Unmanned aerial vehicle-based target detection method and system, target tracking method and system

Abstract

The application relates to the technical field of unmanned aerial vehicle target detection, in particular to a target detection method and system based on an unmanned aerial vehicle and a target tracking method and system, which comprises the following steps: acquiring a double-detection-head deep learning network model, the double-detection-head deep learning network model comprising a high-recall detection head for initially screening all potential detection targets, a high-precision detection head for finely classifying the detection targets, and a self-adaptive weight fusion module connected to the output ends of the high-recall detection head and the high-precision detection head; acquiring a current frame image taken by the unmanned aerial vehicle and calculating the real-time image complexity; inputting the current frame image into the double-detection-head deep learning network model, outputting detection results by the high-recall detection head and the high-precision detection head respectively, and dynamically adjusting the fusion weights of the output results of the two detection heads by the self-adaptive weight fusion module according to the real-time image complexity to generate the final target detection result.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) target detection technology, specifically to UAV-based target detection methods and systems, and target tracking methods and systems. Background Technology

[0002] With the rapid development of UAV technology, UAV aerial imagery has been widely used in security monitoring, emergency rescue, and traffic inspection. Deep learning-based target detection technology has become a core tool in UAV visual perception systems, providing crucial information for autonomous decision-making by locating and identifying targets in aerial images in real time. In recent years, to adapt to the limitations of UAV onboard computing resources, power consumption, and payload weight constraints, related research has mainly focused on three aspects: lightweight model design, multi-scale detection head optimization, and edge model deployment. Regarding lightweight models, researchers have replaced traditional backbones with lightweight networks such as depthwise separable convolutions and channel shuffling to reduce the number of parameters and computational complexity. In terms of detection head design, some studies have alleviated the detection difficulties caused by the dense concentration of small targets and drastic scale changes in UAV aerial images by adding small target detection heads or using dual prediction heads to process targets of different scales. Regarding edge deployment, trained models are converted to formats such as ONNX, TensorRT, or RKNN, and INT8 quantization and multi-core NPU parallel inference are employed to improve inference speed on embedded platforms.

[0003] However, the aforementioned existing technologies still have the following shortcomings in practical applications: First, existing UAV target detection methods mostly adopt a single detection head or a multi-detector head structure with tasks assigned according to target scale. There is a lack of dynamic collaborative mechanisms between the detection heads based on scene features, making it impossible to adaptively balance recall and precision in complex and varied aerial photography scenarios. This results in a high false detection rate in complex scenarios with dense targets and severe occlusion, and an increased false negative rate in scenarios with sparse targets and simple backgrounds. Second, existing methods typically use fixed non-maximum suppression (NMS) intersection-over-union (IoU) thresholds and confidence thresholds in the post-processing stage. These threshold settings are difficult to consider the detection requirements of scenarios with varying complexity. Finally, regarding lightweight deployment for domestic edge computing platforms, existing research mainly focuses on engineering implementations such as model format conversion and quantization acceleration, lacking collaborative optimization of algorithm-level detection strategies and hardware inference characteristics. This makes it difficult to simultaneously meet the application requirements of high accuracy and real-time performance under limited computing power constraints. Summary of the Invention

[0004] This invention aims to solve the problem of adaptively adjusting the detection strategy according to the image complexity of the aerial photography scene under the computing power constraints of the UAV edge computing platform, so as to achieve a dynamic balance between recall and precision, while ensuring the lightweight deployment and real-time inference of the model.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following solution: The target detection method based on UAVs includes the following steps: S1: Obtain the dual-detection-head deep learning network model. The dual-detection-head deep learning network model includes: a high-recall detection head for initial screening of all potential detection targets, a high-precision detection head for fine classification of detection targets, and an adaptive weight fusion module connected to the output of the high-recall detection head and the high-precision detection head. S2: Obtain the current frame image captured by the drone and calculate the real-time image complexity of the current frame image; S3: Input the current frame image into the dual-detection-head deep learning network model. The high-recall detection head and the high-precision detection head output the detection results respectively. The adaptive weight fusion module dynamically adjusts the fusion weight of the output results of the two detection heads according to the real-time image complexity to generate the final target detection result. The target detection result includes the detection box set, confidence score and target category.

