Target detection system, method and equipment based on airborne scene of unmanned aerial vehicle and medium

Abstract

The invention discloses a target detection system, method, equipment and medium based on an airborne scene of an unmanned aerial vehicle, relates to the technical field of target monitoring, and solves the technical problems of low computing power and slow response caused by mismatching of a target detection model and hardware deployment. The video stream delay is reduced to be within 50ms, and the pan-tilt response delay is reduced to be within 30ms, so that the problems of stream cutoff and slow response are solved; layered weight reduction is carried out on a target detection model on the software level, the precision is kept while the computing power is greatly saved, cache blockage is avoided by adopting multi-thread frame skip detection, control logic is optimized, the target detection rate reaches 19.2 FPS by combining TensorRT acceleration, high precision is kept, airborne computing power constraints are adapted, and the real-time performance and reliability of airborne target monitoring of the unmanned aerial vehicle are improved.

Description

Technical Field

[0001] This application relates to the field of target monitoring technology, and in particular to a target detection system, method, device and medium based on UAV airborne scenarios. Background Technology

[0002] Current UAV-based target detection technologies primarily revolve around algorithm selection and deployment optimization. Algorithm-wise, standard target detection models such as YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector) are often directly adopted without customization for onboard computing power. Deployment-wise, embedded onboard computers like the Jetson TX2 are used, with processes mainly involving pre-trained model porting and basic environment configuration, lacking deep adaptation to the hardware architecture. In practical applications, optimization strategies are simplistic: software only adjusts training parameters, hardware relies on default settings, and redundant processes are retained, impacting computing power utilization.

[0003] Existing technologies have several shortcomings: First, the models are mismatched with airborne computing power. Traditional object detection algorithms such as Faster R-CNN and SSD, when deployed on a Jetson TX2, have an inference speed of less than 6 FPS, resulting in insufficient real-time performance. Second, the models are poorly adapted. Models trained on general datasets have low detection accuracy and weak robustness for small targets and targets with complex backgrounds in airborne scenes. Third, the deployment optimization is simplistic. Hardware has not been specifically optimized, and models have not been lightweighted, resulting in low utilization of computing power. Fourth, the system latency is high. Image transmission is prone to interruption, the gimbal control link is long, and video reading cache is blocked, affecting the coordination of monitoring and control.

[0004] Chinese patent application CN115063704A discloses a method for classifying targets monitored by drones. It constructs a convolutional neural network with an encoder-decoder structure and uses the MobilenetV2 network as the backbone network, which improves the model's running efficiency to a certain extent. However, this method does not consider actual deployment and only uses public datasets for algorithm verification without conducting actual deployment experiments.

[0005] Chinese patent application CN108022255A discloses an automatic tracking method for unmanned aerial vehicles (UAVs). This method mainly includes extracting a reference position and calculating the target distance, which can enable UAVs to follow targets even without a mechanical gimbal. However, this method does not mention the actual deployment effect and has not been tested in practice.

[0006] Therefore, how to integrate the target detection model with hardware deployment is a problem that this application aims to solve. Summary of the Invention

[0007] This application provides a target detection system, method, device, and medium based on UAV airborne scenarios. Its technical purpose is to combine the target detection model with hardware deployment, thereby achieving model lightweighting while improving target detection accuracy, and enabling deep matching between the target detection model and hardware architecture to reduce system latency.

[0008] The above-mentioned technical objective of this application is achieved through the following technical solution: A target detection system based on an unmanned aerial vehicle (UAV) airborne scenario includes a control system and an UAV image acquisition system. The UAV image acquisition system includes a flight control module and a gimbal module. The gimbal module is equipped with a camera, which is connected to the control system via a converter. The control system is connected to the flight control module, and a protection resistor is connected in series between the control system and the gimbal module. The control system is equipped with a target detection model, which detects images captured by the camera; the control system uses a feedback control algorithm to control the flight control module and the gimbal module. The control system receives the video stream captured by the camera, caches the video stream through a cache update thread, and caches and reads the video stream through an image processing thread before sending it to the target detection model for target detection; the cache update thread and the image processing thread are independent of each other; The feedback control algorithm is expressed as follows: ; in, express Expected output at any given time express The difference between the real-time visualization results and the expected output. express The difference between the real-time visualization results and the expected output; Indicates a time interval; Represents a saturation function; Indicates the PID proportional coefficient; Indicates the integral coefficient of the PID controller; Represents the PID differential coefficients; express The deviation between expected and actual values ​​at any given moment; This indicates the horizontal deviation, which is the deviation of the UAV's azimuth angle; Indicates the camera's focal length. Indicates the image width. Indicates the distance between the drone and the target. Indicates the image height. This indicates the position coordinates of the center of the target box within the image coordinate system; This indicates the vertical deviation, which is the pitch angle deviation of the gimbal module; Indicates longitudinal distance deviation. Indicates lateral distance deviation. and Used to describe the deviation in the flight speed of a drone; This represents the distance-velocity mapping ratio. Indicates the expected area of ​​the target box. Indicates the area of ​​the target bounding box. This indicates the direction angle of the drone's movement.

