Lightweight intelligent control method for low-altitude operation scene

Abstract

The invention provides a lightweight intelligent control method for a low-altitude operation scene, and the method comprises the steps: focusing a low-altitude key target through an attention mechanism based on a lightweight deep learning model; data enhancement is carried out through a dynamic data enhancement module so as to expand the diversity of training samples; calling a space-time calibration model to process the multi-source data, and controlling a time synchronization error and a space positioning deviation within a preset range; based on a dynamic obstacle trajectory prediction module, historical motion data and environmental constraints are fused, and a lightweight LSTM network is adopted to predict a trajectory of a dynamic target in a future time period; and calling a task priority-computing power load dynamic matching mechanism to realize task allocation, and adjusting computing power occupation of non-critical tasks according to a designed energy consumption perception scheduling strategy. According to the method, the autonomous perception, decision making and resource management capabilities of the low-altitude operation unmanned aerial vehicle in a complex environment are comprehensively improved, so that the low-altitude operation unmanned aerial vehicle better meets the requirements for safety, efficiency and continuity in actual operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-altitude operation control of unmanned aerial vehicles, and specifically provides a lightweight intelligent control method for a low-altitude operation scene. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle technology, low-altitude intelligent operations such as power inspection, emergency rescue, urban low-altitude logistics, and agricultural plant protection have gradually become important application directions in the field of automation. Such operation scenarios put high requirements on the autonomy of unmanned aerial vehicles. However, in the perception link, the existing technology cannot balance the accuracy and edge computing power: complex models can improve the identification accuracy of small targets and occluded targets, but due to the high demand for computing power, they cannot run in real time on low-power devices of small unmanned aerial vehicles; lightweight models adapt to computing power, but greatly reduce the identification accuracy, and have weak anti-interference ability to environmental disturbances such as rain, fog, and sudden changes in light, often resulting in missed detection and false detection, and cannot provide reliable environmental information. At the same time, the spatio-temporal alignment effect of multi-sensor data is poor, and is affected by the vibration of the vehicle body and time synchronization errors, so the data is prone to misalignment and contradiction, and the fusion cannot reflect the true environmental state. In the decision-making link, the existing technology cannot adapt to the low-altitude dynamic scene: trajectory prediction mostly relies on simple models, and the prediction error is large for sudden changes in obstacles such as pedestrians crossing and other unmanned aerial vehicles, and it is difficult to reserve avoidance time in advance; in addition, the avoidance decision is mostly based on a fixed distance threshold to trigger the action, without combining environmental constraints such as whether there are high-rise buildings and no-fly zones for flexible adjustment, which is prone to the problems of "whether to avoid" or "false triggering of avoidance", and the decision response speed is slow, making it difficult to keep up with the dynamic changes of the low-altitude environment.

[0003] Therefore, there is an urgent need for a lightweight intelligent control method for a low-altitude operation scene to solve the above problems. SUMMARY

[0004] In order to overcome the above-mentioned defects, the present application is proposed to solve or partially solve the above-mentioned problems.

[0005] The present application provides a lightweight intelligent control method for a low-altitude operation scene, which comprises: focusing on low-altitude key targets through an attention mechanism based on a lightweight deep learning model; and performing data augmentation through a dynamic data augmentation module to expand the diversity of training samples; calling a spatio-temporal calibration model based on vision, laser radar, millimeter wave radar, and IMU to process multi-source data, so as to control the time synchronization error and spatial positioning deviation within a preset range; based on a dynamic obstacle trajectory prediction module, fusing historical motion data and environmental constraints, and using a lightweight LSTM network to predict the trajectory of a dynamic target in a future time period; based on an edge lightweight intelligent computing module, calling a task priority-computing power load dynamic matching mechanism to realize the allocation of tasks, and adjusting the computing power occupation of non-critical tasks according to a designed energy perception scheduling strategy.

[0006] In one of the technical solutions of the lightweight intelligent control method for the low-altitude operation scene described above, the lightweight deep learning model is YOLOv8-tiny, and based on the lightweight deep learning model, the attention mechanism is used to focus on the low-altitude key target, including: inserting an attention module at the Backbone end and the Neck end of YOLOv8-tiny, and in the low-altitude scene characteristics, adding a target difficulty coefficient in training: dynamically increasing the attention weight of small targets, occluded targets, and complex background targets through a loss function.

