A method for detecting visual targets of a thermal imaging unmanned aerial vehicle for search and rescue missions

Abstract

The present application relates to target detection technical field, specifically to a kind of unmanned aerial vehicle thermal imaging visual target detection method for search and rescue task, comprising the following steps, obtain unmanned aerial vehicle multiple frames thermal imaging image, extract area thermal difference feature according to window, mark non-background area to generate candidate set, fit reconstruction suspected thermal target contour, analyze trajectory and thermal variation rate and exclude background interference, identify jump anomaly and locate barycenter, generate target relocation signal.The present application, by constructing the thermal value range and variance index sequence in image area and combining temperature baseline difference judgment thermal anomaly area, the direction vector of coordinate trajectory in combination with thermal change parameter executes background disturbance comparison, reduces the false detection probability caused by background noise, according to the thermal value and area change rate linkage thermal isoclines closed degree to carry out target jump identification, improve the target discrimination accuracy under the shielding scene, overall improve the robustness of thermal target extraction in complex search and rescue environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a UAV thermal imaging visual target detection method for search and rescue tasks. BACKGROUND

[0002] The technical field of target detection aims to accurately identify and locate multiple targets of interest in images or video sequences through computer vision methods. The field of target detection relies on core algorithm frameworks such as deep learning, convolutional neural networks, feature pyramid networks, and region proposal networks, covering image input, feature extraction, candidate region generation, target classification, and bounding box regression processing procedures. Target detection is widely used in security monitoring, autonomous driving, medical image analysis, industrial detection, and other scenarios, with high requirements for model detection accuracy, real-time performance, and robustness in complex backgrounds. The research focus in this field also includes optimization of adaptability to complex scenarios such as occlusion, scale variation, illumination variation, and multiple target overlap, promoting the deployability and stability of model performance in practical applications.

[0003] Among them, the UAV thermal imaging visual target detection method for search and rescue tasks aims to combine UAV platforms and thermal imaging vision to build a detection method that can perform target identification and positioning in complex environments. Its main use is to automatically detect and label personnel, animals, or other life target positions based on thermal imaging images in emergency search and rescue scenarios such as natural disasters and accidents, providing real-time and efficient information support for rescue deployment, improving search and rescue efficiency and accuracy, and reducing search and rescue delays and risks.

[0004] Traditional detection methods rely on temperature mutation regions in thermal imaging images for static image analysis, lack dynamic quantitative means for local region temperature values and background temperature distribution differences, leading to mislabeling in uneven background temperature difference distribution or presence of high temperature interference sources, and lack of a complete thermal value change sequence model to identify the continuity of target features in consecutive frames, resulting in target shape jumps caused by occlusion or posture changes that cannot be effectively identified and confirmed, and inability to compensate for missing pixel information in image edges or resolution reduction scenarios, causing contour boundary breakage and incomplete target identification, affecting stable tracking and effective positioning efficiency of thermal targets. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art and to provide a UAV thermal imaging visual target detection method for search and rescue tasks.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: a UAV thermal imaging visual target detection method for search and rescue tasks, comprising the following steps:

[0007] S1: acquire multiple frames of thermal imaging images collected by a UAV in a search and rescue task area, mark image sub-regions with non-background features according to the degree of thermal difference deviation from the background thermal equilibrium baseline temperature difference, and generate a high-thermal candidate positioning region set;

[0008] S2: based on the high-thermal candidate positioning region set, determine whether the interval thermal value meets the fitting condition, if yes, perform linear gradient fitting processing on the pixel thermal value in the interval, reconstruct the suspected human body contour region with continuous boundaries in the region, and generate a suspected target thermal form region;

[0009] S3: based on the spatial position of each contour region in the suspected target thermal form region, acquire the coordinate change trajectory in the continuous image frames, and compare it with the path record in the constructed thermal background disturbance trajectory table, screen out the background disturbance type hot spot region, and generate an effective thermal target candidate region set;

[0010] S4: call the effective thermal target candidate region set, respectively calculate the thermal value change rate and the pixel area change rate of the region between frames, judge whether there is local shielding or target switching situation, acquire the thermal isotheral closed degree of the corresponding region in the frame as an auxiliary index, and generate an abnormal trajectory target recognition marker set.