[0006] Furthermore, in S2, the real-time image complexity of the current frame image is calculated, specifically including: obtaining the edge density, texture entropy, and number of targets detected in the previous frame image of the current frame image; and performing a weighted summation of the edge density, texture entropy, and the normalized number of targets detected in the previous frame image to obtain the real-time image complexity characterizing the background complexity of the current frame image.

[0007] Furthermore, in S3, the adaptive weight fusion module dynamically adjusts the fusion weights of the outputs of the two detection heads based on the real-time image complexity. Specifically, this includes: using the real-time image complexity, the average confidence of the high-recall detection head output, and the average confidence of the high-precision detection head output as input features; mapping the output to an adaptive fusion weight value between 0 and 1 through the fully connected layer and non-linear activation function in the adaptive weight fusion module; and using the adaptive fusion weight value to perform a weighted summation of the outputs of the high-recall detection head and the high-precision detection head to generate the final set of detection boxes, confidence, and target category.

[0008] Furthermore, the model construction process for S1 also includes a nonmaximum suppression step: Based on the real-time image complexity, the cross-union ratio (CUP) threshold and confidence threshold in the non-maximum suppression step are dynamically adjusted. Both the CUP threshold and the confidence threshold are positively correlated with the real-time image complexity. When the real-time image complexity is greater than or equal to the preset complexity threshold, the intersection-union ratio threshold is adjusted to the first intersection-union ratio threshold, and the confidence threshold is adjusted to the first confidence threshold. When the real-time image complexity is less than the preset complexity threshold, the intersection-over-union (IoU) threshold is adjusted to the second IoU threshold, and the confidence threshold is adjusted to the second confidence threshold. Among them, the first crossover ratio (CUP) threshold is greater than the second CUP threshold, and the first confidence threshold is greater than the second confidence threshold.

[0009] Furthermore, S1 also includes staged training of the dual-detection-head deep learning network model, specifically: In the first stage, the parameters of the adaptive weight fusion module are frozen, and the high recall detection head and the high precision detection head are trained independently using the complete dataset, so that the two detection heads each have their own emphasis characteristics. In the second stage, the parameters of the adaptive weight fusion module are unfrozen, and end-to-end joint training is performed on all parameters of the high recall detector, the high precision detector, and the adaptive weight fusion module with a second learning rate lower than that in the first stage. The network parameters are then optimized using the total loss function.

[0010] The target tracking method based on drones includes the following steps: SA1: Obtain the target detection result as described in any of the above items; SA2: Using the center coordinates of the detection boxes in the detection box set, perform target ID matching with the center coordinates of the detection boxes in the current frame image. If the match is successful, a tracking box with an ID is obtained. SA3: Performs state prediction, data association, and trajectory update on the tracking box, and outputs the results to the visualization module.

[0011] Furthermore, in SA3, the tracking frame performs data association: Specifically, it includes: Association with the detection data of the current frame image: Match the prediction bounding box output by the state prediction module with the detection bounding box of the current frame image; Association with historical trajectory data: Combine the appearance features and motion trajectory of the target in historical frame images to determine whether the detection box in the current frame image belongs to an existing target ID; By calculating the spatial and appearance similarity between each predicted bounding box and the detected bounding box, the optimal matching scheme is found. Once a match is successful, the state of the target ID can be updated and the trajectory can be continued.

[0012] Unmanned aerial vehicle (UAV) based target detection systems include: Model acquisition module: used to acquire the dual-detection head deep learning network model, which includes a high-recall detector head and a high-precision detector head, as well as an adaptive weight fusion module connected to the output of the high-recall detector head and the high-precision detector head; Image complexity calculation module: used to acquire the current frame image of the drone aerial photography and calculate the real-time image complexity of the current frame image; The target detection module is used to input the current frame image into the dual-detection-head deep learning network model. The high-recall detection head and the high-precision detection head output the detection results respectively. The adaptive weight fusion module dynamically adjusts the fusion weight of the output results of the two detection heads according to the real-time image complexity to generate the final target detection result. The target detection result includes the detection box set, confidence score and target category.

[0013] Unmanned aerial vehicle (UAV) based target tracking systems include: Target detection result acquisition module: used to acquire target detection results as described in any of the above items; Target tracking module: Used to match the center coordinates of the detection boxes in the target detection results with the center coordinates of the detection boxes in the current frame image to obtain the target ID. If the match is successful, a tracking box with an ID is obtained. Visualization module: This module takes the tracking box as input to the target tracking module for state prediction, data association, and trajectory update, and outputs the results to the visualization module.

