An unmanned aerial vehicle emergency warning lighting method and system combined with an AI identification function

Abstract

The present application relates to the technical field of unmanned aerial vehicle emergency warning, and particularly relates to an unmanned aerial vehicle emergency warning lighting method combined with AI recognition function, which comprises the following steps: recognizing ground abnormal features; adjusting flight attitude; secondary hovering scanning; light source directional projection; and calculating warning level. The present application fuses the ground dynamic features monitored by the airborne visual sensor in real time with the target detection of the AI recognition model, dynamically adjusts the flight attitude according to the ground abnormal features, and performs secondary hovering scanning, so as to compensate for the recognition deviation caused by light changes or shielding. Then, the segmented lighting program is used to separate the environmental interference area from the target area, and the number of false positive targets is deducted from the recognition results, so as to obtain a more accurate emergency warning level. Through multi-source data fusion and physical environment compensation mechanism, the reliability of the unmanned aerial vehicle emergency warning in a complex scene is directly improved.

Description

Technical Field

[0001] This invention relates to the field of drone emergency early warning technology, and in particular to a drone emergency early warning lighting method and system that combines AI recognition function. Background Technology

[0002] With the widespread application of drone technology in emergency rescue, nighttime patrols, and security monitoring, the emergency warning capability of its onboard visual recognition system in complex lighting and dynamic scenarios has become a core bottleneck for performance improvement in this field. However, the drastic fluctuations in light intensity, the diversity of ground textures, and dynamic interference within the target area in actual operating environments constitute extremely non-uniform and randomly changing visual recognition scenarios. This poses a severe technical challenge to the recognition accuracy, response timeliness, and reliability of drone warning methods based on traditional image processing and fixed lighting strategies in complex scenarios.

[0003] Chinese Patent Publication No. CN112633326A discloses a method for UAV target detection based on Bayesian multi-source fusion. The method includes the following steps: S1: Target detection is performed on the optical image of the detection area, and a first vectorized result is obtained; S2: Target detection is performed on the communication signal of the detection area, and a second vectorized result is obtained; S3: The first vectorized result and the second vectorized result are used as inputs to a Bayesian function, respectively, to perform multi-source fusion and complete the UAV target detection.

[0004] Therefore, the existing technology has the following problems: it relies on a target detection network with fixed parameters and does not dynamically adjust the recognition strategy according to changes in actual scene lighting or environmental interference, resulting in a significant decrease in recognition rate under drastic lighting changes or low light conditions at night; although its multi-source fusion strategy uses confidence for fusion decision-making, it is mainly used to improve target positioning accuracy and does not establish a closed-loop feedback adjustment mechanism for dynamic environmental interference, making it difficult to effectively deal with false alarms caused by dynamic ground interference; its early warning decision relies on single or static fusion recognition results and lacks a secondary confirmation and interference deduction mechanism based on real-time sensor feedback, resulting in insufficient early warning reliability. Summary of the Invention

[0005] To address this, the present invention provides an emergency warning lighting method and system for drones that incorporates AI recognition capabilities. This system overcomes the problems of poor recognition accuracy and insufficient warning reliability caused by changes in illumination and ground interference in the prior art by monitoring ground image features in real time and dynamically adjusting flight attitude and lighting strategies.

[0006] To achieve the above objectives, on the one hand, the present invention provides an emergency early warning lighting method for drones that incorporates AI recognition functionality, comprising: Ground anomaly features are identified based on ground image sequences, where the ground image sequences are determined based on ground images collected by the onboard visual sensor of the monitoring UAV during its cruise. The flight attitude is adjusted based on the ground anomaly features and anomaly coverage area, wherein the anomaly coverage area is obtained based on the record after stopping cruise, and stopping cruise is determined based on whether the identification status within the expected identification time period is stable. The UAV is controlled to hover over the target area for secondary hovering scanning, and a segmented lighting program is set based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment to perform segmented lighting projection, thereby obtaining the environmental interference zone and the target warning zone. AI recognition models are used to detect false alarm signals in the environmental interference zone and target warning zone respectively, in order to obtain the confidence level of target features in the target warning zone and the confidence level of target features in the target warning zone. The confidence level of target features is then converted into the number of effective targets by recognizing confidence factors. The warning level is determined based on the number of effective targets and the area of ​​the target region, wherein the number of effective targets is determined by subtracting the equivalent number of false alarm signals detected in the environmental interference area from the number of targets corresponding to the confidence level of the target features.

[0007] Furthermore, the process of determining the ground image sequence based on ground images collected by the onboard visual sensor of the monitoring UAV during cruise includes: Record the initial ground images before the UAV enters cruise mode; Real-time ground images are acquired at fixed time intervals during the cruise. Calculate the difference features between the real-time ground image and the initial ground image at each time point, and arrange the difference features in chronological order to form a ground image sequence.

