A method and system for identifying unmanned aerial vehicle targets at low altitudes in cities

Abstract

The present invention relates to the field of computer vision technology and discloses a method and system for target recognition of unmanned aerial vehicles (UAVs) at low altitudes in cities. The method comprises: collecting recognition data through a sensor network, performing target detection based on the recognition data, and outputting a master control decision based on the target detection result; calculating the interference level of target detection in real time, initiating auxiliary target recognition if the interference level exceeds a threshold, acquiring auxiliary data to supplement the recognition data, and performing heat map presence detection and lattice deformation detection based on the supplemented recognition data; outputting a perception correction based on the heat map presence detection and lattice deformation detection to cover and execute the conflicting parts of the master control decision, thereby achieving UAV target recognition in high-interference environments. The method of the present invention can significantly improve the perception robustness of UAVs at low altitudes in cities in non-ideal environments such as rain, fog, and backlight, reduce missed detections and misjudgments, avoid collision risks, and enhance the overall stability and security of the system.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method and system for identifying unmanned aerial vehicle (UAV) targets at low altitudes in cities. Background Art

[0002] As urban low-altitude airspace management policies gradually open up, multi-rotor drones (UAVs) for tasks such as logistics distribution, basic inspections, and emergency response are rapidly entering urban core areas. However, the complex structure of urban low-altitude environments is characterized by numerous high-frequency interference sources, such as glass curtain wall reflections, backlight shadows, rain and fog, and electromagnetic obstruction. This places higher demands on the real-time and robustness of UAV target recognition and obstacle avoidance systems. Existing recognition systems based on single visual or infrared sensors are prone to problems such as reduced recognition accuracy, target misjudgment, and obstacle avoidance failures under high-interference conditions, seriously impacting flight safety and mission stability.

[0003] To address these challenges, some systems have attempted to employ multimodal fusion or end-cloud collaborative recognition solutions, but these often suffer from heavy computational overhead, significant response delays, and unsuitability for lightweight deployment. Therefore, a drone recognition method that balances recognition accuracy with real-time responsiveness is urgently needed. This method, without relying on complex or heavy modeling, can dynamically perceive environmental interference and flexibly modify the primary recognition path, thereby improving drone recognition and decision-making capabilities in low-altitude urban environments. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to improve the target presence perception capability of urban low-altitude drones in complex visual interference environments such as rain, fog, and backlight.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a method for identifying targets of unmanned aerial vehicles (UAVs) at low altitudes in cities, comprising: collecting identification data through a sensor network, performing target detection based on the identification data, and outputting a master control decision based on the target detection results;

[0007] Calculate the interference degree of target detection in real time. If the interference degree exceeds the threshold, start auxiliary target recognition, obtain auxiliary data to supplement the recognition data, and perform heat map presence detection and lattice deformation detection based on the supplemented recognition data;

[0008] Based on the output of thermal image presence detection and lattice deformation detection, perception correction is implemented to cover and execute the conflicting parts of the main control decision, thus realizing UAV target recognition in high-interference environments.

[0009] As a preferred solution of the method for identifying unmanned aerial vehicle targets at low altitudes in cities described in the present invention, the identification data includes primary visual camera data, posture data, positioning data, and auxiliary data;

[0010] Auxiliary data includes infrared thermal imaging data and structured light projection data. The sensor network used to collect auxiliary data is initially in standby mode, only performing device power-on and handshake confirmation, and no data collection tasks. Auxiliary data collection will only begin after auxiliary target recognition is started.

[0011] As a preferred embodiment of the method for identifying urban low-altitude drone targets according to the present invention, the target detection according to the identification data includes preprocessing the primary visual camera data, including normalization, scaling, and color format conversion, matching the input requirements of the target detection model, and generating an input image;

[0012] Call the target detection model deployed on the drone to infer the input image and output the target bounding box, category label, and confidence score of each detection result;

[0013] The output results of the target detection model are processed by non-maximum suppression to remove redundant target boxes, and are filtered according to the confidence threshold to output the target detection results.

[0014] As a preferred solution of the method for identifying drone targets at low altitudes in cities described in the present invention, the calculation of the interference degree of target detection includes calculating the interference degree of the input image affected by the external environment while detecting the target, and the interference degree includes three indicators: image contrast interference degree, backlight interference degree and dark area coverage.