[0007] In one of the technical solutions of the lightweight intelligent control method for the low-altitude operation scene described above, the dynamic data enhancement module includes: a static enhancement module for performing geometric transformation and color transformation on the training set images; and a dynamic enhancement module for selecting a corresponding enhancement strategy according to low-altitude scene interference, wherein the low-altitude scene interference includes rain, fog, sudden changes in illumination, and dynamic occlusion.

[0008] In one of the technical solutions of the lightweight intelligent control method for the low-altitude operation scene described above, the space-time calibration model includes a time synchronization sub-model, a space calibration sub-model, and a vibration compensation sub-model; a space-time calibration model based on vision, lidar, millimeter wave radar, and IMU is called to process multi-source data, and time synchronization error and spatial positioning deviation are controlled within a preset range, and the control of the time synchronization error and the spatial positioning deviation within the preset range specifically includes: calibrating the consistency of the acquisition time of each sensor through the time synchronization sub-model, mapping the original data to a unified body coordinate system through the space calibration sub-model, and correcting the vibration misplacement through the vibration compensation sub-model, thereby completing the space-time alignment of multi-source data.

[0009] In one of the technical solutions of the lightweight intelligent control method for the low-altitude operation scene described above, the LSTM network has 2 layers of hidden layers, each layer has 64 hidden units, the input features are the position data, velocity data, and acceleration data of the target in the past n seconds, and the output is the position prediction in the future m seconds.

[0010] In one of the technical solutions of the lightweight intelligent control method for the low-altitude operation scene described above, the dynamic obstacle trajectory prediction module introduces an attention mechanism to real-time correct the prediction result in a sudden turning scene, and specifically, the attention mechanism gives higher weights to the historical data before and after the turning time by focusing on the target motion direction change characteristics.

[0011] In one technical solution of the lightweight intelligent control method for the aforementioned low-altitude operation scenario, the task priority-computing load dynamic matching mechanism is as follows: priority quantification and type classification of edge tasks are performed to clarify the scheduling priority and hardware adaptation direction of different tasks; GPU / CPU load is monitored in real time, and tasks are dynamically allocated according to the load scenario based on task priority and type to complete the initial matching of tasks and hardware; after the initial task allocation, the real-time load of GPU, CPU and memory is continuously monitored, and if the load abnormality threshold is triggered, the adjustment mechanism is activated to correct the initial allocation deviation and ensure the balance of computing power supply and demand and the stable operation of the system.

[0012] In one technical solution of the lightweight intelligent control method for the aforementioned low-altitude operation scenario, the edge task includes core tasks and auxiliary tasks; wherein, the core tasks include obstacle avoidance decision-making, target detection, and trajectory planning, and the auxiliary tasks include data storage and status monitoring.

[0013] In one technical solution of the lightweight intelligent control method for the aforementioned low-altitude operation scenario, the adjustment mechanism is as follows: when the GPU utilization rate is detected to trigger a high load threshold and the CPU is in a low load state, a GPU high load mitigation mechanism is triggered, that is, the core task currently being executed by the GPU is broken down into stages, and the non-real-time computing stages are extracted; when the GPU utilization rate is detected to recover redundant computing power and have the ability to take on tasks again, and the CPU utilization rate is greater than or equal to a preset value, a task rollback mechanism is triggered after the GPU load is restored, that is, the non-critical stages of the core tasks that were previously migrated to the CPU due to the high GPU load are traced back; when the memory utilization rate is detected to trigger an overload threshold, a memory overload protection mechanism is triggered, that is, auxiliary tasks at the edge are identified and filtered first.

[0014] In one technical solution of the lightweight intelligent control method for the aforementioned low-altitude operation scenario, the energy consumption sensing and scheduling strategy is as follows: when the real-time battery power is lower than a preset value, the computing power occupation of auxiliary tasks is automatically reduced.

