA sample matching strategy. These innovative designs jointly improve the detection performance of small targets from the perspective of drones. The small target detection head performs small target detection on low-level feature maps, improving the positioning and classification accuracy; the lightweight decoupling detection head reduces the impact of bounding box disturbances, allowing the model to focus more on the feature extraction of small targets; the C3-SepVIT module enhances the global feature modeling capability of the model and effectively improves the problem of mutual interference between similar targets; the SimOTA sample matching strategy ensures that each target can be assigned a high-quality positive sample, thereby improving the detection accuracy.

[0029] 2. This method not only improves the detection accuracy of small targets, but also enhances the robustness and reliability of the system through multi-target tracking algorithms and secondary verification mechanisms. The combined use of Kalman filter and Hungarian algorithm makes target tracking more stable and can effectively deal with problems such as target occlusion and rapid movement in complex scenes. The spatial and temporal verification further verifies the recognition results of the YOL0v5-improved model, ensuring the accuracy and reliability of the final output results.

[0030] 3. This solution introduces a spatial and temporal verification model to perform a secondary verification on the recognition results of the YOLOv5-improved model. The spatial verification uses the height estimation model to estimate the actual distance and size of the target, while the temporal verification tracks the movement trajectory and morphological changes of the target through the target tracking model. This dual verification mechanism greatly improves the accuracy and reliability of the recognition results and reduces the risk of false positives and false negatives.

[0031] 4. In this solution, since there are many difficulties in the target detection task from the perspective of drones, such as small target size, few features, interference from similar objects, etc., the adaptability and robustness of the model are crucial. This invention can better deal with these difficult problems by introducing the spatial and temporal verification model, thereby enhancing the adaptability and robustness of the model.

[0032] 5. In this scheme, in the data preprocessing stage, the method performs pixel matrix normalization on the video data captured by the drone and converts the image data type into a format more suitable for model input, which helps to reduce data noise and improve model training efficiency. At the same time, this method also reduces the amount of calculation and video memory usage by optimizing the model structure and algorithm flow, making the entire system more efficient and practical in practical applications.

[0033] 6. This scheme not only improves the accuracy of the recognition results through the space and time verification model, but also provides intuitive verification results. Through the representation of binary variables and logical AND operations, the comprehensive verification results can clearly show the reliability of the recognition results. This provides strong data support and analysis basis for subsequent algorithm optimization, tactical planning and combat command. Since this method performs well in target detection and tracking, it can significantly improve the capabilities of drones in real-time monitoring and tactical decision-making. This is of great significance for military reconnaissance, target tracking and other tasks, and can provide commanders with more accurate and timely target information, so as to make more informed decisions.

[0034] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the method steps of an embodiment of a method for intelligently identifying and tracking small targets based on drone video data of the present invention;

[0036] Figure 2 A schematic diagram of a small target optimization model YOLOv5-improved according to an embodiment of a small target intelligent recognition and tracking method based on drone video data of the present invention;

[0037] Figure 3 A schematic diagram of a tank recognized from a close side view of a model in an embodiment of a small target intelligent recognition and tracking method based on unmanned aerial vehicle video data of the present invention;

[0038] Figure 4 A schematic diagram of a tank recognized from a long distance by looking down at a model of an embodiment of a small target intelligent recognition and tracking method based on unmanned aerial vehicle video data of the present invention;

[0039] Figure 5 Schematic diagram of model recognition of tanks and people in an embodiment of the small target intelligent recognition and tracking method based on drone video data of the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0042] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0043] The following is further described in detail through specific implementation methods:

[0044] Embodiment 1:

[0045] As attached Figure 1 As shown: A small target intelligent recognition and tracking method based on drone video data includes the following steps:

[0046] Step 1: Data preprocessing: Perform pixel matrix normalization on the images in the video data captured by the drone. During pixel matrix normalization, the pixel values ​​are normalized from the range [0, 255] to the range [0.0, 1.0], and the image data type is converted from uint8 to float32 as model input.