[0015] As a preferred solution of the method for identifying drone targets at low altitudes in cities described in the present invention, the heat map presence detection includes preprocessing infrared thermal imaging data to obtain a grayscale heat map I t In order to adapt to the complex and changeable light and heat conditions in the urban low altitude, a dynamic threshold is used instead of a fixed threshold. The dynamic threshold is calculated based on the intra-frame statistics:

[0016] T t =μ t +k·σ t ;

[0017] Among them, T t represents the dynamic threshold; μ t represents the average grayscale; k represents the adjustment coefficient; σ t represents standard deviation; t represents time;

[0018] According to the dynamic threshold T t Will I tConvert to a binary mask, perform a 3×3 corrosion and a 5×5 expansion operation in sequence, extract all connected domains {C i}, calculate area A i and the minimum bounding rectangle height h i ;

[0019] Since infrared thermal imaging is usually a monocular system and cannot directly obtain depth information, the target height inversion method based on projection geometry is used to estimate the distance to close-range targets. According to the pinhole imaging relationship of the thermal imager and the minimum target height calibrated at the factory, the distance from the center of the connected domain to the camera is calculated:

[0020] ;

[0021] Among them, d i represents the distance from the center of the connected domain to the camera; f represents the focal length pixel value of the thermal imager; H represents the minimum target height; h i Indicates the minimum circumscribed rectangle height; i indicates the i-th point;

[0022] Based on the hot spot area and image ratio, the detected hot spot is double-judged. i >α×A f and d i <d safe When , it is marked as risky, where α represents the parameter, A f Represents the area of ​​the entire heat map frame, d safe Indicates the safety distance, output hot spot range and viewing direction of the hot plate center area.

[0023] As a preferred solution of the method for identifying drone targets at low altitudes in cities described in the present invention, the lattice deformation detection includes locating sub-pixel coordinates (x i ,y i ), calculate the expected dot matrix template in real time based on the posture data ;

[0024] The changes in the structured light dot matrix are judged by the dot matrix regularity and occlusion rate. The dot matrix regularity is used to measure the degree of matching between the dot matrix extracted in the current frame and the expected dot matrix template, which is expressed as:

[0025] ;

[0026] Where R represents the dot matrix regularity index; N represents the number of points in the structured light projection data that have been identified and matched with the points in the expected dot matrix template; d0 represents the theoretical average spacing between adjacent points in the expected dot matrix template; ‖·‖2 represents the calculated Euclidean distance;

[0027] The occlusion rate is used to measure the difference between the number of visible points and the theoretical expected number, which is expressed as:

[0028] ;

[0029] Among them, O represents the occlusion rate index; N vis Indicates the number of points actually detected in the current frame; N exp Indicates the number of points that should be projected and visible in theory;

[0030] A dual-threshold risk assessment is performed based on the lattice regularity and occlusion rate. When lattice deformation is detected, the deformation range and the viewing direction of the deformation center area are output.

[0031] As a preferred solution of the method for UAV target recognition at low altitude in cities described in the present invention, the perception correction includes judging the consistency of the target detection result and the auxiliary target recognition result. If there is an recognition conflict, the target detection result is corrected, and a corrected recognition map is generated to correct the main control decision based on the hot spot range and the viewing angle direction of the central area of ​​the hot plate output by the thermal map detection output, as well as the deformation range and the viewing angle direction of the deformation center area output by the lattice deformation detection output.

[0032] A target recognition system for unmanned aerial vehicles (UAVs) at low altitudes in cities using any of the methods described in the present invention, wherein: a recognition module collects recognition data through a sensor network, performs target detection based on the recognition data, and outputs a master control decision based on the target detection result;

[0033] The auxiliary module calculates the interference degree of target detection in real time. If the interference degree exceeds the threshold, auxiliary target recognition is started, and auxiliary data is obtained to supplement the recognition data. The heat map presence detection and lattice deformation detection are performed based on the supplemented recognition data.

[0034] The correction module outputs perception correction based on heat map presence detection and lattice deformation detection, covers the conflicting part of the main control decision and executes it, realizing UAV target recognition in high interference environment.

[0035] Beneficial Effects of the Invention: This invention provides a method for urban low-altitude UAV target recognition that integrates interference assessment, thermal image detection, structured light lattice deformation analysis, and recognition correction mechanisms. This method dynamically activates auxiliary channels when significant interference is detected, performs local presence detection and spatial perception correction, and then integrates and optimizes the primary recognition results. This method is computationally lightweight, flexible in deployment, and responsive, making it particularly suitable for flight obstacle avoidance and target recognition tasks in complex urban low-altitude environments. While improving recognition robustness, it effectively ensures UAV flight safety and mission continuity. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 An overall flow chart of a method for identifying unmanned aerial vehicle targets at low altitudes in cities, provided as an embodiment of the present invention. DETAILED DESCRIPTION

[0038] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0039] Example 1, reference Figure 1 , as one embodiment of the present invention, provides a method for identifying unmanned aerial vehicle targets at low altitudes in cities, comprising:

[0040] S1: Collect identification data through the sensor network, perform target detection based on the identification data, and output the main control decision based on the target detection results.