[0009] Preferably, in the feedback control algorithm, multi-target tracking adaptation is performed based on the target score. When the score of a new target exceeds a threshold of the current target score, target switching is initiated. The target score is represented as: ; in, This represents the target score; the higher the target score, the higher the target priority. Indicates the confidence weight. This represents the confidence level of object detection. Indicates scale weight, This indicates the time weight of the target being continuously monitored. This indicates the number of frames in which the target is detected consecutively. .

[0010] Preferably, the thread priority of the target detection system is set as follows: visual deviation calculation thread > UAV PID control thread > flight control module IMU data acquisition thread; The visual deviation calculation thread is used to process the video stream acquired by the camera, and the visual deviation calculation thread and the flight control module IMU data acquisition thread serve as inputs to the UAV PID control thread.

[0011] Preferably, after the target detection model obtains the target coordinates, the control module calculates the relative position between the target and the UAV based on the target coordinates, and calculates the UAV flight speed, UAV azimuth angle, and gimbal module pitch angle based on the relative position. The feedback control algorithm then controls the flight control module and the gimbal module based on the UAV flight speed, UAV azimuth angle, and gimbal module pitch angle.

[0012] Preferably, the step of caching the video stream via a cache update thread includes: The latest frame stored in the OpenCV library cache queue is continuously read through the RTSCapture thread; Check for cache blocking. If it is, replace the first frame in the cache queue. Otherwise, take the first frame in the cache queue, which is the latest frame, and continue reading. The process of caching and reading the video stream through an image processing thread includes: The received video stream is buffered using the OpenCV library; The cached video stream is read sequentially and then fed into the object detection model in turn.

[0013] Preferably, the target detection model is a YOLOv5s fine-tuning model, and the construction process of the YOLOv5s fine-tuning model includes: In the backbone network of the original YOLOv5s model, convolutional layers with absolute weight values ​​less than a preset threshold are filtered to obtain redundant channels. The weights of the original YOLOv5s model were pre-trained using the COCO dataset, and then sparsified training was performed on the VisDrone2019 dataset to make the weights of redundant channels approach 0, thus obtaining the first basic YOLOv5s model. The first basic YOLOv5s model backbone network with channels whose absolute weight values ​​are less than a preset threshold is removed by using a batch pruning tool to obtain the pruned second basic YOLOv5s model. By embedding an ECA attention mechanism module at the output of each CSP module of the second basic YOLOv5s model, a third basic YOLOv5s model is obtained. The accuracy of the third basic YOLOv5s model is adjusted to obtain the YOLOv5s fine-tuning model.

[0014] Preferably, the training process of the YOLOv5s fine-tuning model includes: Preprocessing: The YOLOv5s fine-tuned model is subjected to stratified freezing, including: the bottom layer CSP module is fully frozen, the middle layer CSP module has 70% of its weights frozen, and the high layer CSP module and its ECA attention mechanism module are not frozen; the YOLOv7-tiny model trained on the UAVDT dataset is selected as the teacher model. Training process: Distillation training is performed on the high-level layers and the weights of the high-level layers are updated; after the high-level distillation training is completed, the middle layers are unfrozen and fine-tuned to obtain the trained YOLOv5s fine-tuned model. Precision quantization and acceleration process: The trained YOLOv5s fine-tuning model is subjected to precision quantization and acceleration to obtain the final YOLOv5s fine-tuning model.

[0015] A target detection method based on UAV-borne scenarios, implemented through a UAV-borne scenario target detection system, includes: The UAV image acquisition system acquires target images through a camera, and the camera transmits the target images to the control system through a converter; The control system detects the target image using a deployed target detection model, and controls the flight control module and the gimbal module based on the image detection results using a feedback control algorithm.