[0014] The beneficial effects of this invention are: 1. By calculating the real-time image complexity and inputting it into the adaptive weight fusion module, the system can automatically generate a fusion weight value that matches the complexity of the input image, thereby achieving differentiated fusion of the output results of the high recall detection head and the high precision detection head.

[0015] 2. When the image complexity is high (e.g., dense edges, rich textures, and a large number of targets), the fusion weights automatically favor the output of the high-precision detection head, reducing false detections; when the image complexity is low (e.g., flat background and sparse targets), the fusion weights automatically favor the output of the high-recall detection head, reducing false negatives. The final generated set of detection boxes achieves a dynamic balance between recall and precision overall.

[0016] 3. After format conversion and INT8 quantization, the dual-detection-head deep learning network model can run efficiently on airborne embedded devices (such as RK3588 NPU), completing the entire process of complexity calculation, dual-head parallel inference, weight fusion and non-maximum suppression for each frame of the input video stream, achieving the frame rate requirements for real-time object detection.

[0017] 4. By first freezing the adaptive weight module, training the two detection heads separately, and then unfreezing them for end-to-end joint fine-tuning, the high recall detection head and the high precision detection head can each fully learn the feature representation they emphasize. At the same time, the adaptive weight module can learn the optimal fusion strategy, avoiding the training instability or local optima problems that may be caused by joint training from scratch. Attached Figure Description

[0018] Figure 1 This is a flowchart of the target detection method based on unmanned aerial vehicles (UAVs) of the present invention; Figure 2 This is a flowchart of the processing of the dual-detection-head deep learning network model of the present invention; Figure 3 The image shows the result of drone aerial image recognition using the method of this invention. Figure 4 The image shows the results of drone aerial image recognition at different times using the method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0021] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0022] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0023] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0024] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: Example 1: This example provides a method and system for target detection and tracking based on unmanned aerial vehicles (UAVs), applicable to UAV platforms equipped with an RK3588 onboard computer. For example... Figure 1 As shown, the method includes the following steps: Construct a custom drone aerial image dataset, annotate targets, and divide it into training, validation, and test sets. Calculate the real-time image complexity. The bounding boxes should be as accurate as possible, and the visible parts of partially occluded targets should also be annotated.

[0026] Build as Figure 2 The image shows a dual-detection-head deep learning network model. This model includes a shared feature extraction backbone network (e.g., a lightweight MobileNetV3 or EfficientNet-Lite), and two parallel detection heads: High-recall detection head: In this embodiment, a three-layer convolutional layer is used with a kernel size of 5×5 and a stride of 1. Dilated convolution (with a dilation rate of 2) expands the receptive field to more than 60% of the image size, which is used to initially screen all potential targets and output a first set of detection boxes. and its average confidence score .

[0027] High-precision detection head: In this embodiment, ten convolutional layers are used, with a kernel size of 3×3, a stride of 1, and a receptive field of 30% to 40% of the image size. These layers are used to perform fine classification and bounding box regression on the candidate boxes output by the high-recall detection head, outputting a second set of detection boxes. and its average confidence score .

[0028] The outputs of the two detectors are connected to an adaptive weight fusion module.

[0029] The model is trained in the following three stages: Phase 1: Freeze the adaptive weight fusion module and set its output fusion weights. α The initial value is fixed at 0.5. High recall and high precision detection heads are trained independently using the training set.

[0030] When training the high-recall detector head, the confidence threshold is set to 0.1, and the cross-union ratio (CUP) threshold for non-maximum suppression is set to 0.3 to encourage the model to output more candidate boxes. When training the high-precision detector head, the confidence threshold is set to 0.5, the CUP threshold is set to 0.6, and the weight coefficient of the classification loss in the total loss function is set to twice the weight coefficient of the localization loss to improve classification accuracy. This stage allows the two sub-networks to initially possess their respective focus characteristics.