[0041] In order to clearly illustrate the beneficial effects of the method of the present invention on target recognition, this embodiment is explained using the example of automatic obstacle avoidance during drone flight. That is, the main control decision is the automatic obstacle avoidance decision and path made by the drone's main recognition module (corresponding to the auxiliary recognition module proposed in the present invention) after detecting an obstacle.

[0042] Furthermore, the present invention provides an auxiliary recognition module, which consists of an infrared thermal imaging submodule and a structured light projection submodule. During system initialization, this module completes basic communication handshakes and power consumption configuration, but remains in standby mode and does not collect data. It activates collection tasks only when the main recognition process enters a state of interference (such as when image contrast decreases or target confidence falls below a threshold), thus achieving conditionally triggered auxiliary recognition without increasing additional energy consumption.

[0043] Upon system startup, the primary visual acquisition unit is initialized. This involves activating the RGB camera and loading the camera's intrinsic parameter file. This includes performing camera distortion correction, adaptive white balance adjustment, and exposure control to ensure good brightness contrast and geometric consistency in the captured image. Simultaneously, the drone's inertial measurement unit and GNSS module are activated and synchronized to collect current flight attitude data and absolute position information for subsequent motion compensation and flight path estimation in image recognition.

[0044] Furthermore, in one embodiment of the present invention, the primary recognition module is used by a low-altitude urban drone to perform real-time obstacle detection and avoidance during autonomous flight. The drone is equipped with a primary visual image acquisition module (e.g., a 1080p high-definition RGB camera), which captures images of the scene ahead in real time at a constant frame rate (e.g., 30 frames per second).

[0045] The image is input to the object detection module deployed in the drone's processing unit. This module uses lightweight object detection models for fast inference, such as YOLOv5-Nano, YOLOv7-Tiny, and MobileNet-SSD. Models can be loaded statically or dynamically. During inference, the image is first normalized and resized before entering the neural network backbone for feature extraction.

[0046] The model output includes the bounding box coordinates of multiple candidate objects, their category labels (such as pedestrians, vehicles, utility poles, buildings, billboards, etc.), and their corresponding confidence scores. The system uses non-maximum suppression to filter out overlapping areas, remove redundant objects, and retain only valid detections with a confidence score greater than a set threshold.

[0047] Subsequently, the system spatially aligns the identified obstacle target with the current drone posture and flight path, predicts the target's relative position and distance, and generates flight control decisions based on its occupancy on the flight path, including deceleration, heading adjustment, lateral detour, or hovering and waiting, and sends them to the flight control system to avoid potential collision risks.

[0048] The main recognition module in this embodiment serves as a conventional obstacle avoidance detection means and does not participate in image interference judgment and presence detection logic. It is only used for conventional flight recognition tasks in a clear environment.

[0049] Furthermore, after the main recognition module completes target detection, it transmits the results to the main control module for obstacle avoidance navigation. This module uses the bounding box position, category, and confidence level of each target in the recognition output, combined with flight control parameters such as the drone's current position, heading angle, and flight speed, to perform real-time path adjustment calculations.

[0050] Specifically, the system first calculates the spatial position coordinates of the target relative to the drone based on the target position and camera field of view parameters in the recognition results using geometric projection. The system can also call binocular structured light depth or IMU height data to compensate for and estimate the distance to major obstacles.

[0051] Once the detected target is determined to be an obstacle that "may block the flight path" (for example, the center offset does not exceed the set angle threshold, and the distance is less than the safety tolerance), the system will generate corresponding control instructions based on the preset obstacle avoidance strategy, such as:

[0052] Reduce flight speed;

[0053] Fine-tune the heading angle laterally to avoid the center of the obstacle;

[0054] If an obstacle is detected approaching for multiple frames, the aircraft will enter the "Hover Obstacle Avoidance" mode and wait for manual confirmation or the risk to be resolved.

[0055] Based on the above control decisions, the flight control module outputs corresponding motor drive instructions, adjusts the flight attitude in real time, and achieves autonomous dynamic obstacle avoidance.