[0008] Furthermore, the process of obtaining ground anomaly features based on ground image sequences includes: Using the acquisition time point as the horizontal axis and the difference features between the real-time ground image and the initial ground image as the vertical axis, the data points in the ground image sequence are connected sequentially to form the original change curve; The original variation curve is subjected to moving average filtering to eliminate high-frequency noise, thereby obtaining the ground anomaly characteristic curve.

[0009] Furthermore, the process of identifying ground anomaly features based on ground anomaly feature curves includes: Calculate the first derivative of the ground anomaly characteristic curve to obtain the curve of the anomaly change rate over time; The ground anomaly is determined to be a normal scene based on the fact that the rate of abnormal change is monotonically decreasing and always higher than the baseline noise. If the abnormal rate of change decreases and then increases twice before decreasing again, the ground anomaly is determined to be a dynamic interference scenario. If the rate of abnormal change is still higher than the baseline noise at the end of the expected identification period, the ground anomaly feature is determined to be a continuously changing scene.

[0010] Furthermore, the process of determining whether to stop cruise based on real-time acquisition of the recognition status within the expected recognition duration includes: The system acquires real-time data on light intensity, flight altitude, and ground texture complexity during the cruise process in order to retrieve the corresponding expected recognition time from a pre-stored recognition time lookup table. Monitor the current patrol duration in real time during the patrol process; In response to the fact that the real-time abnormal change rate has dropped below the baseline noise when the cruise duration reaches the expected recognition duration, the current recognition state is determined to be stable and cruise is stopped.

[0011] Furthermore, the process of adjusting flight attitude based on the aforementioned ground anomaly features and anomaly coverage includes: In response to the ground anomaly being a dynamic interference scenario, it is determined that there are interfering objects competing with the early warning target in the target area; The degree of interference is determined based on the deviation between the abnormal coverage area and the expected coverage area, wherein the expected coverage area is predicted from a pre-stored scene model based on real-time acquired light intensity, flight altitude, and ground texture complexity; In response to the interference level exceeding a preset interference threshold, the hovering altitude in the flight attitude is adjusted to change the recognition field of view between the interfering object and the airborne visual sensor; In response to the fact that the ground anomaly features are a continuously changing scene, it is determined that the image signal transmitted from the target area to the airborne visual sensor is affected by changes in ambient light. The degree of light impact is determined based on the deviation between the abnormal coverage area and the expected coverage area. In response to the light impact exceeding a preset light threshold, the yaw angle in the flight attitude is adjusted to improve the uniformity of light incidence from the target area to the airborne vision sensor. The attitude adjustment amplitude is recorded to characterize the degree of influence of flight attitude changes on recognition behavior.

[0012] Furthermore, the process of setting segmented illumination procedures based on the reacquired ground image sequence and the amplitude of flight attitude adjustment includes: The total amount of identification information is determined based on the total difference in the reacquired ground image sequence when the abnormal rate of change drops below the baseline noise. The scene complexity level is determined based on the decay rate of the ground image sequence before the abnormal change rate drops below the baseline noise and the total amount of identification information, wherein the smaller the decay rate and the larger the total amount of identification information, the higher the scene complexity level. The scene complexity level is corrected based on the attitude adjustment range to obtain the target lighting intensity level; The duration of the low-illuminance segment, the duration of the high-illuminance segment, and the termination light intensity of the high-illuminance segment are determined based on the target illumination intensity level. The zoned lighting procedure shall be executed according to the determined duration of low illumination and high illumination. In low-illuminance areas, environmentally adaptive lighting is projected as an environmental disturbance zone; High-brightness identification lighting is projected in high-illuminance areas to serve as target warning zones.

[0013] Furthermore, the process of converting the confidence level of target features into the number of effective targets by identifying confidence factors includes: Obtain the identification confidence factor obtained in advance through standard target sample calibration. The identification confidence factor is used to determine the number of effective targets corresponding to a unit target feature confidence level. Multiply the confidence level of the target feature by the recognition confidence factor to obtain the number of valid targets.

[0014] Furthermore, the process of determining the warning level based on the number of effective targets and the area of ​​the target region includes: Divide the number of effective targets by the area of ​​the target region to obtain the target density per unit area of ​​the target region, and determine the warning level based on the target density per unit area.