[0031] Phase 2: Unfreeze the adaptive weight fusion module and set the learning rate to one-tenth of the learning rate in Phase 1 (e.g., the learning rate in Phase 1 was 10%). −3 At that time, the learning rate for the second stage was set at 10%. −4 The entire network is jointly trained end-to-end. The total loss function includes the loss of the high-recall detection head, the loss of the high-precision detection head, and the final detection loss after fusion, with the weight ratio of the three losses set to 1:1:2. During this stage, the adaptive weight module learns to dynamically adjust based on image features. α The optimal mapping relationship. Third stage: Evaluate the model's performance on the validation set under different image complexities. Collect typical scene data (simple, medium, complex) for a small number of iterative training iterations, or perform targeted fine-tuning on manually selected boundary samples to further optimize the fully connected layer parameters of the adaptive weight module.

[0032] The trained PyTorch model was exported to ONNX format. Using the RKNN-Toolkit2 toolchain, the ONNX model was converted to RKNN format, with INT8 quantization enabled during the conversion process to quantize the FP32-precision model parameters and activation values ​​to INT8 precision. The converted RKNN model was deployed on the NPU of the RK3588 onboard computer, utilizing its 6 TOPS computing power for efficient inference.

[0033] During flight, the drone uses its onboard camera to capture video streams in real time, obtaining the current frame image frame by frame.

[0034] For each frame of the image, calculate its image complexity score. The specific calculation method is as follows: First, extract three metrics, including: Edge density The Canny operator is used to extract image edges, and the proportion of edge pixels to the total number of pixels in the image is calculated. Texture Entropy Texture entropy is calculated using the gray-level co-occurrence matrix (GLCM), which reflects the randomness and complexity of image texture. Number of targets detected in the previous frame The number of detection boxes in the final output of the previous frame is normalized to the [0,1] interval; Then, the image complexity is obtained by weighted summation according to the following formula. : ; in, The weighting coefficients for each indicator, and the image complexity. This will be used to dynamically adjust the fusion weights of the subsequent adaptive weighting module.

[0035] Based on image complexity The intersection-over-union (IoU) threshold and confidence threshold in the non-maximum suppression (NMS) operation are dynamically adjusted. The adjustment formula is as follows: ; ; in, The adjusted IoU threshold, Based on the IoU threshold, This is the adjustment factor for the IoU threshold. The adjusted confidence threshold. Based on the confidence threshold, This is the adjustment coefficient for the confidence threshold. This strategy allows for a stricter threshold to improve precision and suppress false positives when the image complexity is high, and a more lenient threshold to improve recall and avoid false negatives when the complexity is low.

[0036] Base threshold and The coefficients can be set based on experience and adjusted. and All are positive numbers. Therefore, based on the real-time image complexity, the cross-union ratio (CUP) threshold and confidence threshold in the non-maximum suppression operation are dynamically adjusted, and both the CUP threshold and confidence threshold are positively correlated with the real-time image complexity. When the real-time image complexity is greater than or equal to the preset complexity threshold, the intersection-union ratio threshold is adjusted to the first intersection-union ratio threshold, and the confidence threshold is adjusted to the first confidence threshold. When the real-time image complexity is less than the preset complexity threshold, the intersection-over-union (IoU) threshold is adjusted to the second IoU threshold, and the confidence threshold is adjusted to the second confidence threshold. Specifically, the first cross-union ratio (CUP) threshold is greater than the second CUP threshold, and the first confidence threshold is greater than the second confidence threshold. Increasing the confidence threshold reduces false positives; decreasing the confidence threshold reduces false negatives. The current frame image is input into the deployed RKNN model for inference. A shared backbone network extracts multi-scale feature maps, which are then fed into a high-recall detection head and a high-precision detection head, respectively. The high-recall detection head outputs a large number of candidate boxes (each containing location coordinates and a confidence score), while the high-precision detection head outputs candidate boxes refined by fine regression. The average confidence score from both detection heads is calculated simultaneously. and .

[0037] Image complexity score Average confidence level of high recall Average confidence level of high-precision head Constructing eigenvectors The input is the adaptive weight fusion module. This module is a small neural network containing two fully connected layers, and its calculation process is as follows: ; in, For adaptive weights, , , This is the weight matrix of the fully connected layer. , For bias vectors, To modify the activation function of the linear unit, This is the Sigmoid function.

[0038] Then, the output box sets of the two detectors are weighted and merged: Multiply the confidence score of each detection frame output by the high recall detector by ; Multiply the confidence level of each detection frame output by the high-precision detection head by ; Merging the two sets yields the intermediate detection box set: ; in, This is the final set of detection boxes. This is the first set of detection boxes; This is the set of the second detection boxes.