[0056] This step serves as a response mechanism for the main recognition result and can be run independently in a normal visual environment; and when the subsequent auxiliary module (corresponding to S2 in this embodiment) is started, the obstacle avoidance control instruction can also be dynamically overwritten and corrected according to the auxiliary recognition result.

[0057] S2: Calculate the interference degree of target detection in real time. If the interference degree exceeds the threshold, start auxiliary target recognition, obtain auxiliary data and supplement it to the recognition data, and perform heat map presence detection and lattice deformation detection based on the supplemented recognition data.

[0058] Furthermore, during the execution of the main recognition module, the quality of the primary visual image is continuously evaluated to determine whether there are interference factors that hinder the accuracy of target recognition. If interference is determined to exist, the auxiliary recognition module is immediately started to enhance the robustness of the system.

[0059] Specifically, the interference degree includes three indicators: image contrast interference degree, backlight interference degree and dark area coverage degree.

[0060] The image contrast interference index reflects the richness of image texture and edge levels by calculating the standard deviation of the overall image grayscale value. When an image is affected by rain, fog, blur, or motion smear, its grayscale distribution tends to smooth out and contrast decreases, resulting in a significant decrease in the image contrast interference index.

[0061] Backlight interference is used to detect strong non-uniform lighting in the image, such as backlighting or specular reflections that cause extreme brightness on one side and darkness on the other. The system divides the image into symmetrical regions (left and right or top and bottom), calculates the average brightness of each, and calculates the relative deviation ratio.

[0062] Dark area coverage reflects the proportion of low-brightness areas in the entire image. The system counts the number of pixels with brightness below a set threshold (e.g., 40) and calculates their percentage relative to the entire image. If this percentage exceeds the set threshold, the image is too dark, potentially leading to recognition failure.

[0063] These three indicators can trigger the activation of the auxiliary recognition module individually or in combination. By comparing the real-time calculated interference level with a preset fixed threshold, the interference level of the image can be quickly determined. If interference is detected, the activation signal of the infrared thermal imaging and structured light modules is triggered, and the auxiliary recognition path proposed by this invention is entered.

[0064] Through this step, the system can independently judge the image environment quality without relying on the confidence of the model output, effectively adapt to common image interference scenarios in urban low altitudes such as rainy and foggy weather, backlit shooting, and low light, provide an automatic startup basis for the subsequent auxiliary presence detection module, and improve the reliability and stability of the overall recognition system.

[0065] Furthermore, thermal image presence detection uses infrared thermal imaging sensors to capture thermal radiation images of the scene ahead, determining whether objects with unusual thermal characteristics are present. This feature is primarily used to address visually blurred scenes such as rain and fog, low-light conditions, and low-contrast textured backgrounds. Under these conditions, traditional visible light images struggle to provide clear outlines, while infrared thermal images can still detect heat sources. This effectively fills in the blind spots of the primary recognition path and improves overall recognition stability.

[0066] Specifically, the thermal image presence detection submodule is used to quickly determine whether there are obstacles to be avoided in the direction of flight in the infrared channel without identifying the category of the obstacle. The system uses the LWIR thermal imager to obtain grayscale thermal images at 25fps. t To eliminate noise and enhance contrast, a 3×3 Gaussian filter is used for denoising, followed by global histogram equalization. The preprocessed heatmap maintains its original resolution of 320×240 to reduce the computational complexity on the client side.

[0067] Traditional methods often use a preset fixed threshold to segment thermal images. However, in complex low-altitude urban environments, thermal imaging data is significantly affected by factors such as weather (such as rain and fog) and backlight, and image brightness and noise distribution vary significantly. Fixed thresholds often cannot adapt to different scenarios and can easily lead to missed detections or false alarms. Therefore, this invention uses a dynamic threshold instead of a fixed threshold. The dynamic threshold is calculated based on intra-frame statistics:

[0068] T t =μ t +k·σ t ;

[0069] Among them, T t represents the dynamic threshold; μ t represents the average grayscale; k represents the adjustment coefficient; σ t This adaptive method can flexibly adjust the threshold based on the current image quality (such as the overall brightness and noise level under environmental interference), preventing the fixed threshold from failing in different environments, thereby improving detection accuracy.

[0070] According to the dynamic threshold T t Will I t Convert it into a binary mask and perform a 3×3 erosion and a 5×5 dilation operation in sequence to remove isolated noise points and close small holes to ensure the continuity of the connected domain contour.

[0071] Use the four-connected labeling algorithm to extract all connected components {C i}, for each connected domain, calculate the area A i and the minimum bounding rectangle height h i , the area is used to reflect the target size, h i Used for distance estimation.