[0016] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of a target detection method based on an unmanned aerial vehicle (UAV) airborne scene.

[0017] A computer storage medium storing a computer program, which, when executed by a processor, implements the steps of a target detection method based on an unmanned aerial vehicle (UAV) airborne scene.

[0018] The above technical solution can achieve at least some of the following technical effects: This application optimizes the image transmission link at the hardware level, reducing video stream latency to within 50ms and gimbal response latency to within 30ms, resolving issues of stream interruption and slow response. At the software level, the target detection model is layered and lightweighted, significantly saving computing power while maintaining accuracy. Multi-threaded frame skipping detection is used to avoid cache blocking, and control logic is optimized. Combined with TensorRT acceleration, the target detection rate reaches 19.2 FPS while maintaining high accuracy, adapting to airborne computing power constraints and improving the real-time performance and reliability of UAV-based target monitoring. Attached Figure Description

[0019] Figure 1 This is a framework diagram of a target detection system based on an unmanned aerial vehicle (UAV) airborne scenario, as described in this application embodiment. Figure 2 This is a schematic diagram showing the comparison before and after image transmission link optimization in the embodiments of this application; Figure 3 This is a schematic diagram showing the comparison before and after line optimization of the gimbal control link in the embodiments of this application; Figure 4 This is a flowchart illustrating how multi-threaded cache optimization addresses cache embolism in this application embodiment. Detailed Implementation

[0020] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0021] like Figure 1 , Figure 2 and Figure 3 As shown, the target detection system based on UAV airborne scene described in this application is characterized in that the target detection system includes a control system and a UAV image acquisition system. The UAV image acquisition system includes a flight control module and a gimbal module. The gimbal module is equipped with a camera. The camera is connected to the control system through a converter. The control system is connected to the flight control module. A protection resistor is connected in series between the control system and the gimbal module.

[0022] The control system is equipped with a target detection model, which is used to detect images captured by the camera. The control system controls the flight control module and the gimbal module through a feedback control algorithm.

[0023] like Figure 1 As shown in the embodiment of this application, the UAV is also equipped with a GPS module, a data transmission module, and an optical flow module.

[0024] The feedback control algorithm is expressed as follows: ; in, express The expected output at any given moment is used to control the movement of the drone in three directions; express The difference between the real-time visualization results and the expected output. express The difference between the real-time visualization results and the expected output; This represents a time interval, taken as 0.02s; Represents a saturation function; This represents the PID proportional coefficient. Represents the integral coefficient of the PID controller. Represents the PID derivative coefficients. This represents the deviation between the expected and actual values. This indicates the horizontal deviation, which is the deviation of the UAV's azimuth angle; Indicates the camera's focal length. Indicates the image width. Indicates the distance between the drone and the target. Indicates the image height. This indicates the position coordinates of the center of the target bounding box within the image coordinate system. This indicates the vertical deviation, which is the pitch angle deviation of the gimbal module; Indicates longitudinal distance deviation. Indicates lateral distance deviation. and That is, the deviation in the drone's flight speed; This represents the distance-velocity mapping ratio. Indicates the expected area of ​​the target box. Indicates the area of ​​the target bounding box. This indicates the direction angle of the drone's movement.

[0025] Preferably, in the feedback control algorithm, multi-target tracking adaptation is performed based on the target score. When the score of a new target exceeds a threshold of the current target score, target switching is initiated. The target score is represented as: ; in, This represents the target score; the higher the target score, the higher the target priority. Indicates the confidence weight. This represents the confidence level of object detection. Indicates scale weight, This indicates the time weight of the target being continuously monitored. This indicates the number of frames in which the target is detected consecutively. .

[0026] In this embodiment of the application, when the score of the new target exceeds 20% of the score of the current target, the target switching is initiated.

[0027] In the embodiments of this application, the control objectives of the feedback control algorithm include: (1) Horizontal direction (UAV heading angle): Deviation between the X coordinate of the target box center and the X coordinate of the image center ≤±3 pixels; (2) Vertical direction (pan-tilt angle): Deviation between the Y coordinate of the target box center and the Y coordinate of the image center ≤±2 pixels; (3) Distance control (UAV speed): Deviation between target frame area and desired area ( and ≤±10% (the expected area corresponds to the optimal distance of 10 meters); (4) Basic PID (proportional) The initial values ​​of the (integral I-differential D) parameters (based on measured calibration) are: UAV azimuth angle PID: =0.9, =0.01, =0.2; Gimbal tilt angle PID: =1.2, =0.08, =0 (PI control only); Drone speed PID: =0.6, =0.02, =0.06.