[0011] The application improves that the high-thermal candidate positioning region set includes a multi-window thermal difference significant block, a region thermal response standard deviation value, a thermal difference deviation background reference value mapping matrix and an abnormal thermal difference region boundary coordinate, the suspected target thermal form region includes a thermal value gradient continuous image block, a linear fitting thermal contour distribution, a bilinear interpolation pixel layer and a potential target form closed boundary in the fitting block, the effective thermal target candidate region set includes a thermal stable region trajectory set, a spatial target position set after excluding disturbance, a thermal response continuous feature data and a non-background matching identification coordinate point, and the abnormal trajectory target recognition marker set includes a thermal value jump identification mark, a shielding behavior inference mark, an area mutation region number and a form closed degree judgment mark.

[0012] The application improves that the acquisition step of the high-thermal candidate positioning region set is specifically:

[0013] S111: acquire multiple frames of thermal imaging images collected by a UAV in a search and rescue task area, divide each frame of image into a fixed size window region by using a regular grid, extract the pixel difference value between the maximum thermal value pixel in each window and the minimum thermal value pixel in the corresponding peripheral window, obtain the thermal response difference in the window region, and acquire a window thermal difference response set;

[0014] S112: Based on the set of thermal difference responses between windows, the range and variance of the pixel difference sequence in each image window are calculated to establish a difference statistical index. The statistical value in the window area is offset and compared with the reference temperature difference of the overall background of the thermal imaging image. The thermal difference response deviation value of the image window area is calculated and obtained.

[0015] S113: Based on the thermal response deviation value, compare it with the thermal response offset threshold, filter the window numbers that exceed the thermal response offset threshold, extract the boundary coordinate information of the corresponding image region, perform region numbering and structure classification processing, and uniformly mark the region set as non-background response region to obtain the high-heat candidate positioning region set.

[0016] The present invention is improved in that the step of obtaining the suspected target thermal morphology region is specifically as follows:

[0017] S211: Based on the set of high-heat candidate positioning regions, extract the maximum and minimum heat value pixels in each region, construct heat value interval boundaries, and perform interval judgment on each interval boundary according to the set human body surface temperature reference range. Select image regions whose heat value intervals fall within the human body surface temperature reference range to obtain a target region set that conforms to the heat reference interval.

[0018] S212: Call the target region set that conforms to the thermal reference interval, perform coordinate mapping and numerical extraction on the effective pixels with recorded thermal values ​​inside each image region, construct a set of two-dimensional coordinate and thermal value pairs, calculate the thermal value gradient difference between adjacent pixels in the horizontal and vertical directions respectively, estimate the continuous trend of pixel response in the region according to the linear increment rule, construct the thermal value distribution function surface, and generate thermal gradient fitting distribution information.

[0019] S213: Based on the thermal gradient fitting distribution information, identify pixel thermal value void regions in the fitted surface caused by missing, occluded or noise suppression, and perform bilinear interpolation on the neighborhood around the void to fill the thermal value missing. In the filled image area, extract the closed boundary curve according to the thermal value change gradient direction, construct the contour structure map, and obtain the suspected target thermal morphology region.

[0020] The present invention is improved in that the step of obtaining the effective thermal target candidate region set is specifically as follows:

[0021] S311: Based on the geometric boundary coordinates of each contour structure in the suspected target thermal morphology region, extract the positioning point coordinates of each target in consecutive image frames, construct the spatial position change trajectory within the frame sequence, record the pixel displacement direction and amplitude between adjacent frames, calculate the multi-frame path vector, and obtain the contour trajectory direction parameter set.

[0022] S312: Call the contour trajectory direction parameter set, combine it with the corresponding frame number, count the number of times each coordinate point appears, calculate the number of frames in which the target exists in the image sequence, extract the thermal response value change sequence within the corresponding time period, calculate the average temperature change rate according to the temperature increase and decrease trend between frames, construct continuous thermal dynamic features bound to each target contour, and obtain target thermal time change information.

[0023] S313: Based on the target thermal time change information, the three feature parameters of trajectory direction vector, duration frame number and temperature change rate are normalized to construct a three-dimensional feature vector. The three-dimensional feature vector is then matched with the perturbation path set in the pre-stored thermal background perturbation trajectory table using Euclidean distance. Background perturbation items with similarity higher than the similarity threshold are filtered out to obtain a set of effective thermal target candidate regions.

[0024] The present invention is improved in that the process of normalizing the three feature parameters of trajectory direction vector, number of consecutive frames and temperature change rate is specifically as follows: the trajectory direction vector in each image frame is represented as a two-dimensional coordinate difference sequence, the number of consecutive frames is normalized proportionally according to the total number of frames, and the temperature change rate is normalized based on the average value of pixel heat value change per unit time.

