Low-altitude economy unmanned aerial vehicle data processing method and system
By calculating the information entropy value of UAV images and filtering redundant frames based on the difference between adjacent frames, the problem of storage and computing resource consumption in UAV image data processing is solved, achieving efficient data compression and real-time target recognition, and adapting to diverse application scenarios.
Patent Information
- Application Number
- CN202610916783.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-24
- Publication Date
- 2026-11-10
AI Technical Summary
Existing UAV image data processing systems, operating at a fixed frame rate, result in redundant frames consuming storage space and computing resources, impacting system real-time performance and continuous execution capabilities, especially during long-duration cruise missions where storage devices quickly become saturated.
By calculating the information entropy value of multiple frames and determining redundancy based on the information entropy difference between adjacent frames, redundant frames with information entropy differences lower than a preset threshold are discarded, while key frames with information entropy differences higher than or equal to the preset threshold are retained and then processed for target recognition and storage.
It significantly reduces storage space usage by 40% to 60%, reduces the computational burden of target recognition algorithms by 30% to 45%, improves system real-time performance, ensures the integrity of key frames and detection accuracy, and adapts to the characteristics of different application scenarios.
Smart Images

Figure CN123071883A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to a data processing method and system for low-altitude economical UAVs. Background Technology
[0002] With the booming development of the low-altitude economy, drone technology is increasingly being applied in fields such as geographic surveying, environmental monitoring, precision agriculture, logistics and distribution, and urban management. Equipped with high-definition cameras, drones can complete large-scale, long-duration patrol missions in complex environments, and the massive amounts of image and video data they collect are crucial data sources for intelligent decision-making.
[0003] When performing monitoring, inspection, or real-time tracking tasks, UAVs need to perform target identification and information extraction on continuously acquired image frames. Continuous frame image data often contains a large amount of similar or repetitive content. How to efficiently filter out key information frames from these continuous frames has become one of the core technical challenges in improving the performance of UAV data processing systems.
[0004] Current technologies primarily employ a fixed frame rate for image acquisition and processing. This approach suffers from several significant drawbacks: First, fixed frame rate acquisition fails to differentiate between frames, continuing to sample at a fixed frequency even in relatively static or low-lying scenes. This results in a large number of redundant frames being retained, consuming valuable storage space and transmission bandwidth. Second, the continuous accumulation of redundant frames significantly increases the computational burden on subsequent target recognition algorithms, leading to increased processing latency and a substantial decrease in system real-time performance. Third, on resource-constrained embedded processing platforms, excessive processing of redundant data can crowd out computational resources for critical target detection, easily causing recognition delays or missed detections. Finally, the indiscriminate full-data storage strategy can lead to rapid storage saturation during long-duration cruise missions, severely impacting the continuous execution capability of UAV missions. These technical issues collectively limit the potential value of UAVs in low-altitude economic applications. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a data processing method and system for low-altitude economic unmanned aerial vehicles (UAVs), which solves the problems mentioned in the background section.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A low-altitude economic unmanned aerial vehicle (UAV) data processing method includes the following steps: S1. Continuously acquire multi-frame image data output by the drone's camera device; S2. Calculate the information entropy value of each frame in the multi-frame image data; S3. Determine redundancy based on the information entropy difference between adjacent frames; S4. Based on the redundancy determination result, discard redundant frames with an information entropy difference lower than a preset threshold, and retain key frames with an information entropy difference higher than or equal to the preset threshold. S5. Perform target recognition and storage processing on the keyframes.
[0007] Preferably, the low-altitude economic unmanned aerial vehicle (UAV) data processing system includes: an image acquisition module for continuously acquiring multiple frames of image data output by the UAV camera device; an information entropy calculation module for calculating the information entropy value of each frame in the multiple frames of image data; a redundancy determination module for redundancy determination based on the information entropy difference between adjacent frames; a frame filtering module for discarding redundant frames and retaining key frames according to the redundancy determination result; and a target recognition and storage module for target recognition and storage processing of the key frames.
[0008] Preferably, step S2 specifically includes the following steps: S21. Divide each frame of image into multiple sub-regions; S22. Calculate the probability distribution of pixel grayscale values for each sub-region; S23. Calculate the information entropy value of each sub-region based on the probability distribution of pixel grayscale values; S24. The information entropy values of each sub-region are weighted and summed to obtain the overall information entropy value of the frame image.
[0009] Preferably, step S3 includes the following specific steps: S31. Calculate the information entropy difference between the current frame and the previous frame; S32. Compare the information entropy difference with a preset redundancy threshold; S33. When the information entropy difference is less than the redundancy threshold, the current frame is determined to be a redundant frame; S34. When the information entropy difference is greater than or equal to the redundancy threshold, the current frame is determined to be a key frame.