[0028] (5) Control quantity hard limiting includes: Drone heading angle adjustment: maximum angular velocity ±60° / s, single adjustment range ±30°; Gimbal pitch angle: Adjustment range -30° (down) to +10° (up), maximum angular velocity ±45° / s; Unmanned aerial vehicle (UAV) flight speed: Maximum speed along the Xb / Yb axes ±2m / s, acceleration ≤0.5m / s². 2 (Avoid sudden acceleration and deceleration).

[0029] Preferably, the camera is connected to the converter via an HDMI interface, the converter being an HDMI to CSI-2 adapter board, and the control system is connected to the converter via a CSI interface; the control system is a Jetson TX2NX. Figure 2 As shown, the camera's HDMI interface is connected to the Jetson TX2 NX's CSI interface via an HDMI to CSI module.

[0030] Preferably, after the target detection model obtains the target coordinates, the control module calculates the relative position between the target and the UAV based on the target coordinates, and calculates the UAV flight speed, UAV azimuth angle, and gimbal module pitch angle based on the relative position. The feedback control algorithm then controls the flight control module and the gimbal module based on the UAV flight speed, UAV azimuth angle, and gimbal module pitch angle.

[0031] like Figure 4 As shown, after receiving the video stream acquired by the camera, the control system caches the video stream through a cache update thread and caches and reads the video stream through an image processing thread, then sends it to the target detection model for target detection. The cache update thread and the image processing thread are independent of each other.

[0032] Preferably, the step of caching the video stream via a cache update thread includes: The latest frame stored in the OpenCV library cache queue is continuously read through the RTSCapture thread; The system checks for cache blocking. If a cache block exists, the first frame in the cache queue is replaced. Otherwise, the system continuously reads the first frame in the cache queue, which is the latest frame.

[0033] The process of caching and reading the video stream through an image processing thread includes: The received video stream is buffered using the OpenCV library; The cached video stream is read sequentially and then fed into the object detection model in turn.

[0034] In this application, the thread priority is set as follows: visual deviation calculation thread > UAV PID control thread > flight control module IMU data acquisition thread, ensuring that the core control logic responds first. The visual deviation calculation thread is used to process the video stream acquired by the camera. Figure 4 The cache update thread and image processing thread are both part of the visual deviation calculation thread; the UAV PID control thread is part of the UAV control thread; the flight control module IMU data acquisition thread refers to the real-time reading of data from the flight control module's IMU sensors, such as heading angle and flight speed. The visual deviation calculation thread and IMU data acquisition thread serve as inputs to the PID control thread.

[0035] Preferably, the target detection model is a YOLOv5s fine-tuning model, which includes an input layer, a backbone network, a neck network, and an output layer. The backbone network includes high-level, mid-level, and low-level layers. The dataset used by the YOLOv5s fine-tuning model is the VisDrone dataset, and the number of categories output by the output layer is 10.

[0036] Preferably, the construction process of the YOLOv5s fine-tuning model includes: (1) Filter the convolutional layers in the backbone network of the original YOLOv5s model whose absolute weight values ​​are less than a preset threshold to obtain redundant channels.

[0037] Preferably, the preset threshold for the absolute value of the weights when filtering redundant channels is 0.01.

[0038] (2) The weights of the original YOLOv5s model are pre-trained using the COCO dataset, and then sparsified training is performed on the VisDrone2019 dataset to make the weights of redundant channels approach 0, thus obtaining the first basic YOLOv5s model.

[0039] Preferably, during sparsity training, an L1 regularization term (with a weight coefficient of 0.0001) is added to the loss function to force the weights of redundant channels to approach 0.

[0040] Preferably, the training parameters are: epochs=50, batch-size=16, initial learning rate 0.005, cosine annealing learning rate decay, freezing the bottom 3 CSP modules, and training only the middle, high and output layers.

[0041] (3) The channels with absolute weight values ​​less than a preset threshold in the backbone network of the first basic YOLOv5s model are removed by using a batch pruning tool to obtain the pruned second basic YOLOv5s model.

[0042] Preferably, the batch pruning tool is torch.nn.utils.prune. For the first base YOLOv5s model trained with sparsity, channels with absolute weight values ​​less than a threshold in each candidate layer are removed at a pruning rate of 30%, generating the pruned second base YOLOv5s model.