[0047] Step 2, small target detection: Based on the improved small target detection algorithm and multi-target tracking algorithm of YOLOv5, the improved YOLOv5 model, namely the YOLOv5-improved model, is adopted. The improved small target detection algorithm is: the image preprocessed in step 1 is passed into the model to obtain the predicted box. Then the loss function is calculated. The loss function consists of three parts: classification loss, confidence loss, and positioning loss. The classification loss is used to measure the gap between the predicted category and the true category, the confidence loss is used to measure the gap between the confidence of the predicted box and the confidence of the true box, and the positioning loss is used to measure the gap between the position of the predicted box and the position of the true box. The total loss function is Loss = λ1Lcls + λ2Lobj + λ3Lloc, where λ1 = 0.5, λ2 = 1, and λ3 = 0.05.

[0048] In mixed precision training, first create a gradient scaler object to manage the gradient scaling operation. This is done by using the amp.GradScaler() function, and then using scaler.scale(loss).backward(). Mixed precision training can improve the speed and stability of model training.

[0049] In each round, the gradient and learning rate are updated, and the learning rate strategy is applied to update the learning rate. It is important to update the gradient. By adjusting the step size of model training, the model can converge better.

[0050] EMA (Exponential Moving Average) is a technique used to stabilize and improve the model training process. By updating EMA, the model parameters can be made smoother and the performance of the model on the test set can be improved.

[0051] During the training process, the trained model is saved regularly so that it can be evaluated and tested after the training is completed. Here, the last.pt model is saved and the best.pt model is output. Usually, best.pt is the model that performs best on the validation set.

[0052] The algorithm process continuously iterates the above steps to train an object detection model that can perform well on the test set.

[0053] Small target detection is performed on the preprocessed image. The YOLOv5-improved model includes a small target detection head, a lightweight decoupling detection head, a C3-SepVIT module, and a SimOTA sample matching strategy:

[0054] The tiny target detection head is set on the low-level feature map of the YOLOv5 model, and uses the precise location information of the low-level feature map and the preset tiny-size anchor frame design to improve the positioning and classification accuracy of small target objects.

[0055] The lightweight decoupled detection head decouples the coupled detection head of YOLOv5 to separate the positioning task and the classification task, thereby reducing the impact of bounding box perturbations on small target detection performance.

[0056] The C3-SepVIT module combines the separable self-attention calculation method in the SepVIT model with the hybrid CNN and Transformer models to enable the model to have the ability to model global features without increasing the amount of calculation too much, thereby improving the problem of interference between similar targets.

[0057] The SimOTA sample matching strategy is used to improve the positive and negative sample matching strategy of YOLOv5. By assigning high-quality positive samples to each GT box and determining the remaining samples as negative samples, loss calculation is performed to improve the model's detection performance for small targets.

[0058] Step 3, multi-target tracking: Input the results of small target detection into the multi-target tracking algorithm, using the Kalman filter and Hungarian algorithm as tracking components.

[0059] The Kalman filter is used to predict the current state of the target and update the target's state estimate by fusing the dynamic model and observation data. The Kalman filter is divided into prediction and update. The prediction predicts the current state through the previous state quantity and distribution, and the update updates the current state quantity and covariance and other parameters using the observed value and the predicted value.

[0060] The Hungarian algorithm is used to solve the data association problem. It matches the target of the current frame with the tracked target of the previous frame by minimizing the association cost between target detection and target tracking. In the multi-target tracking algorithm, it matches the target of the current frame with the tracked target of the previous frame by minimizing the association cost between target detection and target tracking (such as position overlap, appearance similarity, etc.), thereby effectively associating the targets and maintaining the continuity and consistency of the targets. The Hungarian algorithm assigns the target ID in the current frame by calculating the minimum iou loss between the targets in the current frame and the previous frame. If the iou of the two frames is large (i.e., the loss is small), then they are very likely to be the same target of the frame, so that the tracking ID in the previous frame is assigned to the current new value, thereby performing association matching.

[0061] The multi-target tracking algorithm process is:

[0062] S31, predict the state of the next frame through the Kalman filter.

[0063] S32, using the Hungarian algorithm to solve the data association problem, that is, how to match the detection target in the current frame with the tracked target in the previous frame. In this process, both unmatched tracks and unmatched detections will be processed. If the adaptation lasts T times, the target ID will be deleted from the target to be tracked. If no track can match the detection, then a new track is assigned to the detection.