[0025] The similarity threshold is a distance boundary value calculated during the training phase based on the mean and standard deviation of the Euclidean distance of historical perturbation paths in the thermal background perturbation trajectory table. The distance boundary value is the standard deviation of the mean weighted by a set ratio.

[0026] The present invention is improved in that the step of obtaining the abnormal trajectory target identification tag set is specifically as follows:

[0027] S411: Based on the set of effective thermal target candidate regions, call the maximum heat value, minimum heat value and corresponding frame number information recorded in each region in consecutive frames, calculate the heat value difference between consecutive frames, normalize the region time span, calculate the temperature change amplitude per unit time, and obtain the heat value inter-frame change rate set.

[0028] S412: Call the set of heat value inter-frame change rate, combine it with the number of pixels in the corresponding region in each image frame, count the area change value between consecutive frames, calculate the jump trend of the heat target region, and compare it with the occlusion jump threshold. If the jump trend of the heat target region is greater than the occlusion jump threshold, mark it as meeting the abnormal change standard and generate a jump sensitive region number set.

[0029] S413: Based on the set of abrupt change sensitive regions, retrieve the thermal contour line closure status of the corresponding region in the image, perform boundary integrity judgment on each thermal contour line, extract the curve closure rate as a structural stability index, mark the unstable region of the structural boundary and assign it an identification status code, and establish an abnormal trajectory target identification tag set.

[0030] The present invention has an improvement, wherein the method further includes the following steps:

[0031] S5: Based on the abnormal trajectory target identification marker set, the hot target area that has changed is marked as the hot target area. The position and thermal response distribution of the area in the current frame are called. The hot spot dense center point is extracted and weighted center positioning processing is performed. The in-frame area rescan window is constructed. The normal image of the area, the thermal imaging image and the area positioning information are sent to the rescue personnel. The search and rescue target relocation trigger signal is generated.

[0032] The search and rescue target relocation trigger signal includes the thermal response centroid coordinate point, regional rescan window parameters, trigger condition judgment value, and target detection module activation command.

[0033] The present invention is improved in that the step of obtaining the search and rescue target relocation trigger signal is specifically as follows:

[0034] S511: Based on the regions marked as having abrupt changes in the abnormal trajectory target identification marker set, extract the thermal response matrix of the corresponding region in the current image frame, and search for pixels with heat values ​​greater than the median heat value within the region. Calculate the centroid coordinates of the hotspots and assign weighted values ​​to the heat values ​​of the hotspots. Perform a two-dimensional weighted average to obtain the focal point of the region's thermal intensity, and obtain the set of thermal response weighted centroid coordinates.

[0035] S512: Call the thermal response weighted centroid coordinate set, construct a square region window with a side length twice that of the original target outline with each centroid coordinate as the center, and extract the infrared thermal imaging image, visible light image and heat value data frame corresponding to the window, construct the intra-frame target rescan information structure indexed by spatial thermal density, and obtain the region rescan window data block set.

[0036] S513: Based on the area rescan window data block set, write the corresponding infrared thermal image, visible light image and centroid positioning coordinates into the search and rescue intermediate interaction frame structure, and send them synchronously to the front-end rescue node as target information to be confirmed. Encode the target data and register the status field, and establish a search and rescue target relocation trigger signal.

[0037] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0038] In this invention, by constructing a sequence of thermal value range and variance indices within an image region and combining it with temperature baseline differences to determine thermal anomaly regions, the accuracy of locating significant thermal features of living targets is improved. Linear gradient fitting and bilinear interpolation methods are used to complete missing image information, enhancing the boundary continuity and integrity of suspected target contours. Background perturbation comparison is performed by combining the direction vector of coordinate trajectories and thermal change parameters in multiple frames of images, reducing the probability of false detection caused by background noise. Target jump recognition is performed based on the inter-frame thermal value and area change rate linked to the thermal contour closure, improving the accuracy of target discrimination in occluded scenarios. The centroid coordinates of densely populated hotspot areas are extracted for precise positioning, and target information is reported by combining regional rescanning with multimodal images. Overall, this improves the robustness of thermal target extraction, the continuity of target recognition, and the real-time performance of relocation response in complex search and rescue environments. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is a flowchart illustrating the process of obtaining a set of high-temperature candidate positioning regions according to the present invention;