[0010] Preferably, the preset redundancy threshold is dynamically adjusted according to the UAV's flight altitude, camera resolution, and target detection accuracy requirements.
[0011] The specific dynamic adjustment mechanism is as follows: When the drone's flight altitude decreases, the proportion of pixels occupied by ground objects in the image increases, making changes in information entropy more sensitive. In this case, the redundancy threshold should be appropriately lowered to capture more detailed changes. When the flight altitude increases, the changes in image content are relatively gradual, and the redundancy threshold should be appropriately increased to filter more similar frames. Simultaneously, as the camera resolution increases, the image contains more detailed information, and the dynamic range of information entropy becomes larger. Therefore, the threshold needs to be adjusted accordingly to adapt to the information content evaluation at higher resolutions. Furthermore, the higher the target detection accuracy requirement, the higher the requirement for the integrity of keyframes. The redundancy threshold setting should be conservative to ensure that no frame that may contain the target is missed.
[0012] Preferably, when discarding redundant frames in step S4, the method further includes recording the timestamp and location information of the redundant frame, and storing the timestamp and location information in association with adjacent key frames to support subsequent data backtracking and integrity verification. The specific association storage mechanism is as follows: when a redundant frame is discarded, the system records the timestamp of that frame and the UAV's position coordinates at that moment, and establishes an association with the nearest key frame at that moment, forming a data chain in the time series. When data backtracking is needed later, the corresponding historical moment can be quickly located through the timestamp of the key frame, and the time range of the discarded redundant frame can be restored through the association information, thereby ensuring that the data throughout the flight process is continuously traceable in the time dimension.
[0013] Preferably, step S5, which involves target recognition and storage processing of the keyframe, specifically includes: S51. Preprocess the keyframes, including denoising and image enhancement; S52. Perform target recognition on keyframes based on a preset target detection model; S53. Associate the recognition results with the corresponding keyframes and store them in the data storage unit.
[0014] The denoising operation in the preprocessing step employs an adaptive filtering algorithm, dynamically adjusting filtering parameters based on the noise characteristics of local image regions to remove noise while preserving image edge details as much as possible. The image enhancement operation uses a contrast adaptive adjustment algorithm, performing independent contrast enhancement processing on each frame based on the image histogram distribution to make target features more prominent, providing more favorable input conditions for subsequent target recognition. The target recognition model employs a deep neural network architecture, capable of automatically learning target feature representations and classifying and locating multiple targets in the image. The storage unit uses a structured data organization method, encapsulating the original data, recognition results, timestamps, and location information of each frame into complete data units, facilitating subsequent data retrieval and analysis.
[0015] Preferably, the information entropy calculation module includes: Sub-region division unit, used to divide each frame of image into multiple sub-regions; The distribution statistics unit is used to separately calculate the probability distribution of pixel grayscale values in each sub-region; The entropy calculation unit is used to calculate the information entropy value of each sub-region based on the probability distribution of the pixel grayscale value and to synthesize the overall information entropy value.
[0016] The sub-region segmentation unit divides the image into a preset number of non-overlapping sub-regions. Information entropy is calculated independently for each sub-region. This partitioned calculation method effectively captures the differences in information content between different regions of the image, especially for complex scenes that simultaneously contain sky and ground, or distant and near views. Partitioned calculation provides a more accurate assessment of information content. The distribution statistics unit traverses all pixels within each sub-region, counts the frequency of occurrence at each gray level, and calculates a normalized probability distribution, providing basic data for information entropy calculation. The entropy calculation unit uses the Shannon entropy formula to calculate the information entropy of the probability distribution of each sub-region and obtains the overall information entropy value through weighted summation. The weight of each sub-region is set according to its importance in object detection, with regions containing more texture information assigned higher weights.
[0017] Preferably, the system further includes a threshold dynamic adjustment module, used to dynamically adjust the preset redundancy threshold used in the redundancy determination module according to the UAV flight parameters and environmental conditions.
[0018] The threshold dynamic adjustment module receives flight parameters such as drone flight altitude, speed, and heading angle, as well as environmental conditions such as ambient light intensity and scene complexity as input. It calculates the current optimal redundancy threshold through a preset adjustment strategy and updates the redundancy determination module with this threshold.
[0019] The adjustment strategy is formulated based on a large amount of experimental data, analyzing the statistical characteristics of information entropy distribution under different parameter combinations, and establishing a mapping relationship between parameters and thresholds. When the UAV enters complex terrain or areas with dense targets, the system automatically lowers the redundancy threshold to capture more details of change; when the UAV cruises in open and flat areas, the system automatically raises the redundancy threshold to filter out more similar frames. This adaptive adjustment mechanism enables the system to maintain optimal data processing efficiency in different application scenarios.
[0020] Preferably, the target recognition and storage module includes: The preprocessing unit is used for denoising and image enhancement of keyframes; The target detection unit is used to identify targets in keyframes based on a preset target detection model; the storage unit is used to associate and store the recognition results with the corresponding keyframes.