[0015] The beneficial effects of the lightweight intelligent control method for low-altitude operation scenarios provided by this invention are as follows: This method combines a lightweight deep learning model with an attention mechanism, accurately focusing on key low-altitude targets while adapting to the limited computing power of the edge. Combined with a dynamic data augmentation module to expand sample diversity for low-altitude-specific interference, it maintains the lightweight characteristics of the model while significantly improving perception accuracy and anti-interference capability in complex environments, breaking through the bottleneck of "lightweight and high precision being mutually exclusive" in traditional technologies. Simultaneously, a spatiotemporal calibration model is used to accurately process multi-source data such as visual and LiDAR data, strictly controlling time synchronization errors and spatial positioning deviations, avoiding data misalignment contradictions caused by aircraft vibration, and providing a reliable fusion data foundation for decision-making. For dynamic scenarios, a lightweight LSTM network fuses historical motion data and environmental constraints to predict target trajectories, meeting the real-time requirements of the edge while accurately capturing complex movements such as sudden changes in direction, reserving sufficient response time for obstacle avoidance decisions, and improving the ability to cope with dynamic scenarios. Furthermore, the edge computing system achieves flexible resource allocation through a dynamic matching mechanism of task priority and computing load, and adjusts the computing power of auxiliary tasks in conjunction with an energy-aware scheduling strategy. This ensures efficient execution of core tasks while significantly improving computing power utilization and extending operational endurance, thus balancing high-performance requirements with low-energy constraints. In summary, this method comprehensively enhances the autonomous perception, decision-making, and resource management capabilities of low-altitude UAVs in complex environments, enabling them to adapt more safely, efficiently, and continuously to various low-altitude operational scenarios such as power line inspection and emergency rescue. Attached Figure Description

[0016] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a schematic flowchart of a lightweight intelligent control method for low-altitude operation scenarios according to an embodiment of the present invention. Detailed Implementation

[0017] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0018] like Figure 1 As shown, this invention provides a lightweight intelligent control method for low-altitude operation scenarios, comprising: Step S1: Based on a lightweight deep learning model, focus on key low-altitude targets through an attention mechanism; and perform data augmentation through a dynamic data augmentation module to expand the diversity of training samples. Step S2: Call the spatiotemporal calibration model based on vision, lidar, millimeter-wave radar, and IMU to process multi-source data and control the time synchronization error and spatial positioning deviation within a preset range; Step S3: Based on the dynamic obstacle trajectory prediction module, historical motion data and environmental constraints are fused together to use a lightweight LSTM network to predict the trajectory of the dynamic target in the future time period; Step S4: Based on the edge lightweight intelligent computing module, the task priority-computing load dynamic matching mechanism is invoked to realize task allocation, and the computing power occupation of non-critical tasks is adjusted according to the designed energy consumption-aware scheduling strategy.

[0019] In this embodiment, the lightweight deep learning model can be YOLOv8-tiny or MobileNet-SSD. Through an attention mechanism, the model automatically focuses on key targets in low-altitude scenarios, such as obstacles, work sites, and no-fly zone boundaries. For example, in a power line inspection scenario, the model can prioritize core targets such as transmission lines and towers, reducing computational resource consumption in non-critical areas.

[0020] Furthermore, spatial attention is employed: highlighting pixel weights in the target region, such as the elongated regions of electric field lines or the motion regions of moving objects. Specifically, a 1×1 convolution (with the same number of output channels as the feature map) + sigmoid activation layer can be added after the output feature map of the model backbone to generate a spatial attention weight map of the same size as the feature map, which is then multiplied by the original feature map.

[0021] Alternatively, channel attention can be employed: enhancing the weights of key feature channels, such as the edge feature channels of electric lines and the motion feature channels of moving objects. Specifically, the feature map (C×H×W) output by the Backbone is subjected to global average pooling (resulting in C×1×1), compressed through two fully connected layers (the first layer has C / 4 neurons, the second layer has C neurons, and ReLU activation is used in between) and then expanded to output C×1×1 channel weights, which are then multiplied with the original feature map.

[0022] In one embodiment, the lightweight deep learning model is YOLOv8-tiny. Based on the lightweight deep learning model, focusing on key low-altitude targets through an attention mechanism includes: inserting attention modules into the Backbone and Neck of YOLOv8-tiny, and adding a target difficulty coefficient during training in the context of low-altitude scene characteristics: dynamically increasing the attention weights of small targets, occluded targets, and targets with complex backgrounds through a loss function.