[0041] Figure 3 This is a flowchart illustrating the process of obtaining a suspected target thermal morphology region according to the present invention;

[0042] Figure 4 This is a flowchart illustrating the process of obtaining a set of effective thermal target candidate regions according to the present invention;

[0043] Figure 5 This is a flowchart illustrating the process of obtaining an abnormal trajectory target identification marker set according to the present invention;

[0044] Figure 6 This is a flowchart illustrating the process of obtaining the search and rescue target relocation trigger signal according to the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0046] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0047] Please see Figure 1 This invention provides a technical solution: a method for detecting unmanned aerial vehicle (UAV) thermal imaging visual targets for search and rescue missions, comprising the following steps:

[0048] S1: Acquire multiple frames of thermal imaging images collected by the UAV in the search and rescue mission area, divide the image area according to a fixed window, extract the difference sequence between the maximum heat value pixel and the corresponding minimum heat value pixel in each area, calculate the range and variance of the difference sequence in each area, mark the image sub-regions with non-background features according to the degree of thermal difference deviation from the background thermal equilibrium baseline temperature difference, and generate a set of high-heat candidate positioning areas.

[0049] Range refers to the difference between the maximum and minimum values ​​in the dataset, used to measure the width of the interval distribution; variance represents the dispersion of the calorific value difference sequence; baseline temperature difference can be calculated from the historical temperature average of non-target areas and is used to construct a thermal background reference.

[0050] S2: Based on the high-heat candidate localization region set, call the maximum and minimum heat values ​​in each region, set the interval as the reference range of human body surface temperature, determine whether the heat value in the interval meets the fitting condition, if it does, perform linear gradient fitting processing on the pixel heat values ​​in the interval, and perform bilinear interpolation on the missing pixel gaps to reconstruct the suspected human body contour region with continuous boundaries in the region and generate the suspected target thermal morphology region.

[0051] Human body surface temperature reference range is a commonly used standard in medicine and thermal imaging; bilinear interpolation is a common local interpolation algorithm in image processing, which performs weighted calculations based on four neighboring pixels.

[0052] S3: Based on the spatial location of each contour region in the suspected target thermal morphology region, obtain the coordinate change trajectory in continuous image frames, extract three sets of parameters: the direction vector of the coordinate point, the number of frames of thermal duration, and the average temperature change rate, and compare them with the path records in the constructed thermal background disturbance trajectory table. Use three difference indicators to calculate the Euclidean distance to determine whether it is a background disturbance hotspot region, filter out background disturbance hotspot regions, and generate a set of effective thermal target candidate regions.

[0053] The direction vector represents the translation trend of the hot target and can be calculated from the positional changes between frames; the hot duration frame number is the number of frames in which the hot spot is maintained in a statistically continuous frame; the average temperature change rate is used to judge the stability of the heat source; and the Euclidean distance is used to measure the similarity of three-dimensional vectors.

[0054] S4: Call the maximum heat value, minimum heat value and area value of each region in the effective thermal target candidate region set in consecutive frames, calculate the heat value change rate and pixel area change rate of the region between frames respectively, and compare them with the set occlusion jump threshold to determine whether there is local occlusion or target switching. If the jump condition is met, obtain the thermal contour closure degree of the corresponding region in the frame as an auxiliary indicator to generate an abnormal trajectory target identification tag set.

[0055] The pixel area change rate is defined as the ratio of the pixel number difference to the initial frame area; the occlusion jump threshold can be determined experimentally, and is generally set at a temperature change of 30% or more and an area change of 25% or more; the thermal contour closure degree is used to determine whether the heat source is complete by the degree of contour geometric closure.

[0056] S5: Based on the abnormal trajectory target identification marker set, the hot target area that has changed is marked. The location and thermal response distribution of the area in the current frame are called. The hot spot dense center point is extracted and weighted center point localization processing is performed to obtain the center point coordinates. The in-frame area rescan window is constructed. The normal image of the area, the thermal imaging image and the area positioning information are sent to the rescue personnel to generate a search and rescue target relocation trigger signal.

[0057] Weighted centroid positioning is performed by calculating the centroid of pixel coordinates based on the magnitude of the heat value; the dense hotspot center point refers to the center of the region where the local heat value is greater than 1.5 times the average heat value.