[0015] On the other hand, the present invention also provides an emergency warning lighting system for drones that incorporates AI recognition functionality, comprising: The identification module is used to identify ground anomalies based on a sequence of ground images, wherein the sequence of ground images is determined based on ground images collected by the onboard visual sensor of the monitoring UAV during its cruise. The adjustment module is used to adjust the flight attitude based on the ground anomaly features and anomaly coverage range, wherein the anomaly coverage range is obtained based on the record after stopping cruise, and stopping cruise is determined based on whether the identification state within the expected identification time period is stable in real time. The projection module is used to control the UAV to hover over the target area for secondary hovering scanning, and to set a segmented lighting program based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment to perform segmented lighting projection, thereby obtaining the environmental interference zone and the target warning zone. The detection module is used to perform AI recognition model detection on the environmental interference zone and the target warning zone respectively, so as to obtain the false alarm signal in the environmental interference zone and the target feature confidence in the target warning zone, and convert the target feature confidence into the number of effective targets by recognizing the confidence factor; The early warning determination module is used to determine the early warning level based on the number of valid targets and the area of ​​the target region, wherein the number of valid targets is determined by subtracting the equivalent number of false alarm signals detected in the environmental interference area from the number of targets corresponding to the confidence level of the target features.

[0016] Compared with existing technologies, the beneficial effects of this invention are that by fusing the ground image features monitored in real time by the airborne visual sensor with the target detection of the AI ​​recognition model, the flight attitude is dynamically adjusted according to the ground anomaly features and a secondary hovering scan is performed. This can compensate for the recognition deviation caused by changes in lighting or occlusion. Furthermore, by using a segmented lighting program to separate the environmental interference zone from the target warning zone, and by deducting the equivalent number of false alarm targets from the recognition results, a more accurate emergency warning level can be obtained. Through multi-source data fusion and physical environment compensation mechanisms, the reliability of UAV emergency warning in complex scenarios is directly improved.

[0017] Furthermore, by using the difference features between real-time ground images acquired by a computer-mounted vision sensor and the initial ground images, a ground image sequence is formed and sorted. Then, the original change curve is formed by connecting the time points as the horizontal axis and the difference features as the vertical axis. After removing high-frequency noise through moving average filtering, a ground anomaly feature curve is obtained. Then, the first derivative of this curve is calculated to obtain the change curve of the anomaly change rate over time. Based on whether the anomaly change rate is monotonically decreasing, whether there is a second increase, or whether it is still higher than the baseline noise at the end of the expected recognition time, normal scene, dynamic interference scene, or continuously changing scene can be determined respectively. This processing method directly transforms the image signal output by the vision sensor into a readable feature of ground anomalies, so that anomalies such as the entry of interference objects or sudden changes in illumination during the cruise can be identified in real time without relying on any preset experience model or statistical classifier.

[0018] Furthermore, by acquiring real-time light intensity, flight altitude, and ground texture complexity during the cruise process and retrieving the expected recognition duration from a pre-stored recognition duration lookup table, and then determining whether the recognition status is stable and stopping the cruise based on whether the real-time abnormal change rate drops below the baseline noise, the cruise duration can adapt to changes in actual environmental conditions. When dynamic interference scenarios are identified, the degree of interference is calculated and the hovering altitude is adjusted accordingly to change the recognition field of view of the interference. When continuously changing scenarios are identified, the degree of light impact is calculated and the yaw angle is adjusted accordingly to improve the uniformity of light incidence. Closed-loop control can be directly performed based on the deviation between the visual measurement value and the expected value under standard conditions to compensate for the loss of recognition efficiency caused by dynamic interference or uneven lighting.

[0019] Furthermore, by determining the scene complexity level based on the abnormal change rate, decay rate, and total recognition information, and then using the attitude adjustment amplitude to correct this level to obtain the target illumination intensity level, it is possible to adaptively determine the duration and termination light intensity of the low-illuminance and high-illuminance segments, so that the segmented lighting program matches the recognition difficulty of the actual scene. By utilizing the inherent differences in the light response characteristics of the environmental interference area and the target warning area, areas that are greatly affected by ambient light can obtain adaptive lighting in the low-illuminance segment, while areas where target features are concentrated can obtain high-brightness recognition lighting in the high-illuminance segment. Thus, selective illumination enhancement of both can be achieved without pre-setting lighting parameters for different scenes.

[0020] Furthermore, the confidence level of the target feature is converted into the number of effective targets by a pre-calibrated identification confidence factor. This number is then divided by the area of ​​the region to obtain the target density per unit area, and the warning level is determined accordingly. This conversion process follows the statistical correspondence between confidence and target detection probability in the AI ​​identification model, thereby directly obtaining a traceable and quantifiable warning level assessment result. Attached Figure Description

[0021] Figure 1 This is a flowchart of the drone emergency early warning lighting method incorporating AI recognition functionality in this embodiment; Figure 2 A flowchart for determining the ground image sequence is provided for this embodiment; Figure 3 This embodiment defines a flowchart for the zoned lighting procedure; Figure 4 This is a schematic diagram of an emergency warning lighting system for drones that incorporates AI recognition functionality, as described in this embodiment. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] Please see Figure 1 The diagram shows a flowchart of the drone emergency warning lighting method incorporating AI recognition functionality. One aspect of this embodiment is a drone emergency warning lighting method incorporating AI recognition functionality, comprising: Step S1: Identify ground anomaly features that characterize ground anomalies based on ground image sequences used to characterize the degree of change in ground images, wherein the ground image sequence is determined based on ground images collected by the onboard visual sensor of the monitoring UAV during the cruise.