[0021] The preprocessing unit integrates a variety of image processing algorithms, including spatial domain filtering, frequency domain filtering, and sharpening, and can adaptively select processing strategies based on image quality.
[0022] The target detection unit adopts a multi-scale feature fusion detection framework, which can simultaneously detect targets of different scales in the image and output the target's category, location, and confidence information.
[0023] The storage unit adopts a tiered storage architecture, storing hot data and cold data in high-speed cache and long-term storage media respectively, thereby reducing storage costs while ensuring data access speed.
[0024] This invention provides a data processing method and system for low-altitude economic unmanned aerial vehicles (UAVs), which has the following beneficial effects: (1) During system operation, compared with the prior art, the beneficial effects of the present invention are as follows: The present invention realizes intelligent filtering and compression of UAV-collected image data by calculating the information entropy value of consecutive frames and making redundancy judgment based on the information entropy difference between adjacent frames. First, information entropy, as a scientific indicator for measuring the amount of image information, can accurately reflect the detailed information and content changes contained in a single frame image. Compared with the traditional fixed frame rate sampling method, it can fundamentally avoid the problem of indiscriminate treatment of information content, so that redundant frames with low information content are automatically discarded, and only key frames containing rich content changes are retained. This mechanism directly leads to a significant reduction in storage space. According to the verification of different application scenarios, the storage space can be reduced by 40% to 60%, effectively solving the prominent problem of limited storage capacity in long-term UAV cruise missions and significantly extending the effective operation time of UAVs.
[0025] (2) Due to the significant reduction in the number of redundant frames that need to be processed, the computational burden of the target recognition algorithm is significantly reduced, allowing the system to concentrate its limited processing resources on the analysis of key frames. This improvement is directly reflected in the reduction of target detection response latency, which can be reduced by approximately 30% to 45%, significantly improving the system's real-time performance. In application scenarios with high real-time requirements, such as monitoring and inspection, this improvement can help UAVs respond to emergencies more quickly, detect and track targets more promptly, and effectively avoid target omissions or recognition delays caused by processing delays. At the same time, the optimized allocation of processing resources also reduces the overall energy consumption of the system, which has significant practical value for UAVs with limited battery capacity.
[0026] (3) This invention strictly judges redundant frames based on the information entropy difference, ensuring that key frames containing significant information changes are fully preserved. Through reasonable threshold settings and dynamic adjustment mechanisms, the decrease in recognition accuracy of key target frames can be controlled within 2%. The key to this accuracy preservation mechanism is that the information entropy difference can effectively reflect the content differences between images. When an important target appears in the scene or significant changes occur, the information entropy value will show obvious jumps. These jumps will be captured by the redundancy judgment module and the corresponding key frames will be preserved. Therefore, even if a large number of redundant frames are discarded, frames containing key targets can still be fully preserved, and the target detection accuracy will not be significantly affected. This effectively balances data compression efficiency and target detection accuracy, meeting the dual requirements of data integrity and detection accuracy in practical applications.
[0027] (4) This invention supports dynamic adjustment of the redundancy judgment threshold based on factors such as UAV flight altitude, camera resolution, and environmental conditions, enabling the system to adapt to the characteristics of different application scenarios. Whether it is traffic monitoring in urban scenarios, power line inspection in the wild, or rescue search under complex terrain conditions, the method of this invention can provide the optimal data processing strategy through parameter adaptation. This adaptability comes from the introduction of the threshold dynamic adjustment module, which can calculate the current optimal threshold based on real-time flight parameters and environmental conditions, so that the system can maintain the best data processing effect in various application scenarios. Especially in practical applications with complex and ever-changing environmental conditions, this dynamic adjustment mechanism ensures that the system will not cause key frame omissions or excessive redundant frame retention due to improper fixed threshold settings, significantly enhancing the adaptability of the UAV data processing system to diverse application scenarios.