[0023] In this embodiment, the core characteristic of a small target is that it occupies a very small area in the image and has little pixel information. For example, when the pixel area (width × height) of the target is less than 32 × 32 pixels, it is determined to be a small target.

[0024] First, mark the complete outline and the actual visible outline of the target, and calculate the occlusion ratio = (complete area - visible area) / complete area; when the occlusion ratio is greater than 30%, it is determined to be an occluded target. If the scene is low-altitude with dense obstacles, it can be reduced to 20% according to actual needs.

[0025] Extract the target region (ROI) and the surrounding background region within a range of 10-20 pixels, and calculate the color histogram similarity between the two, such as the Bach distance. The smaller the value, the more similar the target is, or the edge gradient difference. When the color histogram similarity is <0.3 (Bach distance threshold) or the edge gradient difference is <preset threshold, the target is judged to be a complex background target.

[0026] The base difficulty coefficient is 1.0 (for ordinary, easily detectable targets). Difficulty bonuses are applied to each of the three target types, with the final coefficient controlled between 1.0 and 3.0. The specific settings are shown in Table 1. Table 1

[0027] First, calculate the difficulty coefficient for each labeled target (small target, occluded target, etc.) in the dataset and bind it with information such as target location and category to form a correspondence table. During training for each batch of images, first, let the model find the prediction box corresponding to the labeled target and determine the three types of original losses. Then, retrieve the difficulty coefficient of the target from the correspondence table and multiply the coefficient by the three types of original losses respectively to amplify the loss of difficult targets (for example, the loss of occluded targets changes from 0.2 to 0.32). Take the average of the weighted losses of all targets to obtain the total loss, and use the total loss to update the model parameters to avoid abnormal loss due to too many difficult targets.

[0028] In one embodiment, the dynamic data augmentation module includes: a static augmentation module for performing geometric transformations and color transformations on the training set images; and a dynamic augmentation module for selecting a corresponding augmentation strategy based on low-altitude scene interference, including: rain, fog, sudden changes in illumination, and dynamic occlusion.

[0029] In this embodiment, geometric transformations, such as random cropping, rotation, scaling, and flipping, simulate the appearance effect of the target at different angles and distances; color transformations, such as adjusting brightness, contrast, and saturation, or adding Gaussian noise, simulate image changes under different lighting conditions.

[0030] Simulated low-altitude scene interference: Rain: Generate rain lines in random directions (5-20 pixels in length, density increases with the probability of the "rainy day" tag; probability 0-0.3 corresponds to low-density rain lines (100-300 lines per frame), probability 0.3-0.7 corresponds to medium-density rain lines (300-800 lines per frame), and probability 0.7-1 corresponds to high-density rain lines (800-1500 lines per frame)). The enhanced image output after superimposing the rain lines is represented as: I'=I*(1-rain_mask)+rain_line*rain_intensity Where I is the original low-altitude scene image without rain lines, rain_mask is the rain line region mask, with a value of 0-1; rain_intensity is the rain line brightness intensity, with a value of 0.2-0.8, which increases synchronously with the rain density; Fog: Enhanced image output after adding fog effect using an atmospheric scattering model: I' = I*t + A*(1-t) Where I is the original low-altitude scene image without fog effect, t is the transmittance, and A is the atmospheric light. The value of t is dynamically adjusted according to the fog concentration, which is [0.3, 0.8]. Fog concentration levels 1-5 correspond to t values ​​of 0.8 to 0.3. The value of A is 1.2-1.8 times the average global brightness of the image.

[0031] Sudden Illumination Changes: Backlight: Brightness in local areas (such as the sky) is increased by 2-3 times, while the brightness of the target area is reduced by 30% (simulating direct sunlight). The sky area is automatically identified through image segmentation algorithms, and the brightness adjustment range is 200-255 (pixel value 0-255). Nighttime: Overall brightness is reduced by 50% (overall pixel value multiplied by 0.5), and random noise is added (Gaussian noise, variance 0.01-0.05, variance 0.03-0.05 in low light environment, variance 0.01-0.03 in low light environment), combined with local infrared effects (brightness of the target area is enhanced, and the pixel value is increased to 150-200).