[0058] The high-temperature candidate location region set includes multi-window significant thermal difference blocks, regional thermal response standard deviation, thermal difference deviation from background reference value mapping matrix, and boundary coordinates of abnormal thermal difference regions. The suspected target thermal morphology region includes continuous image blocks of thermal value gradient, thermal contour distribution after linear fitting, bilinear interpolation completed pixel layer, and potential target morphology closed boundary in the fitted block. The effective thermal target candidate region set includes thermally stable zone trajectory set, spatial target location set after excluding disturbances, thermal response continuous feature data, and non-background matching identifier coordinate points. The abnormal trajectory target identification marker set includes thermal value jump identification marker, occlusion behavior inference marker, area abrupt change region number, and morphological closure judgment marker. The search and rescue target relocation trigger signal includes thermal response centroid coordinate points, regional rescanning window parameters, trigger condition judgment value, and target detection module activation command.

[0059] Please see Figure 2 The specific steps for obtaining the high-heat candidate location region set are as follows:

[0060] S111: Acquire multiple frames of thermal imaging images collected by the UAV in the search and rescue mission area. Divide each frame of the image into a fixed-size window area using a regular grid. Extract the pixel difference between the maximum thermal value pixel in each window and the minimum thermal value pixel in the corresponding surrounding window to obtain the thermal response difference in the window area and obtain the set of thermal difference responses between windows.

[0061] Acquire multiple frames of thermal imaging images taken by the drone in the search and rescue mission area, for example, the drone captures these images at a rate of [number] frames per second. Frame rate at altitude Meter height for a piece A search was conducted over a forest area of ​​square kilometers, obtaining data at a resolution of [resolution missing]. A sequence of consecutive thermal imaging images of pixels, with the first captured thermal imaging image divided into pixels using a regular grid. One, that is Each size is A window area of ​​pixels, and each window is uniquely numbered, starting from the top left window. Go to the bottom right window Then, extract any window, for example, window number 1. The maximum heat value pixel within the area, the window contains a total of Each pixel, after traversal, the window The highest pixel heat value within, i.e., grayscale value At the same time, lock the window. The eight adjacent windows, i.e., windows , , , , , , , and within the coverage of these eight adjacent windows Among the pixels, find and extract the lowest pixel heat value, i.e., the grayscale value. Calculate the difference between these two pixel values ​​to obtain This value serves as the window. A measure of thermal response difference, applying this process to all For each window region, the pixel difference between the window and its surrounding windows is calculated once, ultimately forming a region containing... A set of difference data is used to obtain the set of thermal difference responses between windows.

[0062] S112: Based on the inter-window thermal difference response set, the range and variance of the pixel difference sequence in each image window are calculated to establish a difference statistical index. The statistical value in the window area is then offset and compared with the reference temperature difference of the overall background of the thermal imaging image, using the following formula:

[0063] ;

[0064] The thermal response deviation value of the image window region is obtained through calculation;

[0065] in, Indicates the first The thermal response deviation values ​​of each image window region are used to quantify the degree of thermal offset of that window relative to the background. Indicates the first The highest pixel heat value within each window is derived from image grayscale heat value sampling. Indicates the first The lowest pixel heat value in the neighborhood of each window is derived from the gray-level heat value sampling of adjacent windows in the same frame image. Indicates the first The first window The absolute value of the difference in thermal values ​​between groups of pixels. This represents the mean of the absolute values ​​of the differences within all windows in the current image. Indicates the first The number of pixel difference samples in each window, and the thermal response offset threshold is the value obtained by adding one standard deviation to the mean of the standard deviation of the pixel thermal response in the background area in the panoramic thermal image recorded during the initial flight phase of the UAV, which is used to distinguish the response difference boundary between the target and the background.

[0066] Based on the inter-window thermal response set, statistical calculations are performed on the pixel difference sequence associated with each image window in the set, using the window as the basis. For example, randomly select within this window The absolute value of the thermal difference between a group of pixels, i.e. The constructed sequence is Calculate the range of the sequence as follows: And calculate its variance, establish a difference statistical index, and set the window The statistical value in the region is offset and compared with the reference temperature difference of the overall background of the thermal imaging image. The calculation is performed using the following formula:

[0067] ;