[0028] (5) By employing sub-region division and weighted summation to calculate information entropy, the information content of a single frame image can be accurately quantified, providing a reliable basis for redundancy determination. The sub-region division strategy can effectively capture the differences in information content among different regions of the image, avoiding the problem of insufficient sensitivity of global information entropy to local changes; the weighted summation mechanism evaluates the information content differently based on the importance of each region in target detection, making the calculation of information entropy more in line with actual application needs. This technical detail optimization further improves the accuracy and reliability of redundancy determination. By retaining the timestamp and location information of key frames and storing them in association with adjacent key frames, the traceability and integrity verification capabilities of the data are ensured. In subsequent data analysis and task review, the flight status at any moment can be quickly located using timestamps and location information, and the complete flight trajectory and data sequence can be reconstructed. This mechanism plays an important supporting role in subsequent work such as accident investigation, task evaluation, and algorithm optimization, significantly improving the practical value of the entire data processing system. Attached Figure Description
[0029] Figure 1 This is a flowchart illustrating the low-altitude economic unmanned aerial vehicle (UAV) data processing method of the present invention. Figure 2 This is a schematic diagram of the module composition of the low-altitude economic unmanned aerial vehicle (UAV) data processing system of the present invention. Detailed Implementation
[0030] 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. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0031] Example 1 This embodiment uses an urban traffic monitoring scenario as an example to illustrate in detail the specific implementation of the low-altitude economic unmanned aerial vehicle (UAV) data processing method and system of the present invention. In this embodiment, a city traffic management department has deployed a UAV patrol system on key road sections for real-time monitoring of road traffic conditions. When the UAV performs its monitoring task, its onboard camera captures images of the road ahead at a rate of 30 frames per second. The system needs to efficiently filter out target frames containing key traffic information from these continuously acquired image frames and perform target recognition and storage processing.
[0032] In step S1, the image acquisition module continuously acquires multiple frames of image data output by the UAV camera. Specifically, the camera uses a visible light sensor with a 12-megapixel resolution, and the acquired raw image data is in RGB format, with a total pixel count of 4032×3024 for each frame. The image acquisition module continuously acquires image frames at a 30-millisecond acquisition cycle and transmits each frame of image data to the information entropy calculation module for subsequent processing via a high-speed data transmission channel. The image acquisition module integrates a frame buffer memory to temporarily store several recently acquired frames of image data, facilitating parallel processing by the information entropy calculation module. The frame buffer memory is designed to cache at least 10 frames of raw image data simultaneously, ensuring that there is no processing bottleneck due to insufficient data supply during the information entropy calculation process.
[0033] In step S2, the information entropy calculation module calculates the information entropy value for each frame of the acquired image. This embodiment uses a combination of sub-region division and weighted summation to calculate the information entropy. The sub-region division unit divides each frame of the image into 16 non-overlapping sub-regions of 4×4 pixels, each sub-region having a pixel size of 1008×756. The sub-region division unit uniformly cuts the original image using a preset segmentation algorithm to ensure that there are no overlapping areas between the sub-regions, thereby guaranteeing the independence and accuracy of subsequent information entropy calculations. The distribution statistics unit traverses all pixels within each sub-region and counts the frequency of each gray level. In this embodiment, the pixel gray levels are divided into 256 levels from 0 to 255. The distribution statistics unit records the number of pixels at each gray level appearing in the corresponding sub-region and calculates the normalized probability distribution. The specific calculation process is as follows: first, count the number of pixels with gray value i within the sub-region, then divide by the total number of pixels in that sub-region to obtain the probability distribution value corresponding to gray value i. The entropy calculation unit calculates the information entropy value of each sub-region based on the probability distribution of pixel grayscale values, using the Shannon entropy formula. The principle of Shannon entropy is that information entropy equals the negative sum of the negative logarithms of all probability values, reflecting the average amount of information contained in a random variable. In this embodiment, the entropy calculation unit independently calculates the information entropy value of each sub-region, and then performs a weighted summation according to preset weighting coefficients to obtain the overall information entropy value of the frame image. The weighting coefficients of each sub-region are set according to the importance of that region in the traffic monitoring scenario; regions near the center of the image and containing lane information are given higher weights, while the sky regions at the edges of the image are given lower weights.
[0034] In step S3, the redundancy determination module performs redundancy determination based on the information entropy difference between adjacent frames. The redundancy determination module first reads the overall information entropy values of the current frame and the previous frame, and then calculates the difference between them. The calculated difference reflects the degree of information change between two consecutive frames. When the drone is flying at a constant speed and the monitored scene is relatively stable, the information entropy difference between consecutive frames is usually small; when traffic events such as vehicle lane changes or pedestrian crossings occur in the scene, the information entropy value of the current frame will show a significant jump, and the information entropy difference will increase accordingly. The redundancy determination module compares the calculated information entropy difference with a preset redundancy threshold. In this embodiment, the preset redundancy threshold is set to 0.3. This value was obtained through statistical analysis of a large amount of urban scene traffic monitoring experimental data and has good applicability and robustness. When the information entropy difference is less than the redundancy threshold of 0.3, the redundancy determination module determines the current frame as a redundant frame; when the information entropy difference is greater than or equal to the redundancy threshold of 0.3, the redundancy determination module determines the current frame as a key frame.