[0025] In this embodiment, the ground images are acquired in real time based on an airborne vision sensor. The airborne vision sensor is used to continuously acquire ground image frames through an image sensing chip, and the image signal processor set inside it outputs digital image data.

[0026] Step S2: Adjust the flight attitude based on ground anomaly features and anomaly coverage range used to characterize the anomaly coverage range during cruise. The anomaly coverage range is obtained based on the records after cruise is stopped. The stopping of cruise is based on whether the identification state within the expected identification time period is stable.

[0027] Step S3: Control the UAV to hover over the target area to perform a second hover scan, and set a segmented lighting program based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment to perform segmented lighting projection, thereby obtaining the environmental interference zone and the target warning zone. Step S4: Perform AI recognition model detection on the environmental interference zone and the target warning zone respectively to obtain the false alarm signal in the environmental interference zone and the target feature confidence in the target warning zone, and convert the target feature confidence into the number of effective targets through the recognition confidence factor used to characterize the number of effective targets corresponding to the single location confidence. Step S5: Determine the warning level based on the number of effective targets after correcting for false alarms and the area of ​​the target region. The number of effective targets is determined by subtracting the equivalent number of false alarm signals detected in the environmental interference area from the number of targets corresponding to the target feature confidence level.

[0028] By fusing ground image features monitored in real time by airborne vision sensors with target detection by AI recognition models, and dynamically adjusting flight attitude and performing secondary hovering scans based on ground anomaly features, it is possible to compensate for recognition deviations caused by changes in lighting or occlusion. Furthermore, by using a segmented lighting procedure to separate environmental interference zones from target warning zones, and deducting the equivalent number of false alarm targets from the recognition results, a more accurate emergency warning level can be obtained. Through multi-source data fusion and physical environment compensation mechanisms, the reliability of UAV emergency warnings in complex scenarios is directly improved.

[0029] Please see Figure 2 As shown, this is a flowchart for determining the ground image sequence in this embodiment. In this embodiment, the process of determining the ground image sequence based on ground images collected by the onboard visual sensor of the monitoring UAV during cruise includes: Record initial ground images before the drone enters cruise mode; Real-time ground images are acquired at fixed time intervals during the cruise. Calculate the difference features between the real-time ground image and the initial ground image at each time point, and arrange the difference features in chronological order to form a ground image sequence.

[0030] In this embodiment, the difference feature is calculated based on the pixel gradient difference, including: performing Sobel operator processing on the current image and the initial image respectively, and calculating the mean of the gradient magnitude difference as the difference feature.

[0031] Specifically, the process of obtaining ground anomaly features based on ground image sequences includes: Using the acquisition time point as the horizontal axis and the difference features between the real-time ground image and the initial ground image as the vertical axis, the data points in the ground image sequence are connected sequentially to form the original change curve; The original variation curve is filtered by moving average to eliminate high-frequency noise, and the ground anomaly characteristic curve is obtained.

[0032] Specifically, the process of identifying ground anomaly features based on ground anomaly feature curves includes: Calculate the first derivative of the ground anomaly characteristic curve to obtain the curve of the anomaly change rate over time; The ground anomaly was determined to be a normal scene based on the monotonically decreasing rate of change of the abnormal change and its constant higher rate than the baseline noise. The response to the abnormal rate of change decreasing followed by a second increase and then a decrease indicates that the ground anomaly is a dynamic disturbance scenario. If the rate of abnormal change is still higher than the baseline noise at the end of the expected identification period, the ground anomaly feature is determined to be a continuously changing scene.

[0033] In this embodiment, the baseline noise is the statistical fluctuation amplitude of the difference feature signal output by the image signal processor when there are no ground anomalies or the recognition state has reached a stable state. Its value is the standard deviation or peak-to-peak fluctuation range of the difference feature readings of the visual sensor during the stable operation period under standard scenarios.

[0034] By analyzing the differences between real-time ground images acquired by a computer-mounted vision sensor and the initial ground images, a ground image sequence is formed. Then, the original change curve is constructed by connecting the time points as the x-axis and the differences as the y-axis. After removing high-frequency noise through a moving average filter, a ground anomaly feature curve is obtained. The first derivative of this curve is then calculated to obtain the change curve of the anomaly change rate over time. Based on whether the anomaly change rate is monotonically decreasing, whether it shows a second increase, or whether it is still higher than the baseline noise at the end of the expected recognition period, normal scenes, dynamic interference scenes, or continuously changing scenes can be identified respectively. This processing method directly transforms the image signal output by the vision sensor into readable features of ground anomalies, enabling the immediate identification of anomalies such as the entry of interference objects or sudden changes in illumination during the cruise process, without relying on any preset empirical models or statistical classifiers.