[0072] Since infrared thermal imaging is usually a monocular system and cannot directly obtain depth information, the target height inversion method based on projection geometry is used to estimate the distance to close-range targets. According to the pinhole imaging relationship of the thermal imager and the minimum target height calibrated at the factory, the distance from the center of the connected domain to the camera is calculated:

[0073] ;

[0074] Among them, d i represents the distance from the center of the connected domain to the camera; f represents the focal length pixel value of the thermal imager; H represents the minimum target height; h i Represents the minimum enclosing rectangle height; i represents the i-th point.

[0075] Based on the hot spot area and image ratio, the detected hot spot is double-judged. i >α×A f and d i <d safe When , it is marked as risky, where α represents the parameter, A f Represents the area of ​​the entire heat map frame, d safeIndicates the safe distance and outputs the hot spot range and viewing direction of the hot plate center area. Detected hot spots are dual-judged, considering both the hot spot area and image ratio and estimating the actual distance, filtering out distant noise or non-obstructive heat sources caused by environmental interference.

[0076] To suppress occasional false alarms, the system performs a logical OR operation on the risk markers within a sliding window of length M (preferably 5) frames. If at least one frame in the window is marked as 1 (there is a risk), the final output risk marker is 1.

[0077] This invention not only overcomes the instability of conventional infrared detection in dynamic, high-interference urban low-altitude environments due to fixed parameters and single-frame judgment, but also achieves millisecond-level processing without significantly increasing computational complexity, fully meeting the needs of drones for real-time obstacle avoidance and mission execution. This innovative solution improves the system's overall resistance to environmental interference and operational safety while maintaining low power consumption and a lightweight design.

[0078] Furthermore, dot matrix deformation detection involves forming a regular dot matrix pattern in the forward area using a structured light projection device and capturing the reflected image in real time using infrared imaging equipment to determine whether the dot matrix's geometric structure has changed due to occlusion, deformation, or perspective distortion. This is primarily used to address non-thermal target scenarios such as backlighting, high reflections, and transparent occlusions. In such environments, thermal maps may not be able to identify low-reflectivity obstacles. However, the dot matrix distortion feature can quickly determine whether objects are intervening in the light path, enabling the detection of difficult-to-identify obstacles such as glass and plastic. It is an effective geometric supplement to thermal map detection.

[0079] Unlike traditional structured light recognition methods that primarily rely on fixed template matching, point counting, or deep modeling, this paper proposes a structured light dot matrix deformation detection method suitable for urban low-altitude flight environments. This method combines dynamic posture mapping, joint analysis of regularity and occlusion rate, and a temporal consistency mechanism to achieve highly sensitive presence detection of non-thermal obstacles within 3 meters. Without relying on classification or reconstruction, the system focuses solely on the presence or absence of occlusion, significantly reducing the computational burden and enhancing the ability to supplement target recognition in complex environments. This overcomes the limitations of thermal map methods in situations with low thermal reflection or geometric occlusion.

[0080] Specifically, the structured light projection submodule automatically lights up when interference is detected, and the infrared camera collects dot matrix images at 40fps. To suppress the impact of ambient light changes, the dot matrix image is subjected to variable gain adjustment and short exposure fusion to highlight the grayscale peak of the projection point.

[0081] The Difference-of-Gaussian (DoG) operator is used to preliminarily screen candidate bright spots, and then the sub-pixel coordinates (x i,y i ), calculate the expected dot matrix template in real time based on the current posture (pitch, roll, heading angle) of the drone and the structured light geometric model ,It is expected that the lattice template will be dynamically updated with ,random body motion to solve the problem of fixed template ,misalignment due to viewing angle changes.

[0082] The changes in the structured light dot matrix are judged by the dot matrix regularity and occlusion rate. The dot matrix regularity is used to measure the degree of matching between the dot matrix extracted in the current frame and the expected dot matrix template, which is expressed as:

[0083] ;

[0084] Here, R represents the dot matrix regularity index; N represents the number of points identified in the structured light projection data and matched to points in the desired dot matrix template; d0 represents the theoretical average spacing between adjacent points in the desired dot matrix template; and ‖·‖2 represents the calculated Euclidean distance. R = 1 indicates no distortion; smaller R indicates greater distortion. This index accounts for translational, rotational, and scaling distortion.

[0085] The occlusion rate is used to measure the difference between the number of visible points and the theoretical expected number, which is expressed as:

[0086] ;

[0087] Among them, O represents the occlusion rate index; N vis Indicates the number of points actually detected in the current frame; N exp Indicates the number of points that should be projected and visible theoretically. If there is an object blocking the view, part of the dot matrix will be covered, causing O to increase.