[0032] Dynamic occlusion: Randomly add occlusion objects (such as tree branches or building debris), with the occlusion area ≤ 40% of the target area (simulating low-altitude obstacle occlusion). The occlusion object materials are taken from the occlusion sample library collected from real low-altitude scenes, and the occlusion position is determined by a random coordinate generator.

[0033] The core of selecting enhancement strategies based on low-altitude scene interference is on-demand matching: first, determine the type of interference to be simulated (rain, fog, sudden changes in illumination, dynamic occlusion), and then call the corresponding preset rules to generate enhancement effects. The specific logic is as follows: Step 1: Determine the type of interference that needs to be enhanced. First, determine which type (or combination of) of the interference to be simulated is "rain, fog, sudden changes in lighting, or dynamic occlusion." Triggering is primarily done in two ways: Triggering by image label: If the training image is labeled with scene interference labels (e.g., "rainy day," "foggy day," "nighttime backlight"), directly identify the type of interference to be enhanced based on the label. Example: Label "rainy day" → trigger rain interference enhancement; Label "foggy day + dynamic occlusion" → trigger both fog and occlusion enhancement simultaneously.

[0034] Triggered by random probability: If the image is unlabeled, interference will be randomly activated according to a preset probability (simulating the randomness of real low-altitude scenes), such as: rain / fog: each with a 20% probability; sudden changes in lighting (backlight 15%, nighttime 25%); dynamic occlusion: 30% probability (simulating sudden obstacles).

[0035] Step 2: Enhancing strategies based on interference type For different types of interference, the corresponding parameter configuration and generation logic are invoked: Rain interference → Match rain line parameters according to rainy day probability, and use rainy day probability (0-1) to control the density and brightness of rain lines. The higher the probability, the stronger the interference.

[0036] Fog interference → Match fog effect parameters according to fog concentration level.

[0037] Sudden changes in lighting conditions → Match brightness / noise parameters according to "lighting type": First, distinguish between backlight and nighttime, then adjust brightness and noise accordingly. For example: Backlight: Use a segmentation algorithm to identify the sky area, adjust the brightness to 200-255, and reduce the brightness of the target area by 30%; Nighttime: Overall brightness ×0.5, add Gaussian noise with variance of 0.03-0.05 in low light environment, add noise of 0.01-0.03 in low light, and increase the brightness of the target area to 150-200.

[0038] Dynamic occlusion → Matching occlusion parameters according to random rules: Simulating sudden low-altitude occlusion, the core control is the type, location, and area of ​​the occlusion object. Specifically, occlusion objects are randomly selected from a real sample library; a random coordinate generator is used to determine the occlusion position (prioritizing coverage of the target area, but not deliberately blocking the entire area); ensuring that the occlusion area is ≤40% of the target area, and that the edge processing is natural.

[0039] Step 3: The superposition order of composite interference If multiple interferences are triggered simultaneously (such as rain + backlight, fog + occlusion), stack them in the order of basic interference first, then special interference to avoid effect conflicts: First, perform static enhancements: geometric transformations such as cropping / rotation; color transformations such as adjusting brightness / adding basic noise). Then, overlay natural phenomenon interference: rain / fog, either rain followed by fog or fog followed by rain, without affecting the core effect. Next, overlay sudden changes in lighting: backlight / nighttime, adjusting the overall brightness to avoid being covered by subsequent occlusions. Finally, overlay dynamic occlusion: add occlusions to the final image to ensure the occlusion effect is visible.

[0040] In one embodiment, the spatiotemporal calibration model includes: a time synchronization sub-model, a spatial calibration sub-model, and a vibration compensation sub-model. It calls upon a spatiotemporal calibration model based on vision, lidar, millimeter-wave radar, and IMU to process multi-source data, controlling time synchronization errors and spatial positioning deviations within a preset range. Specifically, controlling time synchronization errors and spatial positioning deviations within a preset range includes: calibrating the consistency of acquisition times of each sensor through the time synchronization sub-model; mapping the original data to a unified body coordinate system through the spatial calibration sub-model; and correcting vibration misalignment through the vibration compensation sub-model, thus completing the spatiotemporal alignment of multi-source data.

[0041] In this embodiment, the time synchronization sub-model unifies the acquisition time of all sensor data (e.g., based on the IMU clock) and controls the time difference within a preset range to avoid "some sensors have already recorded data while others have not yet acquired data in the same scenario".