[0043] (4) An ECA (Efficient Channel Attention) attention mechanism module is embedded at the output of each CSP module of the second basic YOLOv5s model. Channel weight calibration is achieved only through 1D convolution and global average pooling without adding extra dimensional transformation, thus obtaining the third basic YOLOv5s model.

[0044] (5) Adjust the accuracy of the third basic YOLOv5s model to obtain the YOLOv5s fine-tuning model.

[0045] Preferably, the pruning model with embedded ECA is fine-tuned with epochs=30, learning rate=0.001, batch-size=16, the weights of the bottom layer of the backbone network are fixed, and the middle layer, high layer and ECA attention mechanism modules are unfrozen.

[0046] Preferably, the training process of the YOLOv5s fine-tuning model includes: (1) Preprocessing: The YOLOv5s fine-tuned model is subjected to layered freezing, including: the bottom CSP module (layers 1-3) is fully frozen (only the general feature extraction capability is retained), the middle CSP module (layer 4) has 70% of its weights frozen, and the high-level CSP module and its ECA attention mechanism module are not frozen; the YOLOv7-tiny model trained on the UAVDT dataset is selected as the teacher model, whose parameter scale is twice that of the student model (YOLOv5s fine-tuned model), and the small target detection accuracy is higher; (2) Training process: Distillation training is performed on the high layer and the weights of the high layer are updated; after the high layer distillation training is completed, the middle layer is unfrozen and fine-tuned to obtain the trained YOLOv5s fine-tuned model.

[0047] Preferably, the distillation loss function is: total loss = detection loss (weight 0.7) + distillation loss (weight 0.3), with KL divergence used for distillation loss, and distillation temperature T=3 (balancing the weights of hard and soft labels).

[0048] Training parameters: epochs=25, batch-size=16, learning rate 0.0008, only update the weights of the unfrozen layer, and use the high-level semantic features of the teacher model to guide the student model to learn.

[0049] (3) Precision quantization and acceleration process: The precision of the trained YOLOv5s fine-tuning model is quantized and accelerated to obtain the final YOLOv5s fine-tuning model.

[0050] Preferably, INT8 quantization is used for the shallow layers (basic feature extraction) and non-critical convolutional layers of the backbone network to significantly save computing power; FP16 quantization is used for the fusion layer, ECA attention layer and output layer of the neck network to preserve accuracy.

[0051] Preferably, the PyTorch model (.pt format) of the YOLOv5s fine-tuning model is converted into a .wts weight file using a tensorrtx script, the quantized weight parameters of each layer are extracted, and an engine (.engine) is built based on TensorRT.

[0052] The target detection method based on UAV airborne scenarios described in this application includes: 100: Image acquisition and transmission process. The UAV image acquisition system acquires target images through a camera, and the camera transmits the target images to the control system through a converter.

[0053] 101: Target detection and feedback control process: The control system detects the target image through the deployed target detection model, and controls the flight control module and the gimbal module based on the image detection results through a feedback control algorithm.

[0054] The above are exemplary embodiments of this application, and the scope of protection of this application is defined by the claims and their equivalents.

Claims

1. A target detection system based on UAV airborne scenarios, characterized in that, The target detection system includes a control system and a UAV image acquisition system. The UAV image acquisition system includes a flight control module and a gimbal module. The gimbal module is equipped with a camera. The camera is connected to the control system through a converter. The control system is connected to the flight control module. A protective resistor is connected in series between the control system and the gimbal module. The control system is equipped with a target detection model, which detects images captured by the camera; the control system uses a feedback control algorithm to control the flight control module and the gimbal module. The control system receives the video stream captured by the camera, caches the video stream through a cache update thread, and caches and reads the video stream through an image processing thread before sending it to the target detection model for target detection; the cache update thread and the image processing thread are independent of each other; The feedback control algorithm is expressed as follows: ； in, express Expected output at any given time express The difference between the real-time visualization results and the expected output. express The difference between the real-time visualization results and the expected output; Indicates a time interval; Represents a saturation function; Indicates the PID proportional coefficient; Indicates the integral coefficient of the PID controller; Represents the PID differential coefficients; express The deviation between expected and actual values ​​at any given moment; This indicates the horizontal deviation, which is the deviation of the UAV's azimuth angle; Indicates the camera's focal length. Indicates the image width. Indicates the distance between the drone and the target. Indicates the image height. This indicates the position coordinates of the center of the target box within the image coordinate system; This indicates the vertical deviation, which is the pitch angle deviation of the gimbal module; Indicates longitudinal distance deviation. Indicates lateral distance deviation. and Used to describe the deviation in the flight speed of a drone; This represents the distance-velocity mapping ratio. Indicates the expected area of ​​the target box. Indicates the area of ​​the target bounding box. This indicates the direction angle of the drone's movement.