[0035] In step S4, the frame filtering module performs a frame filtering operation based on the judgment result of the redundancy determination module. For images determined to be redundant frames, the frame filtering module performs a discard operation and simultaneously activates the timestamp and location information recording mechanism. When a frame is determined to be redundant, the system automatically extracts the timestamp information corresponding to that frame, which is accurate to the millisecond level, and simultaneously reads the current position coordinates of the UAV, including longitude, latitude, and relative altitude. The frame filtering module establishes a correlation between these timestamps and location information and the nearest key frame at that moment, linking the summary information of the redundant frames to the data structure of the key frames. This correlation storage mechanism ensures that even if a large number of redundant frames are discarded, the data throughout the flight process remains continuous and traceable in the time dimension. When it is necessary to trace back the monitoring data of a specific period, the user can quickly locate the data through the timestamp of the key frames and reconstruct the time range of the discarded redundant frames through the correlation information.
[0036] For images identified as keyframes, the frame filtering module performs a retention operation, transferring the complete data of the frame to the target recognition and storage module for further processing.
[0037] In step S5, the target recognition and storage module performs target recognition and storage processing on the keyframes. The preprocessing unit first performs denoising and image enhancement processing on the keyframes. The denoising process uses an adaptive filtering algorithm, which can dynamically adjust the filtering parameters according to the noise characteristics of local areas of the image. Specifically, the preprocessing unit analyzes the noise distribution characteristics of each local area, using a stronger filtering intensity when strong noise is detected and a weaker filtering intensity when weak noise is detected, thereby removing noise while preserving as much edge detail information as possible in the image. The image enhancement process uses a contrast adaptive adjustment algorithm, which analyzes the histogram distribution characteristics of each frame of the image and automatically adjusts the contrast enhancement parameters according to the distribution range and peak position of the histogram. After contrast enhancement processing, the target features of vehicles, pedestrians, etc. in the image are more prominent, providing more favorable input conditions for subsequent target recognition. The target detection unit uses a deep neural network architecture for target recognition. This network model uses a combination structure of convolutional layers, pooling layers, and fully connected layers, which can automatically learn the feature representations of various targets in traffic scenes. The target detection unit can simultaneously identify multiple vehicles, pedestrians, and traffic signs in an image, and output the category, location coordinates, and confidence score for each target. The storage unit adopts a structured data organization method, encapsulating the original image data of the keyframe, target recognition results, timestamps, and location information into complete JSON format data units, and writing them into high-speed solid-state storage media.
[0038] Through the verification of this embodiment, after adopting the data processing method described in this invention, the storage space occupied by UAV traffic monitoring data is reduced by about 52% compared with the traditional fixed frame rate method, while the recognition accuracy of key traffic event frames only decreases by 1.2%, which fully demonstrates the effectiveness of the method of this invention in significantly reducing storage space while ensuring target detection accuracy.
[0039] Example 2 This embodiment uses a power line inspection scenario as an example to illustrate the specific implementation of the present invention in a relatively stable background environment. In the power line inspection scenario, the drone needs to conduct inspection flights along transmission lines that are tens of kilometers long. The main objective of the inspection is to detect the operating status of key equipment such as towers, insulators, and conductors.
[0040] Since power transmission lines mostly run in straight lines, the background environment near the lines changes very slowly over a short period of time. Using traditional fixed frame rate acquisition methods would generate a large number of similar frames, resulting in significant waste of storage space. This embodiment addresses the specific characteristics of power line inspection scenarios by adjusting the system parameters accordingly.
[0041] Regarding the parameter configuration of the redundancy determination module, this embodiment adjusts the preset redundancy threshold from 0.3 for urban scenes to 0.25. This is because the background of power inspection scenes is relatively more stable, and the information content difference between consecutive frames is smaller. Therefore, a lower threshold is needed to avoid misjudging valid frames as redundant frames. The threshold dynamic adjustment module receives UAV flight altitude data. When the UAV flight altitude is 50 meters, the proportion of pixels occupied by ground objects in the image is relatively large, and the sensitivity to information entropy changes is high. The redundancy determination module automatically adjusts the threshold to 0.28. When the UAV flight altitude is 100 meters, the image content changes relatively smoothly, and the threshold automatically increases to 0.32. The camera device has a resolution of 20 megapixels, which contains more detailed information compared to a 12-megapixel camera device. The threshold dynamic adjustment module correspondingly lowers the baseline value of the redundancy threshold to adapt to the information content evaluation requirements of higher resolution images.
[0042] In this embodiment, the information entropy calculation module also employs sub-region division and weighted summation, but the weight settings for the sub-regions differ. In the power inspection scenario, regions containing transmission lines and towers are assigned higher weights; these regions are typically located near the center of the image. Regions containing distant mountains or the sky are assigned lower weights. The distribution statistics unit traverses the pixels of each sub-region and calculates the probability distribution of grayscale values. Based on this, the entropy calculation unit calculates the information entropy value of each sub-region and synthesizes the overall information entropy value.