[0035] Specifically, the process of determining whether to stop cruise based on real-time acquisition of the recognition status within the expected recognition duration includes: The system acquires real-time data on light intensity, flight altitude, and ground texture complexity during the cruise process, and retrieves the corresponding expected recognition time from a pre-stored recognition time lookup table.

[0036] In this embodiment, the light intensity, flight altitude, and ground texture complexity are acquired in real time based on the light intensity sensor, barometric altimeter, and image texture analysis module integrated on the UAV.

[0037] In this embodiment, the pre-stored recognition time reference table is constructed based on the actual stable recognition time measured in visual recognition experiments on standard target scenes under different combinations of light intensity, flight altitude, and ground texture complexity. The light intensity, flight altitude, and ground texture complexity each cover the normal range that may occur in actual operations, and multiple discrete levels are taken in each parameter dimension. The parameter levels are combined through orthogonal experimental design. The measurement is repeated under each parameter group until the abnormal change rate is first continuously lower than the baseline noise. The moment is recorded as the expected recognition time. For parameter combinations that are not directly tested, multiple linear interpolation or Gaussian process regression methods are used to calculate the corresponding relationship between environmental parameters and recognition time, thereby forming a reference table containing the correspondence between environmental parameters and recognition time.

[0038] Monitor the current patrol duration in real time during the patrol process; In response to the fact that the rate of real-time abnormal change has dropped below the baseline noise when the cruise duration has reached the expected recognition duration, the current recognition state is determined to be stable and cruise is stopped.

[0039] Specifically, the process of adjusting flight attitude based on ground anomaly features and anomaly coverage includes: In response to the dynamic interference scenario of ground anomalies, it is determined that there are interfering objects competing with the early warning target in the target area; The degree of interference is determined based on the deviation between the abnormal coverage area and the expected coverage area, where the expected coverage area is predicted from a pre-stored scene model based on real-time acquired light intensity, flight altitude, and ground texture complexity.

[0040] In this embodiment, the abnormal coverage range is the percentage of pixels in the target area that are identified as valid targets or interference areas, specifically the ratio of the sum of the detection box areas to the total image area.

[0041] In this embodiment, the expected coverage range is the theoretically expected normal scene coverage range within the field of view of the onboard visual sensor under the same illumination intensity, flight altitude, and ground texture complexity, assuming no dynamic interference. This value is obtained through a pre-stored scene model. The pre-stored scene model is an empirical mapping relationship constructed by conducting visual acquisition experiments on standard scenes under different combinations of illumination intensity, flight altitude, and ground texture complexity, recording the coverage range value when recognition is stable, and associating the coverage range value with the corresponding environmental parameters. The interference level is the difference between the expected coverage range and the abnormal coverage range, divided by the expected coverage range, to quantify the intensity of the dynamic interference's encroachment on the target recognition field of view.

[0042] In response to interference levels exceeding a preset interference threshold, the hovering altitude in the flight attitude is adjusted to change the recognition field of view between the interfering object and the airborne vision sensor.

[0043] In this embodiment, the preset interference threshold is determined based on the standard deviation of the abnormal coverage obtained by repeated measurements of a standard scenario under the same environmental conditions, and in combination with the confidence interval multiple corresponding to the significance level acceptable in the field of statistics. This multiple makes the deviation exceeding the threshold statistically attributable to dynamic interference rather than random fluctuation.

[0044] In response to the continuously changing scene of ground anomalies, it is determined that the image signal transmitted from the target area to the airborne vision sensor is affected by changes in ambient light. The degree of light impact is determined by the deviation between the abnormal coverage area and the expected coverage area.

[0045] In this embodiment, the degree of illumination impact is the difference between the expected coverage area and the abnormal coverage area, divided by the expected coverage area, to characterize the degree of relative reduction in the recognition area caused by uneven illumination.

[0046] In response to the light intensity exceeding a preset light threshold, the yaw angle in the flight attitude is adjusted to improve the uniformity of light incidence from the target area to the airborne vision sensor.

[0047] In this embodiment, the preset illumination threshold is determined based on the standard deviation of the abnormal coverage obtained by repeated measurements of a standard scene under the same environmental conditions, and in combination with the confidence interval multiple corresponding to the significance level acceptable in the field of statistics. This multiple makes the deviation exceeding the threshold statistically attributable to illumination changes rather than random fluctuations.

[0048] The attitude adjustment amplitude is recorded to characterize the degree of influence of flight attitude changes on recognition behavior.