[0064] Step 4: Spatial and temporal verification: A deep learning-based height estimation model is used for spatial verification. The spatial verification is to estimate the actual distance and size of the target based on its height information in the image through the height estimation model, and compare it with the recognition result of the YOLOv5-improved model to preliminarily judge the recognition result.

[0065] And the target tracking model based on Kalman filter and Hungarian algorithm is used for time verification. The time verification is to track the motion trajectory and morphological changes of the target in the frames before and after the current frame through the target tracking model, and compare them with the recognition results of the YOLOv5-improved model to further verify the recognition results.

[0066] The recognition results of the YOLOv5-improved model are double-checked, and the results of the spatial check and the temporal check are integrated to obtain the comprehensive check results:

[0067] S41, the spatial verification result and the temporal verification result are represented by two binary variables respectively: 0 indicates that the verification is passed, and 1 indicates that the verification is not passed.

[0068] S42, performing a logical AND operation on the two binary variables to obtain a final comprehensive verification result, and outputting a comprehensive verification result.

[0069] Step 5: Output recognition and tracking results: Output the verified recognition and tracking results for subsequent processing and analysis.

[0070] The specific implementation process is as follows: After the model training is completed, the model is deployed to the edge computing device. This embodiment uses RDK X3 4GB as the edge computing device in the example. The steps for deploying the YOLOv5 model are as follows:

[0071] S1, model deployment environment. This includes installing some necessary packages, such as matplotlib, numpy, opencv-python, Pillow, PyYAML, requests, scipy, tqdm, tensorboard, pandas, seaborn, ipython, psutil, and thop. These packages are installed using the pip command to ensure that they are the latest versions.

[0072] S2, download the source code of YOLOv5 from GitHub. Select version v2.0 and download the source code to your desktop.

[0073] S3, unzip the downloaded source code to an appropriate location for subsequent use. Put the video file to be detected into the yolov5-master folder, and put the pre-trained model file into the yolov5-master / models folder.

[0074] S4, enter the yolov5-master folder in the command line and run the detection command.

[0075] S5, after model training, the model is deployed to the edge computing device and tested according to the requirements of the military target detection system from the perspective of the drone.

[0076] After testing, the video transmission frame rate reached 30FPS, which meets the requirements. The real-time video stream can be played smoothly, and the testers can quickly and timely observe and analyze the target information captured by the drone.

[0077] The video resolution adopts the standard of 1920×1080, and the results show that the video quality meets the requirements. The playback picture is clear, and the testers can clearly identify various details of the target and accurately detect and identify the target.

[0078] For target detection in the military field, it can accurately identify more than 10 different target types including tanks, armored vehicles, buildings, etc., with an identification accuracy rate of 85%.

[0079] The test results are as follows Figure 3 , Figure 4 and Figure 5 .

[0080] Embodiment 2:

[0081] As attached Figure 1 As shown in Figure 1, the difference from Example 1 is that in terms of the success rate of continuous frame tracking, the real trajectory of the target is first manually annotated in the test data set, and then matched with the target trajectory output by the model and measured by Euclidean distance. The results show that the success rate of continuous frame tracking reaches 90%.

[0082] Embodiment 3:

[0083] As attached Figure 1 As shown, the difference from Example 2 is that, in order to verify the effectiveness and reliability of the present invention, a large number of experimental tests were carried out. The video file to be detected is placed in the yolov5-master folder, and the pre-trained model file is placed in the yolov5-master / models folder. Then, enter the yolov5-master folder in the command line and run the detection command to perform target detection. Then, the detection result is double-checked using the space and time verification model.

[0084] The experimental results show that after the verification model is introduced, the recognition accuracy is significantly improved. At the same time, the false alarm rate is also significantly reduced. Specifically, in terms of the success rate of continuous frame tracking, the real trajectory of the target is first manually annotated in the test data set, and then matched with the target trajectory output by the model. The results show that after the verification model is introduced, the success rate of continuous frame tracking reaches 95%, which is about 5 percentage points higher than when the verification model is not introduced. In addition, excellent performance has also been achieved in terms of video transmission frame rate, video resolution, and target recognition accuracy. These experimental results fully demonstrate the effectiveness and reliability of the present invention.