[0039] For the merged set of detection boxes, a dynamically adjusted IoU threshold is used. Adjusted confidence threshold Perform nonmaximum suppression: first filter out values ​​with a confidence level lower than 100%. The detection bounding boxes are then sorted in descending order of confidence, and boxes with an IoU greater than 1 with higher confidence bounding boxes are removed one by one. The repeated bounding boxes. The final set of remaining detection boxes is the final target detection result for the current frame, which is output to the UAV ground station or onboard storage device.

[0040] This invention also includes a method for target tracking based on unmanned aerial vehicles: Specifically: SA1: Obtain the target detection results mentioned above; SA2: Using the center coordinates of the detection boxes in the detection box set, perform target ID matching with the center coordinates of the detection boxes in the current frame image. If the match is successful, a tracking box with an ID is obtained. SA3: Performs state prediction, data association, and trajectory update on the tracking box, and outputs the results to the visualization module.

[0041] Specifically, state prediction includes: Position status: The current coordinates of the target in the image, usually the coordinates of the center point of the detection box; Size status: Width and height of the target detection box; Motion state: The target's velocity in the horizontal and vertical directions; Predicted bounding box: Based on the above states, a predicted bounding box is finally output.

[0042] Data associations include: Association of detection data in the current frame image: Match the predicted bounding boxes output by the state prediction module with the detection bounding boxes (including the center coordinates, confidence score, target category, etc.) of the current frame image output by the YOLOv8 model inference.

[0043] Association with historical trajectory data: Combine the appearance features (such as appearance feature vectors) and motion trajectory of the target in historical frame images to determine whether the current detection box belongs to an existing target ID.

[0044] By calculating the spatial and appearance similarity between each predicted bounding box and the detected bounding box, the Hungarian algorithm is used to find the optimal matching scheme. Once a match is successful, the state of the target ID can be updated and the trajectory can be continued.

[0045] Using the above embodiments, the test set was validated on the RK3588 platform. The test results show that compared with the baseline model using a fixed threshold and a single detection head, this embodiment reduces the false detection rate by about 25% in complex scenes (high texture, high target density), reduces the false detection rate by about 18% in simple scenes, improves the overall mAP by about 5.2 percentage points, and achieves an inference speed of 32 frames / second, meeting the requirements for real-time detection.

[0046] like Figure 3 , Figure 4 The images shown are target recognition result frames obtained by processing actual UAV aerial video using the method of this invention. The specific processing flow is as follows: During the flight of the UAV, the onboard camera collects video streams in real time and pushes the video streams to the local server; the inference module deployed on the RK3588 onboard computer pulls the video streams from the local server in real time, uses the dual-detection head deep learning network model trained by this invention to perform target detection and recognition, and outputs result images with detection boxes and category labels.

[0047] Those skilled in the art should understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For example, the airborne embedded device is not limited to RK3588, but can also be other embedded platforms with an NPU (such as Jetson Orin, BM1684, etc.); the model format conversion is not limited to ONNX to RKNN, but can also use corresponding intermediate representations (such as TensorRT, TFLite, etc.) depending on the target platform. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A target detection method based on unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1: Obtain the dual-detection-head deep learning network model. The dual-detection-head deep learning network model includes: a high-recall detection head for initial screening of all potential detection targets, a high-precision detection head for fine classification of detection targets, and an adaptive weight fusion module connected to the output of the high-recall detection head and the high-precision detection head. S2: Obtain the current frame image captured by the drone and calculate the real-time image complexity of the current frame image; S3: Input the current frame image into the dual-detection-head deep learning network model. The high-recall detection head and the high-precision detection head output the detection results respectively. The adaptive weight fusion module dynamically adjusts the fusion weight of the output results of the two detection heads according to the real-time image complexity to generate the final target detection result. The target detection result includes the detection box set, confidence score and target category.

2. The target detection method based on unmanned aerial vehicles according to claim 1, characterized in that, In S2, the real-time image complexity of the current frame image is calculated, specifically including: obtaining the edge density, texture entropy, and number of targets detected in the previous frame image of the current frame image; and performing a weighted summation of the edge density, texture entropy, and the normalized number of targets detected in the previous frame image to obtain the real-time image complexity that characterizes the background complexity of the current frame image.