2. The target detection system as described in claim 1, characterized in that, In the feedback control algorithm, multi-target tracking adaptation is performed based on the target score. When the score of a new target exceeds the threshold of the current target score, target switching is initiated. The target score is represented as: ； in, This represents the target score; the higher the target score, the higher the target priority. Indicates the confidence weight. This represents the confidence level of object detection. Indicates scale weight, This indicates the time weight of the target being continuously monitored. This indicates the number of frames in which the target is detected consecutively. .

3. The target detection system as described in claim 2, characterized in that, The thread priority of the target detection system is set as follows: visual deviation calculation thread > UAV PID control thread > flight control module IMU data acquisition thread; The visual deviation calculation thread is used to process the video stream acquired by the camera, and the visual deviation calculation thread and the flight control module IMU data acquisition thread serve as inputs to the UAV PID control thread.

4. The target detection system as described in claim 3, characterized in that, After the target detection model obtains the target coordinates, the control module calculates the relative position between the target and the UAV based on the target coordinates, and calculates the UAV flight speed, UAV azimuth angle, and gimbal module pitch angle based on the relative position. The feedback control algorithm then controls the flight control module and the gimbal module based on the UAV flight speed, UAV azimuth angle, and gimbal module pitch angle.

5. The target detection system as described in claim 4, characterized in that, The method of caching the video stream through a cache update thread includes: The latest frame stored in the OpenCV library cache queue is continuously read through the RTSCapture thread; Check for cache blocking. If it is, replace the first frame in the cache queue. Otherwise, take the first frame in the cache queue, which is the latest frame, and continue reading. The process of caching and reading the video stream through an image processing thread includes: The received video stream is buffered using the OpenCV library; The cached video stream is read sequentially and then fed into the object detection model in turn.

6. The target detection system as described in claim 5, characterized in that, The target detection model is a YOLOv5s fine-tuning model, and the construction process of the YOLOv5s fine-tuning model includes: In the backbone network of the original YOLOv5s model, convolutional layers with absolute weight values ​​less than a preset threshold are filtered to obtain redundant channels. The weights of the original YOLOv5s model were pre-trained using the COCO dataset, and then sparsified training was performed on the VisDrone2019 dataset to make the weights of redundant channels approach 0, thus obtaining the first basic YOLOv5s model. The first basic YOLOv5s model backbone network with channels whose absolute weight values ​​are less than a preset threshold is removed by using a batch pruning tool to obtain the pruned second basic YOLOv5s model. By embedding an ECA attention mechanism module at the output of each CSP module of the second basic YOLOv5s model, a third basic YOLOv5s model is obtained. The accuracy of the third basic YOLOv5s model is adjusted to obtain the YOLOv5s fine-tuning model.

7. The target detection system as described in claim 6, characterized in that, The training process of the YOLOv5s fine-tuning model includes: Preprocessing: The YOLOv5s fine-tuned model is subjected to stratified freezing, including: the bottom layer CSP module is fully frozen, the middle layer CSP module has 70% of its weights frozen, and the high layer CSP module and its ECA attention mechanism module are not frozen; the YOLOv7-tiny model trained on the UAVDT dataset is selected as the teacher model. Training process: Distillation training is performed on the high-level layers and the weights of the high-level layers are updated; after the high-level distillation training is completed, the middle layers are unfrozen and fine-tuned to obtain the trained YOLOv5s fine-tuned model. Precision quantization and acceleration process: The trained YOLOv5s fine-tuning model is subjected to precision quantization and acceleration to obtain the final YOLOv5s fine-tuning model.

8. A target detection method based on an unmanned aerial vehicle (UAV) airborne scenario, wherein the target detection method is implemented by the target detection system described in any one of claims 1 to 7, characterized in that, include: The UAV image acquisition system acquires target images through a camera, and the camera transmits the target images to the control system through a converter; The control system detects the target image using a deployed target detection model, and controls the flight control module and the gimbal module based on the image detection results using a feedback control algorithm.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the target detection method based on UAV airborne scene as described in claim 8.

10. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the target detection method based on the UAV airborne scene as described in any one of claims 8.

Images