[0068] in, Indicates the first The thermal response deviation values ​​of each image window region, where the subscript A unique number representing the window, from arrive This value is used for the quantization window. The degree of thermal shift relative to its neighborhood and the overall image background Indicates the first The highest pixel heat value within a window, its value is derived from the value of the highest pixel heat value within the window. The grayscale thermal values ​​of each pixel are obtained by sampling and traversing the data. Indicates the first The lowest pixel heat value in the neighborhood of a window, whose value is derived from the window A total of eight adjacent windows The grayscale thermal values ​​of each pixel are obtained by sampling and traversing the data. Indicates the first The first window is randomly selected The absolute value of the thermal difference between pixel pairs. Indicates all in the current image The arithmetic mean of the absolute values ​​of the differences within a window Indicates the first The number of pixel difference samples used to calculate the internal thermal value difference in each window; the logic of the formula lies in the first term under the square root. It characterizes the macroscopic thermal contrast between the window and its neighborhood; its square operation amplifies the significant temperature difference. The second term... This represents the comparison between the dispersion of heat value distribution within the window and the average dispersion of the entire image, i.e., the complexity of the "thermal texture" within the window. The square root of the sum of these two values ​​comprehensively evaluates the degree of thermal anomaly of a window at both macroscopic and microscopic levels. The setting of the thermal response offset threshold is used during the initial flight phase before the official start of the UAV search and rescue mission, continuously monitoring a confirmed inanimate background area (such as open ground). A flight scan of seconds, collecting a total of Frame thermal imaging image, calculate this The mean of the standard deviation of the thermal response of all background region pixels in the frame image; for example, the calculated mean is... The standard deviation is The value obtained by adding one standard deviation to the mean is: The threshold is set to Now in window Taking this as an example, we will calculate based on... , Set the number of samples ,window Inside Pixel difference samples for Set to view the entire map The mean of the absolute values ​​of the differences calculated from each window for Then substitute it into the formula:

[0069] ;

[0070] The advantage of this formula is that by combining the macroscopic thermal difference between windows with the microscopic thermal texture complexity within the window, it can more robustly identify thermal anomaly areas. Simply relying on the maximum heat value may misjudge small areas of solar reflective hot spots as targets, while combining the internal heat value variance can better distinguish between real heat sources and environmental noise, and calculate the thermal difference response deviation value of the image window area.

[0071] S113: Based on the thermal response deviation value, compare it with the thermal response offset threshold, filter the window numbers that exceed the thermal response offset threshold, extract the boundary coordinate information of the corresponding image region, perform region numbering and structure classification processing, and uniformly mark the region set as non-background response region to obtain the high thermal candidate positioning region set;

[0072] A series of calculated thermal difference response deviations, such as window values. The deviation value is Compare it with the preset thermal response offset threshold. In comparison, because The window The system determined that the threshold was exceeded and then changed the window number. Record this; this filtering process applies to all... The thermal response deviation value for each window is set, except for the window... Outside, window , , , The deviation values ​​are also all greater than The deviation values ​​of other windows are all smaller than 1. The set of window numbers selected is Next, extract these windows in Boundary coordinate information in a pixel image, if the window The coordinates of the top left corner are Its size is Then the pixel area it covers is and Since the five selected windows are spatially adjacent, the system performs structural classification and merges them into a single, larger image region. This new region is assigned a unique identifier, such as "Region A." This merging process calculates the smallest rectangular boundary that can completely enclose the five windows; for example, its top-left corner coordinates are [window coordinates missing]. The coordinates of the top left and bottom right corners of the window are In the lower right corner, the set of "Region A" is then uniformly marked as a non-background response region, which together with other regions that are also filtered and merged using this method constitutes a high-heat candidate localization region set.

[0073] Please see Figure 3 The specific steps for obtaining the suspected target thermal morphology region are as follows:

[0074] S211: Based on the high-heat candidate localization region set, extract the maximum and minimum heat value pixels in each region, construct heat value interval boundaries, and perform interval judgment on each interval boundary according to the set human body surface temperature reference range. Select image regions whose heat value intervals fall within the human body surface temperature reference range to obtain the target region set that meets the heat reference interval.