[0088] A dual-threshold risk assessment is performed based on the lattice regularity and occlusion rate. When lattice deformation is detected, the deformation range and the viewing direction of the deformation center area are output.

[0089] The structured light dot matrix deformation detection submodule of this invention can complete a frame analysis within 10ms, accurately detecting close-range occlusions or localized optical path disturbances. Compared to detection relying solely on thermal images, this solution is more sensitive to non-thermal obstacles (such as transparent glass and low-reflective plastics) in backlit scenes. It also provides supplementary geometric verification in rainy and foggy scenes, significantly improving the overall reliability of drone obstacle avoidance.

[0090] S3: Based on the output of thermal image presence detection and lattice deformation detection, perception correction is performed to cover and execute the conflicting parts of the main control decision, realizing drone target recognition in high-interference environments.

[0091] Furthermore, during flight, the drone uses a vision-based target detection method to identify obstacles in its flight path, with the results serving as the basis for master control decisions. If the system detects interference risk in the current frame or detects a decrease in the stability of the detection results, it automatically activates the perception correction fusion mechanism proposed in this invention.

[0092] The system receives auxiliary recognition results from the heat map presence detection module and the lattice deformation detection module, including a Boolean flag indicating the presence of an obstacle and its location. The system performs a consistency analysis between the obstacle map from the primary recognition module and the output from the auxiliary module. If an area ahead is identified as passable by the primary recognition module but is repeatedly identified as an obstacle by the auxiliary recognition module, the system determines that there is an identification conflict in that area.

[0093] Based on this, the correction module marks the conflict area in the main recognition result as a "corrected risk area" and generates a corrected obstacle map. Simultaneously, based on the risk angle and recognition time window, the correction module calculates a response strategy, including temporary yaw, trajectory shift, and flight speed adjustment, and transmits the correction results to the main control decision execution module.

[0094] The master control decision execution module receives the revised obstacle map and flight control recommendations and performs a risk review of the current flight plan and target waypoint path. If a high-priority obstacle avoidance requirement exists, the system will temporarily interrupt the original flight path and execute risk avoidance instructions, prioritizing flight safety. After multiple consecutive frames of risk-free input, the system can resume the original mission path.

[0095] Through the perception correction fusion and main control execution strategy proposed in the present invention, the accuracy and stability of the drone's obstacle avoidance behavior can be effectively improved in complex environments such as dense urban buildings, backlight reflection, rain and fog interference, and the risk of collision caused by main recognition failure can be avoided.

[0096] Example 2: In another embodiment of the present invention, the main recognition module is configured to enable the urban low-altitude mission-type UAV to identify preset specific targets during flight, and guide the UAV to complete the flight to the target location and automatic landing mission.

[0097] Furthermore, the drone is equipped with a primary visual image acquisition module (a high-definition RGB monocular camera) that continuously captures images of the area ahead and feeds them to an object detection system deployed in an onboard edge processing unit. This object detection module uses detection models adapted for specific tasks, such as those trained to recognize logistics delivery signs, pattern identification plates, people wearing designated clothing colors, and unmanned vehicle platforms.

[0098] The image input process is the same as for general obstacle avoidance: first, the image is normalized and resized, then fed into the model for category-specific recognition. The model outputs a limited number of target categories, such as "receipt sign A," "person in a red vest," or "recycle bin icon." Each target result includes a bounding box, category ID, and confidence score.

[0099] Once a target's confidence level is detected above a mission threshold, the system projects its coordinates into three-dimensional space and, using the IMU and altimeter, estimates the target's position relative to the drone. The flight controller then issues a path adjustment based on the target's position, slowing the drone's flight toward the target and automatically entering a hovering state within a certain distance.

[0100] If the target is successfully locked and stable in space for several consecutive frames, the landing program is triggered, the vertical descent control logic is started, and the aircraft automatically flies to the top of the target and completes the soft landing.

[0101] The main recognition module in this embodiment also uses existing image detection technology and does not involve the auxiliary recognition structure of the present invention. It only serves as the main visual support module for designated target search and task positioning.

[0102] Furthermore, after the primary recognition module identifies a target, it uses the identified target location to guide the drone through navigation and automatic landing. Based on the target bounding box coordinates and confidence information output by the primary recognition result, combined with the current drone's attitude and positioning data, the system calculates the target's relative position in the world coordinate system and constructs a flight path adjustment vector.