[0085] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A small target intelligent recognition and tracking method based on drone video data, characterized in that: The following steps are involved: Step 1: Data preprocessing: normalize the pixel matrix of the images in the video data captured by the drone, normalize the pixel value range, and convert the image data type from uint8 to float32 as model input; Step 2: Small target detection: Based on the improved small target detection algorithm and multi-target tracking algorithm of YOLOv5, the improved YOLOv5 model, namely the YOL0v5-improved model, is used to detect small targets in the preprocessed images. The YOLOv5-improved model includes a small target detection head, a lightweight decoupling detection head, a C3-SepVIT module, and a SimOTA sample matching strategy. Step 3: Multi-target tracking: Input the results of small target detection into the multi-target tracking algorithm, using the Kalman filter and Hungarian algorithm as tracking components; Step 4: Space and time verification: Use the height estimation model based on deep learning for space verification, and the target tracking model based on Kalman filter and Hungarian algorithm for time verification. Perform a secondary verification on the recognition result of the YOL0v5-improved model, merge the results of the space verification and time verification, and output a comprehensive verification result. Step 5: Output recognition and tracking results: Output the verified recognition and tracking results for subsequent processing and analysis.

2. The small target intelligent recognition and tracking method based on drone video data according to claim 1 is characterized in that: In step 1, when the pixel matrix is ​​normalized, the pixel values ​​are normalized from the range [0, 255] to the range [0.0, 1.0].

3. The small target intelligent recognition and tracking method based on drone video data according to claim 2 is characterized in that: In step 2, the small target detection head is set on the low-level feature map of the YOL0v5 model, and the precise location information of the low-level feature map and the preset small-size anchor frame design are used to improve the positioning and classification accuracy of small target objects.

4. The small target intelligent recognition and tracking method based on drone video data according to claim 3 is characterized in that: In step 2, the lightweight decoupled detection head is to separate the positioning task and the classification task by decoupling the coupled detection head of YOLOv5, thereby reducing the impact of bounding box disturbance on small target detection performance.

5. The small target intelligent recognition and tracking method based on drone video data according to claim 4 is characterized in that: In step 2, the C3-SepVIT module combines the separable self-attention calculation method in the SepVIT model, and by mixing the CNN and Transformer models, enables the model to have the ability to model global features without increasing the amount of calculation too much, so as to improve the problem of interference between similar targets.

6. The small target intelligent recognition and tracking method based on drone video data according to claim 5 is characterized in that: In step 2, the SimOTA sample matching strategy is used to improve the positive and negative sample matching strategy of YOLOv5. By assigning high-quality positive samples to each GTbox and determining the remaining samples as negative samples, loss calculation is performed to improve the model's detection performance for small targets.

7. The small target intelligent recognition and tracking method based on drone video data according to claim 6 is characterized in that: In step 3, the Kalman filter is used to predict the current state of the target and update the target state estimate by fusing the dynamic model and observation data; The Hungarian algorithm is used to solve the data association problem by minimizing the association cost between target detection and target tracking, and matching the target of the current frame with the tracked target of the previous frame.

8. The small target intelligent recognition and tracking method based on drone video data according to claim 7 is characterized in that: In step three, the multi-target tracking algorithm process is: S31, predicting the state of the next frame through the Kalman filter; S32, uses the Hungarian algorithm to solve the data association problem, that is, how to match the detected target in the current frame with the tracked target in the previous frame.

9. The small target intelligent recognition and tracking method based on drone video data according to claim 8 is characterized in that: In step 4, the spatial verification is to estimate the actual distance and size of the target based on its height information in the image through the height estimation model, and compare it with the recognition result of the YOLOv5-improved model to preliminarily judge the recognition result; The time verification is to track the motion trajectory and morphological changes of the target in the frames before and after the current frame through the target tracking model, and compare them with the recognition results of the YOLOv5-improved model to further verify the recognition results.

10. The small target intelligent recognition and tracking method based on drone video data according to claim 9 is characterized in that: In step 4, the steps to obtain the comprehensive verification results are: S41, the spatial verification result and the temporal verification result are represented by two binary variables respectively: 0 indicates that the verification is passed, and 1 indicates that the verification is not passed; S42, performing a logical AND operation on the two binary variables to obtain a final comprehensive verification result.

Images