[0042] Space calibration sub-model: Transform all sensor data into a unified coordinate system centered on the vehicle to eliminate positional deviations caused by different installation locations and ensure that "the same target corresponds in position in images and radar data".

[0043] Vibration compensation sub-model: Using vehicle vibration data detected by IMU, the offset of the camera and radar is corrected in reverse (for example, if the camera shakes, the image is adjusted back) to avoid data misalignment caused by vibration.

[0044] In one embodiment, the LSTM network contains two hidden layers, each with 64 hidden units. The input features are the target's position data, velocity data, and acceleration data over the past n seconds, and the output is the predicted position over the next m seconds.

[0045] In one embodiment, the dynamic obstacle trajectory prediction module introduces an attention mechanism to correct the prediction results in real time in sudden change-of-direction scenarios. Specifically, the attention mechanism assigns higher weight to historical data before and after the change-of-direction moment by focusing on the characteristics of the target's movement direction change.

[0046] In this embodiment, if the rate of change of velocity direction angle is greater than 10° / s, it is determined to be a sudden change of direction. The historical data before and after the change of direction are given higher weight (weight coefficient 1.2-1.5), and the obstacle avoidance response time is shortened by 0.2s after correction.

[0047] In one embodiment, the task priority-computing load dynamic matching mechanism is as follows: prioritize and classify edge tasks to clarify the scheduling priority and hardware compatibility of different tasks; monitor GPU / CPU load in real time, and dynamically allocate tasks according to the load scenario based on task priority and type to complete the initial matching of tasks and hardware; after the initial task allocation, continuously monitor the real-time load of GPU, CPU and memory, and if the load abnormality threshold is triggered, start the adjustment mechanism to correct the initial allocation deviation and ensure the balance of computing power supply and demand and the stable operation of the system.

[0048] In one embodiment, the edge task includes a core task and auxiliary tasks; wherein the core task includes obstacle avoidance decision-making, target detection, and trajectory planning, and the auxiliary task includes data storage and status monitoring.

[0049] In this embodiment, the task priority quantification standard is set as follows: obstacle avoidance decision (priority 10-12), target detection (priority 8-10), trajectory planning (priority 7-9), data storage (priority 3-5), and status monitoring (priority 4-6). The higher the priority value, the more urgent the task.

[0050] In one embodiment, the adjustment mechanism is as follows: when the GPU utilization rate is detected to trigger a high load threshold and the CPU is in a low load state, a GPU high load mitigation mechanism is triggered, that is, the core task currently being executed by the GPU is broken down into stages, and the non-real-time computing stages are extracted; when the GPU utilization rate is detected to recover redundant computing power and have the ability to take on tasks again, and the CPU utilization rate is greater than or equal to a preset value, a task rollback mechanism is triggered after the GPU load is restored, that is, the non-critical stages of the core tasks that were previously migrated to the CPU due to the high GPU load are traced back; when the memory utilization rate is detected to trigger an overload threshold, a memory overload protection mechanism is triggered, that is, auxiliary tasks at the edge are identified and filtered first.

[0051] In one embodiment, the energy consumption sensing scheduling strategy is: when the real-time battery power is lower than a preset value, the computing power usage of auxiliary tasks is automatically reduced.

[0052] In this embodiment, the real-time battery level is compared with a preset battery level threshold to determine whether to trigger the computing power adjustment mechanism. This threshold can be set according to device requirements, such as 20% or 30% remaining battery level.

[0053] Specific ways to reduce computing power usage include reducing the CPU / GPU resource allocation ratio of auxiliary tasks (such as reducing the original CPU computing power usage from 20% to 5%), reducing the task execution frequency (such as changing background synchronization from once every 5 minutes to once every 30 minutes), suspending some non-urgent auxiliary tasks, etc. By reducing the energy consumption of auxiliary tasks, more computing power and power resources can be allocated to core tasks.