[0043] The redundancy determination module calculates and compares the information entropy difference between consecutive frames. When the information entropy difference is less than the current redundancy threshold, the current frame is determined to be a redundant frame; when the information entropy difference is greater than or equal to the threshold, it is determined to be a key frame. In actual inspection, the UAV collects image data at a rate of 25 frames per second. After processing by the method of this invention, the retention rate of key frames is approximately 42%, and the storage amount is reduced by approximately 58% compared to the traditional method.
[0044] The target recognition and storage module performs target recognition processing on the retained keyframes. After the preprocessing unit performs noise reduction and contrast enhancement on the keyframes, the target detection unit uses a deep neural network model optimized for power equipment for target recognition. This model can accurately identify power equipment such as poles, insulators, conductors, and vibration dampers, and detect abnormalities such as damage, corrosion, and hanging foreign objects. The association and storage of recognition results with keyframes is also performed by the storage unit, which encapsulates the original data of each frame, the recognition result, the timestamp, and the UAV's position coordinates into a complete data unit.
[0045] Practical applications show that the method in this embodiment has not resulted in any missed defects in critical equipment during power line inspections, significantly improving inspection efficiency and data utilization, and effectively solving the prominent problem of limited storage space in long-distance UAV inspection missions.
[0046] Example 3 This embodiment uses a precision agriculture pest and disease monitoring scenario as an example to illustrate the specific implementation of the dynamic threshold adjustment strategy adopted by the present invention. In precision agriculture applications, drones need to monitor pests and diseases over large areas of farmland. The main objective of the monitoring is to promptly detect abnormal phenomena such as discoloration, curling, and insect spots on crop leaves.
[0047] While changes in farmland scenes may not be significant in a short period, the amount of information in the image increases dramatically once pests, diseases, or abnormal growth occur. Therefore, this embodiment adds a dynamic threshold adjustment mechanism based on changes in image content to the basic redundancy determination process.
[0048] In this embodiment, the threshold dynamic adjustment module receives not only UAV flight parameters as input but also the overall information entropy value of the current frame output by the information entropy calculation module. The threshold dynamic adjustment module maintains a sliding window containing historical information entropy values from the most recent frames. When a significant increase in the information entropy of a frame is detected compared to the previous frame, for example, if the increase exceeds a preset sensitivity threshold, the threshold dynamic adjustment module determines that a significant change may have occurred in the current scene and automatically lowers the redundancy threshold for subsequent frames. The magnitude of the threshold adjustment is dynamically determined based on the increase in information entropy; the larger the increase, the greater the threshold reduction. In this embodiment, when an increase in information entropy exceeding 0.4 is detected, the redundancy threshold is reduced from a baseline value of 0.3 to 0.2 to capture potential pest and disease anomalies; once the information entropy stabilizes, the redundancy threshold gradually returns to the baseline value.
[0049] The information entropy calculation module also uses a 16-sub-region division in the farmland scenario, but the weight settings for each sub-region have been optimized for agricultural monitoring. Sub-regions containing crop growth areas are assigned higher weights, while non-crop areas such as field ridges and roads are assigned lower weights. This differentiated weight setting ensures that the information entropy calculation can more accurately reflect changes in the amount of information about crop growth status.
[0050] The redundancy determination module performs redundancy determination based on a dynamically adjusted threshold. Frames with an information entropy difference lower than the current threshold are identified as redundant and discarded; frames with an information entropy difference higher than or equal to the threshold are retained as keyframes. The target recognition and storage module performs target recognition on the keyframes. The recognition model uses a deep neural network optimized for agricultural pest and disease detection, capable of identifying symptoms of dozens of common crop diseases and pests.
[0051] Test results show that the method in this embodiment reduces data storage by approximately 45% while maintaining the accuracy of pest and disease detection. The dynamic threshold adjustment mechanism enables the system to efficiently filter redundant frames when the farmland scene is normal and to capture key information in a timely manner when anomalies occur, effectively solving the contradiction between excessive data volume and limited storage space in UAV agricultural monitoring.
[0052] Example 4 This embodiment uses a mountain rescue scenario as an example to illustrate the specific implementation of the present invention in application scenarios with high real-time requirements. In mountain rescue scenarios, drones need to quickly search for missing persons under complex terrain conditions. At this time, the real-time requirements for target detection are extremely high, and any processing delay may lead to a delay in the rescue opportunity.
[0053] The information entropy calculation module uses the same sub-region division and weighted summation method as the aforementioned embodiments to calculate the information entropy value. In the mountain rescue scenario, the weight settings of the sub-regions have been adjusted according to the characteristics of the search task. Areas including mountain terrain, vegetation coverage, and areas where people may appear are given higher weights, while areas such as clouds and the sky are given lower weights.
[0054] The redundancy determination module uses a relatively low preset redundancy threshold of 0.25 to ensure that no keyframes containing missing persons are missed. The threshold dynamic adjustment module receives drone flight parameters and automatically lowers the redundancy threshold to 0.2 when it detects that the drone has entered an area where people may be, in order to capture more possible detection opportunities.