[0049] By acquiring real-time data on light intensity, flight altitude, and ground texture complexity during the cruise process and retrieving the expected recognition duration from a pre-stored recognition duration lookup table, and then determining whether the recognition status is stable and stopping the cruise based on whether the real-time abnormal change rate drops below the baseline noise, the cruise duration can adapt to changes in actual environmental conditions. When dynamic interference scenarios are identified, the degree of interference is calculated and the hovering altitude is adjusted accordingly to change the recognition field of view of the interference. When continuously changing scenarios are identified, the degree of light impact is calculated and the yaw angle is adjusted accordingly to improve the uniformity of light incidence. Closed-loop control can be performed directly based on the deviation between the visual measurement value and the expected value under standard conditions to compensate for the loss of recognition efficiency caused by dynamic interference or uneven lighting.

[0050] Please see Figure 3 As shown, this is a flowchart of the segmented lighting procedure in this embodiment. In this embodiment, the process of setting the segmented lighting procedure based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment includes: The total amount of identification information is determined based on the total difference in the re-acquired ground image sequence when the rate of abnormal change drops below the baseline noise. The scene complexity level is determined based on the decay rate of the ground image sequence before the abnormal change rate drops below the baseline noise and the total amount of recognizable information. The smaller the decay rate and the larger the total amount of recognizable information, the higher the scene complexity level. The scene complexity level is adjusted based on the attitude adjustment range to obtain the target lighting intensity level.

[0051] In this embodiment, the scene complexity level is corrected by adjusting the scene complexity level up or down based on the magnitude of the attitude adjustment. The greater the attitude adjustment, the greater the offset of the target illumination intensity level relative to the original scene complexity level, so as to compensate for the actual impact of the flight attitude adjustment process on the visual recognition effect.

[0052] The duration of low illuminance, duration of high illuminance, and termination light intensity of high illuminance are determined based on the target lighting intensity level. The zoned lighting procedure shall be executed according to the determined duration of low illumination and high illumination. In low-illuminance areas, environmentally adaptive lighting is projected as an environmental disturbance zone; High-brightness identification lighting is projected in high-illuminance areas to serve as target warning zones.

[0053] The scene complexity level is determined by combining the abnormal change rate, decay rate, and total recognition information. The level is then corrected using the attitude adjustment amplitude to obtain the target illumination intensity level. This allows for the adaptive determination of the duration and termination intensity of low-illuminance and high-illuminance segments, ensuring that the segmented illumination program matches the recognition difficulty of the actual scene. By utilizing the inherent differences in light response characteristics between environmental interference areas and target warning areas, areas more affected by ambient light receive adaptive illumination in the low-illuminance segment, while areas with concentrated target features receive high-brightness recognition illumination in the high-illuminance segment. This achieves selective illumination enhancement for both without the need to pre-set illumination parameters for different scenes.

[0054] Specifically, the process of converting the confidence level of target features into the number of effective targets by identifying confidence factors includes: Obtain the identification confidence factor obtained in advance through standard target sample calibration. The identification confidence factor is used to characterize the number of effective targets corresponding to the confidence level of a unit target feature.

[0055] In this embodiment, the identification confidence factor is obtained through statistical analysis of historical labeled samples. Specifically, it includes: using pre-labeled training samples to obtain the confidence of each detected target output by the AI ​​recognition model and the corresponding number of real targets, establishing a fitting relationship between the cumulative value of the detected target confidence and the number of real targets, and determining the identification confidence factor based on the fitting results.

[0056] Multiply the target feature confidence score by the identification confidence factor to obtain the corresponding number of valid targets.

[0057] In this embodiment, the AI ​​recognition model is a pre-trained target detection model used to detect targets in ground images acquired by an airborne vision sensor and output target detection boxes, target categories, and corresponding target feature confidence scores. The AI ​​recognition model includes a shared feature extraction network, a target candidate generation branch, and a detection output branch. The shared feature extraction network is used to extract multi-scale features from the input image, the target candidate generation branch is used to generate candidate target detection boxes, and the detection output branch is used to output the target category, target location, and target feature confidence score corresponding to each candidate target detection box. The AI ​​recognition model is trained offline using a pre-labeled UAV scene image dataset and is directly called to perform target detection during actual early warning processes.

[0058] Specifically, the process of determining the warning level based on the number of effective targets and the area of ​​the target region includes: Divide the number of effective targets by the area of ​​the target region to obtain the target density per unit area of ​​the target region, and determine the warning level based on the target density per unit area.

[0059] By converting the confidence level of target features into the number of effective targets through a pre-calibrated identification confidence factor, and then dividing the number by the area of ​​the region to obtain the target density per unit area, the warning level is determined accordingly. This conversion process follows the statistical correspondence between confidence and target detection probability in the AI ​​identification model, thereby directly obtaining traceable and quantifiable warning level assessment results.