3. The target detection method based on unmanned aerial vehicles according to claim 1, characterized in that, In S3, the adaptive weight fusion module dynamically adjusts the fusion weights of the outputs from the two detection heads based on the real-time image complexity, specifically including: The real-time image complexity, the average confidence score of the high-recall detection head, and the average confidence score of the high-precision detection head are used as input features. Through the fully connected layer and non-linear activation function in the adaptive weight fusion module, an adaptive fusion weight value between 0 and 1 is mapped to the output. The output results of the high-recall detection head and the high-precision detection head are weighted and summed using the adaptive fusion weight value to generate the final set of detection boxes, confidence scores, and target categories.

4. The target detection method based on unmanned aerial vehicles according to claim 1, characterized in that, The model construction process of S1 also includes a nonmaximum suppression step: Based on the real-time image complexity, the cross-union ratio (CUP) threshold and confidence threshold in the non-maximum suppression step are dynamically adjusted. Both the CUP threshold and the confidence threshold are positively correlated with the real-time image complexity. When the real-time image complexity is greater than or equal to the preset complexity threshold, the intersection-union ratio threshold is adjusted to the first intersection-union ratio threshold, and the confidence threshold is adjusted to the first confidence threshold. When the real-time image complexity is less than the preset complexity threshold, the intersection-over-union (IoU) threshold is adjusted to the second IoU threshold, and the confidence threshold is adjusted to the second confidence threshold. Among them, the first crossover ratio (CUP) threshold is greater than the second CUP threshold, and the first confidence threshold is greater than the second confidence threshold.

5. The target detection method based on unmanned aerial vehicles according to claim 1, characterized in that, S1 also includes phased training of the dual-detection-head deep learning network model, specifically: In the first stage, the parameters of the adaptive weight fusion module are frozen, and the high recall detection head and the high precision detection head are trained independently using the complete dataset, so that the two detection heads each have their own emphasis characteristics. In the second stage, the parameters of the adaptive weight fusion module are unfrozen, and end-to-end joint training is performed on all parameters of the high recall detector, the high precision detector, and the adaptive weight fusion module with a second learning rate lower than that in the first stage. The network parameters are then optimized using the total loss function.

6. A target tracking method based on unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: SA1: Obtain the target detection result as described in any one of claims 1-5; SA2: Using the center coordinates of the detection boxes in the detection box set, perform target ID matching with the center coordinates of the detection boxes in the current frame image. If the match is successful, a tracking box with an ID is obtained. SA3: Performs state prediction, data association, and trajectory update on the tracking box, and outputs the results to the visualization module.

7. The target tracking method based on an unmanned aerial vehicle (UAV) according to claim 6, characterized in that, In SA3, tracking frames perform data association: Specifically, it includes: Association with the detection data of the current frame image: Match the prediction bounding box output by the state prediction module with the detection bounding box of the current frame image; Association with historical trajectory data: Combine the appearance features and motion trajectory of the target in historical frame images to determine whether the detection box in the current frame image belongs to an existing target ID; By calculating the spatial and appearance similarity between each predicted bounding box and the detected bounding box, the optimal matching scheme is found. Once a match is successful, the state of the target ID can be updated and the trajectory can be continued.

8. A target detection system based on unmanned aerial vehicles (UAVs), characterized in that, include: Model acquisition module: used to acquire the dual-detection head deep learning network model, which includes a high-recall detector head and a high-precision detector head, as well as an adaptive weight fusion module connected to the output of the high-recall detector head and the high-precision detector head; Image complexity calculation module: used to acquire the current frame image of the drone aerial photography and calculate the real-time image complexity of the current frame image; The target detection module is used to input the current frame image into the dual-detection-head deep learning network model. The high-recall detection head and the high-precision detection head output the detection results respectively. The adaptive weight fusion module dynamically adjusts the fusion weight of the output results of the two detection heads according to the real-time image complexity to generate the final target detection result. The target detection result includes the detection box set, confidence score and target category.

9. A target tracking system based on unmanned aerial vehicles (UAVs), characterized in that, include: Target detection result acquisition module: used to acquire the target detection result as described in any one of claims 1-5; Target tracking module: Used to match the center coordinates of the detection boxes in the target detection results with the center coordinates of the detection boxes in the current frame image to obtain the target ID. If the match is successful, a tracking box with an ID is obtained. Visualization module: This module takes the tracking box as input to the target tracking module for state prediction, data association, and trajectory update, and outputs the results to the visualization module.

Images