[0055] When discarding redundant frames, the frame filtering module also performs a mechanism for recording and associating timestamps and location information. This timestamp and location information provides reliable data support for subsequent rescue mission analysis and search route optimization.
[0056] The target recognition and storage module performs target recognition processing on keyframes. Due to the extremely high real-time requirements of mountain rescue scenarios, the preprocessing unit and target detection unit employ lightweight algorithms optimized for embedded platforms. The preprocessing unit uses a fast denoising algorithm with low computational cost, reducing computational overhead while maintaining basic denoising performance. The target detection unit uses a lightweight deep neural network model, which significantly reduces computational cost and memory usage while maintaining high detection accuracy.
[0057] By significantly reducing the number of redundant frames that need to be processed, the computational burden of the target detection algorithm is greatly reduced, and the system's response time is reduced from an average of 320 milliseconds in the traditional method to about 180 milliseconds, improving real-time performance by about 44%. This means that drones can respond to targets detected during the search process more quickly, providing more timely information support for mountain rescue operations.
[0058] Example 5 This embodiment uses a multi-source data fusion scenario as an example to illustrate the specific implementation of the method of the present invention in a UAV system that simultaneously carries multiple sensing devices. In comprehensive low-altitude missions, UAVs may simultaneously carry multiple sensing devices such as visible light cameras, infrared thermal imagers, and lidar. The image data collected by different sensors have different information entropy distribution characteristics, requiring differentiated processing strategies.
[0059] The information entropy calculation module sets up independent calculation channels for different sensors. The information entropy of RGB images acquired by visible light cameras is calculated using the aforementioned 16-sub-region weighted summation method; the information entropy of thermal images acquired by infrared thermal imagers is calculated using an 8-sub-region division method. This is because thermal images have a narrow grayscale range and relatively little detailed information, and fewer sub-region divisions can better capture changes in the overall amount of information.
[0060] The redundancy determination module maintains independent redundancy thresholds for data from different sensors. For image data acquired by a visible light camera, the preset redundancy threshold is set to 0.3; for image data acquired by an infrared thermal imager, since the temperature difference between the target and the background in thermal imaging images is usually more significant, the change pattern of information entropy is different from that in visible light images, so the preset redundancy threshold is set to 0.25. The threshold dynamic adjustment module independently adjusts the corresponding redundancy threshold according to the data characteristics of each sensor.
[0061] The frame filtering module performs redundancy detection and frame filtering operations on data from different sensors, retaining their respective keyframes. The target recognition and storage module performs target recognition on the keyframes retained by different sensors, with the recognition models for visible light and infrared images optimized for their respective data characteristics. The storage unit associates and stores the recognition results from different sensors with their corresponding keyframes, forming a multimodal data structure to facilitate subsequent information fusion and analysis.
[0062] Experiments show that the method in this embodiment can still effectively reduce data storage by about 50% in multi-source data fusion scenarios, while maintaining the accuracy of multimodal target detection. Redundancy determination of different sensor data is independent and does not interfere with each other, ensuring that all types of sensor data receive optimal processing. This multi-source data collaborative processing capability provides an efficient data processing solution for integrated UAV applications, significantly enhancing the system's adaptability to complex tasks.
[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A data processing method for low-altitude economic unmanned aerial vehicles (UAVs), characterized in that: Includes the following steps: S1. Continuously acquire multi-frame image data output by the drone's camera device; S2. Calculate the information entropy value of each frame in the multi-frame image data; the step of calculating the information entropy value includes: dividing each frame image into multiple sub-regions, statistically analyzing the pixel gray value distribution probability of each sub-region, calculating the information entropy value of each sub-region based on the pixel gray value distribution probability, and weighting and summing the information entropy values of each sub-region to obtain the overall information entropy value of the frame image. S3. Redundancy determination is performed based on the information entropy difference between adjacent frames. The redundancy determination steps include: calculating the information entropy difference between the current frame and the previous frame, comparing the information entropy difference with a preset redundancy threshold, determining the current frame as a redundant frame when the information entropy difference is less than the redundancy threshold, and determining the current frame as a key frame when the information entropy difference is greater than or equal to the redundancy threshold. S4. Based on the redundancy determination result, discard redundant frames with an information entropy difference lower than a preset threshold, and retain key frames with an information entropy difference higher than or equal to the preset threshold; the step of discarding redundant frames further includes: recording the timestamp and location information of the redundant frame, and storing the timestamp and location information in association with adjacent key frames. S5. Perform target recognition and storage processing on the key frame; the steps of target recognition and storage processing include: preprocessing the key frame, performing target recognition on the key frame based on a preset target detection model, and associating and storing the recognition result with the corresponding key frame.