[0060] In this embodiment, the process of determining false alarm signals includes: numbering and recording the target detection results within the environmental interference zone, and saving the position coordinates, target category, and target confidence level of the corresponding target detection box; performing a second hover scan after completing the flight attitude adjustment, and performing target detection again; comparing the two detection results according to the target category and spatial location; when the target detected in the first scan is not detected again in the second scan, and the confidence level of the target corresponding to the first detection is lower than the preset confidence threshold, the target is determined to be a false alarm signal; wherein, the preset confidence threshold is determined based on the verification results of the AI ​​recognition model on a pre-labeled UAV scene dataset. By statistically analyzing the false detection rate and false negative rate under different confidence thresholds, the confidence value corresponding to the best comprehensive recognition performance is selected as the preset confidence threshold, and it remains fixed after the model is deployed.

[0061] Please see Figure 4 As shown, this is a schematic diagram of a drone emergency warning lighting system incorporating AI recognition functionality according to this embodiment. Furthermore, this embodiment also provides a drone emergency warning lighting system incorporating AI recognition functionality, including: The identification module is used to identify ground anomalies based on a sequence of ground images, wherein the sequence of ground images is determined based on ground images collected by the onboard visual sensor of the monitoring UAV during its cruise. The adjustment module, which is connected to the recognition module, is used to adjust the flight attitude based on ground anomaly features and anomaly coverage. The anomaly coverage is obtained based on the record after stopping cruise, and stopping cruise is determined based on whether the recognition status within the expected recognition time is stable. The projection module, which is connected to the recognition module and the adjustment module respectively, is used to control the UAV to hover over the target area for secondary hovering scanning, and to set a segmented lighting program based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment to perform segmented lighting projection, thereby obtaining the environmental interference zone and the target warning zone. The detection module, which is connected to the projection module, is used to perform AI recognition model detection on the environmental interference zone and the target warning zone respectively, so as to obtain the false alarm signal in the environmental interference zone and the target feature confidence in the target warning zone, and convert the target feature confidence into the number of effective targets by recognizing the confidence factor; The early warning determination module, which is connected to the detection module, is used to determine the early warning level based on the number of effective targets and the area of ​​the target region. The number of effective targets is determined by subtracting the equivalent number of false alarm signals detected in the environmental interference area from the number of targets corresponding to the target feature confidence level.

[0062] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention; various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for emergency early warning lighting of drones combined with AI recognition function, characterized in that, include: Ground anomaly features are identified based on ground image sequences, where the ground image sequences are determined based on ground images collected by the onboard visual sensor of the monitoring UAV during its cruise. The flight attitude is adjusted based on the ground anomaly features and anomaly coverage area, wherein the anomaly coverage area is obtained based on the record after stopping cruise, and stopping cruise is determined based on whether the identification status within the expected identification time period is stable. The UAV is controlled to hover over the target area for secondary hovering scanning, and a segmented lighting program is set based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment to perform segmented lighting projection, thereby obtaining the environmental interference zone and the target warning zone. AI recognition models are used to detect false alarm signals in the environmental interference zone and target warning zone respectively, in order to obtain the confidence level of target features in the target warning zone and the confidence level of target features in the target warning zone. The confidence level of target features is then converted into the number of effective targets by recognizing confidence factors. The warning level is determined based on the number of effective targets and the area of ​​the target region, wherein the number of effective targets is determined by subtracting the equivalent number of false alarm signals detected in the environmental interference area from the number of targets corresponding to the confidence level of the target features.

2. The drone emergency early warning lighting method combining AI recognition function according to claim 1, characterized in that, The process of determining a ground image sequence based on ground images collected by the onboard visual sensor of a monitoring drone during its cruise includes: Record the initial ground images before the UAV enters cruise mode; Real-time ground images are acquired at fixed time intervals during the cruise. Calculate the difference features between the real-time ground image and the initial ground image at each time point, and arrange the difference features in chronological order to form a ground image sequence.

3. The drone emergency early warning lighting method combining AI recognition function according to claim 2, characterized in that, The process of obtaining ground anomaly features based on ground image sequences includes: Using the acquisition time point as the horizontal axis and the difference features between the real-time ground image and the initial ground image as the vertical axis, the data points in the ground image sequence are connected sequentially to form the original change curve; The original variation curve is subjected to moving average filtering to eliminate high-frequency noise, thereby obtaining the ground anomaly characteristic curve.

4. The drone emergency early warning lighting method combining AI recognition function according to claim 3, characterized in that, The process of identifying ground anomaly features based on ground anomaly feature curves includes: Calculate the first derivative of the ground anomaly characteristic curve to obtain the curve of the anomaly change rate over time; The ground anomaly is determined to be a normal scene based on the fact that the rate of abnormal change is monotonically decreasing and always higher than the baseline noise. If the abnormal rate of change decreases and then increases twice before decreasing again, the ground anomaly is determined to be a dynamic interference scenario. If the rate of abnormal change is still higher than the baseline noise at the end of the expected identification period, the ground anomaly feature is determined to be a continuously changing scene.