[0103] Specifically, when the system detects a target of a specified category and its confidence exceeds a preset threshold, it will maintain a fixed lock on the target, track its spatial position through a sequence of image frames, and calculate the offset of the target center in the drone's field of view. If this offset exceeds a set tolerance, the system generates lateral or pitch adjustment commands to guide the drone slowly toward the target.

[0104] When the target is continuously detected with high confidence and maintains a stable spatial position for more than a set time window, and the estimated relative distance is less than the set threshold, the system enters the landing preparation state and sends a prepare-to-land command to the flight control system.

[0105] Subsequently, the flight control module executes a slow descent strategy, controlling the drone to descend in the vertical direction of the target, and completes the landing action while ensuring that the target is always in the center area of ​​the image.

[0106] If the main recognition module fails to identify the target or the recognition confidence drops sharply during the flight, the system will automatically interrupt the landing procedure and return to the waiting or hovering state to avoid landing failure or deviation due to incorrect identification or interference.

[0107] When image interference is severe, the auxiliary recognition module proposed in the present invention can also intervene in this link to perform confidence correction or action coverage on the main recognition result to ensure visual reliability and flight control safety during mission execution.

[0108] Furthermore, similar to Example 1, the results of heat map existence detection and dot matrix deformation detection are obtained through the auxiliary module proposed in the present invention as the basis for perception correction, so this part of the method will not be described here.

[0109] Furthermore, in this embodiment, the drone performs designated target recognition and landing tasks in urban scenes. The main recognition module recognizes specific ground landmarks (such as special patterns, lights, target platforms) based on visual feature extraction and matching algorithms, and outputs the recognition position and confidence level for landing guidance.

[0110] If the system determines that the current vision is significantly affected by environmental interference, or the confidence level of the recognition results fluctuates significantly across multiple consecutive frames, the system will automatically activate the Perception Correction Module. This module first collects the data output by the heat map presence detection and lattice deformation detection, and spatially matches them with the main recognition candidate area.

[0111] If the auxiliary recognition results show that the main recognition output location has obvious occlusion, low heat source coverage, or large dot matrix distortion, the correction module will downgrade the confidence of this location and re-evaluate the priority of the target candidate point based on the heat source concentration area or structured light non-occlusion area provided by the auxiliary module.

[0112] The correction module then outputs a revised target candidate map and transmits it to the main control decision execution module. The main control module determines the upcoming landing action. If there is a risk of occlusion in the current target area, the system will suspend the landing process and hover within a certain range. If new credible candidate areas are found in the auxiliary information, the landing point switch operation will be executed, guiding the drone to reposition to the target platform.

[0113] Through this embodiment, the present invention can maintain the stability of the task when target recognition is interfered with by the environment, avoid erroneous landing in non-designated areas, and effectively improve the reliability of drones performing tasks such as fixed-point delivery, platform take-off and landing at low altitude in cities.

[0114] Example 3. In an exemplary embodiment, a drone target recognition system for low altitude in cities is also provided, including a recognition module that collects recognition data through a sensor network, performs target detection based on the recognition data, and outputs a main control decision based on the target detection result.

[0115] The auxiliary module calculates the interference degree of target detection in real time. If the interference degree exceeds the threshold, auxiliary target recognition is started, and auxiliary data is obtained to supplement the recognition data. Heat map presence detection and lattice deformation detection are performed based on the supplemented recognition data.

[0116] The correction module outputs perception correction based on heat map presence detection and lattice deformation detection, covers the conflicting part of the main control decision and executes it, realizing UAV target recognition in high interference environment.