2. The low-altitude economic unmanned aerial vehicle (UAV) data processing method according to claim 1, characterized in that: The calculation of the information entropy value in step S2 includes: S21. Divide each frame of image into multiple sub-regions; S22. Calculate the probability distribution of pixel grayscale values for each sub-region; S23. Calculate the information entropy value of each sub-region based on the probability distribution of pixel grayscale values; S24. The information entropy values of each sub-region are weighted and summed to obtain the overall information entropy value of the frame image.
3. The low-altitude economic unmanned aerial vehicle (UAV) data processing method according to claim 1, characterized in that: The specific steps of the redundancy determination in step S3 include: S31. Calculate the information entropy difference between the current frame and the previous frame; S32. Compare the information entropy difference with a preset redundancy threshold; S33. When the information entropy difference is less than the redundancy threshold, the current frame is determined to be a redundant frame; S34. When the information entropy difference is greater than or equal to the redundancy threshold, the current frame is determined to be a key frame. The preset redundancy threshold is dynamically adjusted according to the drone's flight altitude, camera resolution, and target detection accuracy requirements. When the drone's flight altitude decreases, the redundancy threshold decreases accordingly; when the drone's flight altitude increases, the redundancy threshold increases accordingly.
4. The low-altitude economic unmanned aerial vehicle (UAV) data processing method according to claim 3, characterized in that: The preset redundancy threshold is dynamically adjusted based on the UAV's flight altitude, camera resolution, and target detection accuracy requirements. When the resolution of the camera device is increased, the redundancy threshold is adjusted accordingly to adapt to the information content assessment at higher resolutions.
5. The low-altitude economic unmanned aerial vehicle (UAV) data processing method according to claim 1, characterized in that: Also includes: When discarding redundant frames, record the timestamp and location information of the redundant frame; The timestamp and location information are associated with and stored with adjacent keyframes to support subsequent data backtracking and integrity verification.
6. The low-altitude economic unmanned aerial vehicle (UAV) data processing method according to claim 1, characterized in that: The target identification and storage processing of the keyframes described in step S5 includes: S51. Preprocess the keyframes, including denoising and image enhancement; S52. Perform target recognition on keyframes based on a preset target detection model; S53. Associate the recognition results with the corresponding keyframes and store them in the data storage unit; The preprocessing steps include denoising and image enhancement. The denoising process uses an adaptive filtering algorithm to dynamically adjust the filtering parameters based on the noise characteristics of local areas of the image. The image enhancement process uses a contrast adaptive adjustment algorithm to enhance the contrast of each frame of the image based on the image histogram distribution. The target detection model adopts a deep neural network architecture, which can automatically learn the feature representation of the target and classify and locate multiple targets in the image.
7. A low-altitude economic unmanned aerial vehicle (UAV) data processing system, applied to the low-altitude economic UAV data processing method according to any one of claims 1 to 6, characterized in that: include: The image acquisition module is used to continuously acquire multiple frames of image data output by the drone's camera device; The information entropy calculation module is used to calculate the information entropy value of each frame in the multi-frame image data; The information entropy calculation module includes: a sub-region division unit, used to divide each frame of image into multiple sub-regions; a distribution statistics unit, used to calculate the distribution probability of pixel grayscale values in each sub-region; and an entropy calculation unit, used to calculate the information entropy value of each sub-region based on the pixel grayscale value distribution probability and synthesize the overall information entropy value. The redundancy determination module is used to determine redundancy based on the information entropy difference between adjacent frames. The frame filtering module is used to discard redundant frames and retain key frames according to the redundancy determination result; the frame filtering module is also used to record the timestamp and location information of the redundant frame when discarding the redundant frame, and to associate and store the timestamp and location information with the adjacent key frames. The target recognition and storage module is used to perform target recognition and storage processing on the keyframes.
8. The low-altitude economic unmanned aerial vehicle (UAV) data processing system according to claim 7, characterized in that: The information entropy calculation module includes: Sub-region division unit, used to divide each frame of image into multiple sub-regions; The distribution statistics unit is used to separately calculate the probability distribution of pixel grayscale values in each sub-region; The entropy calculation unit is used to calculate the information entropy value of each sub-region based on the probability distribution of the pixel grayscale value and to synthesize the overall information entropy value.
9. The low-altitude economic unmanned aerial vehicle (UAV) data processing system according to claim 7, characterized in that: Also includes: The threshold dynamic adjustment module is used to dynamically adjust the preset redundancy threshold used in the redundancy determination module according to the UAV flight parameters and environmental conditions.
10. The low-altitude economic unmanned aerial vehicle (UAV) data processing system according to claim 7, characterized in that: The target recognition and storage module includes: a preprocessing unit, used for denoising and image enhancement processing of keyframes; The target detection unit is used to identify targets in keyframes based on a preset target detection model. The storage unit is used to associate and store the recognition results with the corresponding keyframes.