[0054] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A lightweight intelligent control method for low-altitude operation scenarios, characterized in that, include: Based on a lightweight deep learning model, an attention mechanism is used to focus on key low-altitude targets; And data augmentation is performed through a dynamic data augmentation module to expand the diversity of training samples; The system calls upon a spatiotemporal calibration model based on vision, lidar, millimeter-wave radar, and IMU to process multi-source data, controlling time synchronization errors and spatial positioning deviations within a preset range. Based on the dynamic obstacle trajectory prediction module, historical motion data and environmental constraints are fused together, and a lightweight LSTM network is used to predict the trajectory of dynamic targets in the future time period. The edge-based lightweight intelligent computing module invokes a task priority-computing load dynamic matching mechanism to allocate tasks and adjusts the computing power usage of non-critical tasks according to the designed energy consumption-aware scheduling strategy.

2. The method according to claim 1, characterized in that, The lightweight deep learning model is YOLOv8-tiny. Based on the lightweight deep learning model, the attention mechanism focuses on key targets at low altitudes, including: inserting attention modules into the Backbone and Neck of YOLOv8-tiny, and adding target difficulty coefficients during training in the context of low-altitude scene characteristics: dynamically increasing the attention weights of small targets, occluded targets, and targets with complex backgrounds through the loss function.

3. The method according to claim 1, characterized in that, The dynamic data augmentation module includes: a static augmentation module, used to perform geometric transformations and color transformations on the training set images; The dynamic enhancement module is used to select the corresponding enhancement strategy based on low-altitude scene interference, which includes: rain, fog, sudden changes in illumination, and dynamic occlusion.

4. The method according to claim 1, characterized in that, The spatiotemporal calibration model includes: a time synchronization sub-model, a spatial calibration sub-model, and a vibration compensation sub-model; The system calls upon a spatiotemporal calibration model based on vision, lidar, millimeter-wave radar, and IMU to process multi-source data, controlling time synchronization errors and spatial positioning deviations within a preset range. Specifically, this includes: calibrating the consistency of acquisition times of each sensor through a time synchronization sub-model, mapping the raw data to a unified body coordinate system through a spatial calibration sub-model, and correcting vibration misalignment through a vibration compensation sub-model, thereby completing the spatiotemporal alignment of multi-source data.

5. The method according to claim 1, characterized in that, The LSTM network contains two hidden layers, each with 64 hidden units. The input features are the target's position, velocity, and acceleration data over the past n seconds, and the output is the predicted position over the next m seconds.

6. The method according to claim 5, characterized in that, In the scenario of sudden change of direction, the dynamic obstacle trajectory prediction module introduces an attention mechanism to correct the prediction results in real time. Specifically, the attention mechanism focuses on the characteristics of the change in the target's motion direction and assigns higher weight to historical data before and after the change of direction.

7. The method according to claim 1, characterized in that, The task priority-computing load dynamic matching mechanism is as follows: Prioritize and classify edge tasks to clarify the scheduling priority and hardware compatibility of different tasks. Real-time monitoring of GPU / CPU load, combined with task priority and type, dynamically allocates tasks according to the load scenario, and completes the initial matching of tasks and hardware; After the initial task allocation, the real-time load of GPU, CPU and memory is continuously monitored. If the abnormal load threshold is triggered, the adjustment mechanism is activated to correct the initial allocation deviation and ensure the balance of computing power supply and demand and the stable operation of the system.

8. The method according to claim 7, characterized in that, The edge tasks include core tasks and auxiliary tasks; the core tasks include obstacle avoidance decision-making, target detection, and trajectory planning, and the auxiliary tasks include data storage and status monitoring.

9. The method according to claim 7, characterized in that, The adjustment mechanism is as follows: when the GPU utilization rate is detected to trigger the high load threshold and the CPU is in a low load state, the GPU high load mitigation mechanism is triggered, that is, the core task currently being executed by the GPU is broken down into stages and the non-real-time computing stages are extracted. When it is detected that the GPU utilization has recovered to redundant computing power and has the ability to take on tasks again, and the CPU utilization is greater than or equal to the preset value, the task rollback mechanism after the GPU load is restored will trace back the non-critical core tasks that were previously migrated to the CPU due to the high GPU load. When the memory usage rate is detected to have triggered the overload threshold, the memory overload protection mechanism is activated, which prioritizes the identification and filtering of auxiliary tasks at the edge.

10. The method according to claim 8, characterized in that, The energy consumption sensing and scheduling strategy is as follows: when the real-time battery power is lower than a preset value, the computing power usage of the auxiliary task is automatically reduced.

Images