[0117] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0118] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0119] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0120] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying unmanned aerial vehicle targets at low altitudes in cities, characterized in that: include: Collect identification data through the sensor network, perform target detection based on the identification data, and output the main control decision based on the target detection results; Calculate the interference degree of target detection in real time. If the interference degree exceeds the threshold, start auxiliary target recognition, obtain auxiliary data to supplement the recognition data, and perform heat map presence detection and lattice deformation detection based on the supplemented recognition data; Based on the output of thermal image presence detection and lattice deformation detection, perception correction is implemented to cover and execute the conflicting parts of the main control decision, realizing drone target recognition in high-interference environments; The thermal image presence detection includes: pre-processing the infrared thermal imaging data to obtain a grayscale thermal image I t In order to adapt to the complex and changeable light and heat conditions in the urban low altitude, a dynamic threshold is used instead of a fixed threshold, and the dynamic threshold is calculated based on intra-frame statistics: T t =m t +k·s t ; Among them, T t represents the dynamic threshold; μ t represents the average grayscale; k represents the adjustment coefficient; σ t represents standard deviation; t represents time; According to the dynamic threshold T t Will I t Convert to a binary mask, perform a 3×3 corrosion and a 5×5 expansion operation in sequence, extract all connected domains {C i }, calculate area A i and the minimum bounding rectangle height h i ; The target height inversion method based on projection geometry is used to estimate the distance to close-range targets. The distance from the center of the connected domain to the camera is calculated based on the pinhole imaging relationship of the thermal imager and the minimum target height calibrated at the factory: ； Among them, d i represents the distance from the center of the connected domain to the camera; f represents the focal length pixel value of the thermal imager; H represents the minimum target height; h i Indicates the minimum circumscribed rectangle height; i indicates the i-th point; Based on the hot spot area and image ratio, the detected hot spot is double-judged. i >α×A f and d i <d safe When , it is marked as risky, where α represents the parameter, A f Represents the area of ​​the entire heat map frame, d safe Indicates the safety distance, output hot spot range and viewing direction of the hot plate center area; The dot matrix deformation detection includes: locating sub-pixel coordinates (x i ,y i ), calculate the expected dot matrix template in real time based on the posture data ; The changes in the structured light dot matrix are judged by the dot matrix regularity and occlusion rate. The dot matrix regularity is used to measure the degree of matching between the dot matrix extracted in the current frame and the expected dot matrix template, which is expressed as: ； Where R represents the dot matrix regularity index; N represents the number of points in the structured light projection data that have been identified and matched with the points in the expected dot matrix template; d0 represents the theoretical average spacing between adjacent points in the expected dot matrix template; ‖·‖2 represents the calculated Euclidean distance; The occlusion rate is used to measure the difference between the number of visible points and the theoretical expected number, which is expressed as: ； Among them, O represents the occlusion rate index; N vis Indicates the number of points actually detected in the current frame; N exp Indicates the number of points that should be projected and visible in theory; A dual-threshold risk assessment is performed based on the lattice regularity and occlusion rate. When lattice deformation is detected, the deformation range and the viewing direction of the deformation center area are output.

2. The method for identifying unmanned aerial vehicle targets at low altitudes in cities according to claim 1, wherein: The recognition data includes main visual camera data, posture data, positioning data and auxiliary data; Auxiliary data includes infrared thermal imaging data and structured light projection data. The sensor network used to collect auxiliary data is initially in standby mode, only performing device power-on and handshake confirmation, and no data collection tasks. Auxiliary data collection will only begin after auxiliary target recognition is started.

3. The method for identifying unmanned aerial vehicle targets at low altitudes in cities according to claim 2, wherein: The target detection according to the recognition data includes: preprocessing the primary visual camera data, including normalization, scaling and color format conversion, matching the input requirements of the target detection model, and generating an input image; Call the target detection model deployed on the drone to infer the input image and output the target bounding box, category label, and confidence score of each detection result; The output results of the target detection model are processed by non-maximum suppression to remove redundant target boxes, and are filtered according to the confidence threshold to output the target detection results.

4. The method for identifying unmanned aerial vehicle targets at low altitudes in cities according to claim 3, wherein: Calculating the interference degree of target detection includes: calculating the interference degree of the input image affected by the external environment during target detection, and the interference degree includes three indicators: image contrast interference degree, backlight interference degree and dark area coverage degree.

5. The method for identifying unmanned aerial vehicle targets at low altitudes in cities according to claim 4, wherein: The perception correction includes: judging the consistency of the target detection result and the auxiliary target recognition result. If there is an recognition conflict, the target detection result is corrected, and a corrected recognition map is generated to correct the main control decision based on the hot spot range and the viewing angle direction of the central area of ​​the hot plate output by the thermal map detection output, as well as the deformation range and the viewing angle direction of the deformation center area output by the lattice deformation detection output.

6. A UAV target recognition system for low-altitude urban areas, applied to a UAV target recognition method for low-altitude urban areas according to any one of claims 1 to 5, characterized in that: include: The recognition module collects recognition data through the sensor network, performs target detection based on the recognition data, and outputs the main control decision based on the target detection results; The auxiliary module calculates the interference degree of target detection in real time. If the interference degree exceeds the threshold, auxiliary target recognition is started, and auxiliary data is obtained to supplement the recognition data. The heat map presence detection and lattice deformation detection are performed based on the supplemented recognition data. The correction module outputs perception correction based on heat map presence detection and lattice deformation detection, covers the conflicting part of the main control decision and executes it, realizing UAV target recognition in high interference environment.

Images