[0075] Based on the obtained set of high-heat candidate localization regions, "Region A" is processed. First, within the five windows contained in "Region A" (total... In a given set of pixels (n pixels), iterate through the heat values ​​of all pixels and extract the pixel with the highest heat value, whose grayscale value is [value missing]. And the pixel with the lowest heat value, whose grayscale value is These two values ​​constitute the boundary of the calorific value range for "Region A". Next, based on the established reference range for human body surface temperature, interval judgment is performed. This reference range is set based on the following experimental verification process: at different ambient temperatures ( , , Under these conditions, a thermal imaging camera of the same model as the drone is used to... Body surface temperature was measured in participants with different vital signs at rest and during mild activity. The measurements, taken on the forehead, neck, and arms, showed that in most search and rescue scenarios, the effective surface thermal radiation signal of survivors corresponds to a range of values ​​on the thermal imaging grayscale image within the following ranges. to Between (corresponding to approximately) to Therefore, the reference range for human body surface temperature is set as follows: Subsequently, the calorific value range of "Region A" was analyzed. and reference range Compare the two intervals to determine if they intersect. It is their intersection, and the intersection is not empty, so "Region A" is selected as the image region that meets the conditions. This process is applied to all high-heat candidate positioning regions, and finally all regions that overlap with the human body surface temperature reference range are selected to obtain the target region set that meets the thermal reference range.

[0076] S212: Call the target region set that conforms to the thermal reference interval, perform coordinate mapping and numerical extraction on the effective pixels with recorded thermal values ​​within each image region, construct a set of two-dimensional coordinate and thermal value pairs, calculate the thermal gradient difference between adjacent pixels in the horizontal and vertical directions respectively, estimate the continuous trend of pixel response within the region according to the linear increment rule, construct the thermal value distribution function surface, and generate thermal gradient fitting distribution information.

[0077] The system retrieves a target region set that conforms to the thermal reference interval. Taking "Region A" as an example, it performs coordinate mapping and numerical extraction on all valid pixels with recorded thermal values ​​within the region, constructing a set containing... A data set of n elements, each element being of the form of... ,in These are the coordinates of a pixel within the entire image frame. This corresponds to the grayscale heat value, forming a set of two-dimensional coordinate and heat value pairs. Then, the heat value gradient difference between adjacent pixels is calculated for both the horizontal and vertical directions. For example, for a pixel... Its lateral gradient difference is The longitudinal gradient difference is Then, estimation is performed within the pixel grid of "Region A" according to a linear increment rule. Specifically, for any two pixels with known heat values, the heat value between them is assumed to change linearly. This is used to infer the continuous trend of pixel response within the region. Through this process, the discrete pixel heat value data is transformed into a continuous function surface, namely the heat value distribution function surface. This surface, in three-dimensional space, is... As the bottom surface, with calorific value The system provides a high-level, intuitive view of the heat distribution within "Region A" and generates thermal gradient fitting distribution information.

[0078] S213: Based on the thermal gradient fitting distribution information, identify pixel thermal value hole regions in the fitted surface caused by missing, occluded or noise suppression, and perform bilinear interpolation on the neighborhood around the hole to fill the thermal value missing. In the filled image area, extract the closed boundary curve according to the thermal value change gradient direction, construct the contour structure map, and obtain the suspected target thermal morphology area.

[0079] Based on the generated thermal gradient fitting distribution information, the surface of the heat value distribution function of "Region A" is analyzed to identify pixel heat value voids caused by tree branch occlusion or sensor noise suppression algorithms. These regions appear on the surface as abnormal depressions or missing data in the heat value, for example, in the coordinate... A place was found The pixel region has a much lower heat value than its surrounding pixels, forming a "cold hole." The system then processes the neighborhood surrounding the hole, such as a... Within the pixel range, perform bilinear interpolation; specifically, extract the hole points. The heat values ​​of the four nearest non-hole pixels are calculated using a weighted average. The process of filling in the missing heat value is repeated for all hole pixels to fill in the missing heat value. After the filling is completed in the image area, the system extracts isotherms in the filled "region A" according to the gradient direction of the heat value from high to low, that is, the opposite direction of heat diffusion. Then, it connects one or more of the isotherms that best represent the human body contour into a closed boundary curve. This process is achieved by setting a gradient threshold and connecting only the points where the gradient change exceeds the threshold, thereby constructing the contour structure map of the target and obtaining the suspected target thermal morphology area.

[0080] Please see Figure 4 The specific steps for obtaining the effective hot target candidate region set are as follows:

[0081] S311: Based on the geometric boundary coordinates of each contour structure in the suspected target thermal morphology region, extract the positioning point coordinates of each target in consecutive image frames, construct the spatial position change trajectory within the frame sequence, record the pixel displacement direction and amplitude between adjacent frames, calculate the multi-frame path vector, and obtain the contour trajectory direction parameter set.