5. The drone emergency early warning lighting method combining AI recognition function according to claim 4, characterized in that, The process of determining whether to stop cruise based on whether the recognition status within the expected recognition duration has reached stability in real time includes: The system acquires real-time data on light intensity, flight altitude, and ground texture complexity during the cruise process in order to retrieve the corresponding expected recognition time from a pre-stored recognition time lookup table. Monitor the current patrol duration in real time during the patrol process; In response to the fact that the real-time abnormal change rate has dropped below the baseline noise when the cruise duration reaches the expected recognition duration, the current recognition state is determined to be stable and cruise is stopped.

6. The drone emergency early warning lighting method combining AI recognition function according to claim 5, characterized in that, The process of adjusting flight attitude based on the aforementioned ground anomaly features and anomaly coverage includes: In response to the ground anomaly being a dynamic interference scenario, it is determined that there are interfering objects competing with the early warning target in the target area; The degree of interference is determined based on the deviation between the abnormal coverage area and the expected coverage area, wherein the expected coverage area is predicted from a pre-stored scene model based on real-time acquired light intensity, flight altitude, and ground texture complexity; In response to the interference level exceeding a preset interference threshold, the hovering altitude in the flight attitude is adjusted to change the recognition field of view between the interfering object and the airborne visual sensor; In response to the fact that the ground anomaly features are a continuously changing scene, it is determined that the image signal transmitted from the target area to the airborne visual sensor is affected by changes in ambient light. The degree of light impact is determined based on the deviation between the abnormal coverage area and the expected coverage area. In response to the light impact exceeding a preset light threshold, the yaw angle in the flight attitude is adjusted to improve the uniformity of light incidence from the target area to the airborne vision sensor. The attitude adjustment amplitude is recorded to characterize the degree of influence of flight attitude changes on recognition behavior.

7. The drone emergency early warning lighting method combining AI recognition function according to claim 6, characterized in that, The process of setting segmented illumination procedures based on reacquired ground image sequences and adjusting flight attitude amplitude includes: The total amount of identification information is determined based on the total difference in the reacquired ground image sequence when the abnormal rate of change drops below the baseline noise. The scene complexity level is determined based on the decay rate of the ground image sequence before the abnormal change rate drops below the baseline noise and the total amount of identification information, wherein the smaller the decay rate and the larger the total amount of identification information, the higher the scene complexity level. The scene complexity level is corrected based on the attitude adjustment range to obtain the target lighting intensity level; The duration of the low-illuminance segment, the duration of the high-illuminance segment, and the termination light intensity of the high-illuminance segment are determined based on the target illumination intensity level. The zoned lighting procedure shall be executed according to the determined duration of low illumination and high illumination. In low-illuminance areas, environmentally adaptive lighting is projected as an environmental disturbance zone; High-brightness identification lighting is projected in high-illuminance areas to serve as target warning zones.

8. The drone emergency early warning lighting method combining AI recognition function according to claim 7, characterized in that, The process of converting target feature confidence into the number of effective targets by identifying confidence factors includes: Obtain the identification confidence factor obtained in advance through standard target sample calibration. The identification confidence factor is used to determine the number of effective targets corresponding to a unit target feature confidence level. Multiply the confidence level of the target feature by the recognition confidence factor to obtain the number of valid targets.

9. The drone emergency early warning lighting method combining AI recognition function according to claim 8, characterized in that, The process of determining the early warning level based on the number of valid targets and the area of ​​the target region includes: Divide the number of effective targets by the area of ​​the target region to obtain the target density per unit area of ​​the target region, and determine the warning level based on the target density per unit area.

10. An emergency warning lighting system for drones incorporating AI recognition capabilities, based on the emergency warning lighting method according to any one of claims 1-9, characterized in that, include: The identification module is used to identify ground anomalies based on a sequence of ground images, wherein the sequence of ground images is determined based on ground images collected by the onboard visual sensor of the monitoring UAV during its cruise. The adjustment module is used to adjust the flight attitude based on the ground anomaly features and anomaly coverage range, wherein the anomaly coverage range is obtained based on the record after stopping cruise, and stopping cruise is determined based on whether the identification state within the expected identification time period is stable in real time. The projection module is used to control the UAV to hover over the target area for secondary hovering scanning, and to set a segmented lighting program based on the re-acquired ground image sequence and the amplitude of flight attitude adjustment to perform segmented lighting projection, thereby obtaining the environmental interference zone and the target warning zone. The detection module is used to perform AI recognition model detection on the environmental interference zone and the target warning zone respectively, so as to obtain the false alarm signal in the environmental interference zone and the target feature confidence in the target warning zone, and convert the target feature confidence into the number of effective targets by recognizing the confidence factor; The early warning determination module is used to determine the early warning level based on the number of valid targets and the area of ​​the target region, wherein the number of valid targets is determined by subtracting the equivalent number of false alarm signals detected in the environmental interference area from the number of targets corresponding to the confidence level of the target features.

Images