[0082] Based on the acquired suspected target thermal morphology regions, extract "Region A" in the current frame (e.g., frame number 1). The geometric boundary coordinates in the frame image are calculated, and its centroid is used as the positioning point coordinates of that frame, set as... The drone continued flying, in the following consecutive image frames, such as the first... Frame to the In each frame, the entire processing flow from S111 to S213 is repeated to continuously identify and locate "Region A" and obtain its location point coordinate sequence in these frames:

[0083]

[0084] ;

[0085] This coordinate sequence constitutes the spatial position change trajectory of "region A" within the frame sequence. Next, the pixel displacement direction and magnitude between adjacent frames are recorded; for example, from frame number 1... Frame to the Frame, displacement vector is Its amplitude is Pixels, direction is upper right By calculating all displacement vectors between consecutive frames, for example This forms a multi-frame path vector set, and the contour trajectory direction parameter set is obtained.

[0086] S312: Call the contour trajectory direction parameter set, combine it with the corresponding frame number, count the number of times each coordinate point appears, calculate the number of frames in which the target exists in the image sequence, extract the thermal response value change sequence within the corresponding time period, calculate the average temperature change rate based on the temperature increase and decrease trend between frames, construct continuous thermal dynamic features bound to each target contour, and obtain target thermal time change information.

[0087] The obtained contour trajectory direction parameter set of "Region A" is retrieved and combined with the corresponding frame number. The frequency of occurrence of the centroid coordinates of "Region A" is counted. Since "Region A" appears from the 1st... Frame to the Frame continuity It was detected in all frames, therefore the number of frames it exists in is [number missing]. At the same time, extract "Region A" in this... The sequence of thermal response value changes within a frame time period is shown in Table 1. The average thermal value of "Region A" in each frame is recorded.

[0088] Table 1 Thermal dynamic characteristics of continuous frames in region A

[0089]

[0090] As shown in Table 1, based on the inter-frame temperature increase / decrease trend, i.e. Calculate its average rate of temperature change, i.e. Grayscale value / frame, by having a number of frames ( (frames) and average temperature change rate ( The parameters (grayscale value / frame) are bound to the contour of "Region A" to construct the continuous thermal dynamic characteristics of the target and obtain the thermal time change information of the target.

[0091] S313: Based on the target thermal time change information, the three feature parameters of trajectory direction vector, duration frame number and temperature change rate are normalized to construct a three-dimensional feature vector. The three-dimensional feature vector is then matched with the perturbation path set in the pre-stored thermal background perturbation trajectory table using Euclidean distance. Background perturbation items with similarity higher than the similarity threshold are filtered out to obtain a set of effective thermal target candidate regions.

[0092] The process of normalizing the three feature parameters of trajectory direction vector, number of consecutive frames, and temperature change rate is as follows: the trajectory direction vector in each image frame is represented as a two-dimensional coordinate difference sequence, the number of consecutive frames is normalized proportionally according to the total number of frames, and the temperature change rate is normalized based on the average change of pixel heat value per unit time.

[0093] The similarity threshold is the distance boundary value calculated during the training phase based on the mean and standard deviation of the Euclidean distance of the historical perturbation paths in the thermal background perturbation trajectory table. The distance boundary value is the standard deviation of the mean weighted by a set ratio.

[0094] Based on the acquired target thermal time change information, the three feature parameters of "Region A"—trajectory direction vector, number of frames, and temperature change rate—are normalized. First, the trajectory direction vector is represented as a two-dimensional coordinate difference sequence. For example, for a frame... to Its average displacement vector is Secondly, the continuous frame count is proportionally normalized based on the total number of frames currently being processed. If the total number of frames processed is... If the frame is a normalized duration of frames, then the normalized duration of frames is... Secondly, the rate of temperature change is determined based on a unit time (e.g., per second). The average pixel thermal value change of each frame is normalized, and the average temperature change rate of all targets within this time period is set to [value missing]. If the grayscale value is equal to the frame, then the normalized rate is... This allows us to construct the three-dimensional feature vector of "Region A". (This is a four-dimensional example; in practical applications, the two-dimensional vector can be reduced to a smaller dimension.) This three-dimensional feature vector is then matched with the perturbation path set in a pre-stored thermal background perturbation trajectory table using Euclidean distance. This trajectory table stores feature vectors of non-target thermal signals such as wind-blown leaves and flowing streams in historical data. The similarity threshold is set during the training phase based on the thermal background perturbation trajectory table. The eigenvectors of the historical perturbation paths are used to calculate the Euclidean distance between each pair of them, and the mean distance is obtained. The standard deviation is The distance from the boundary value is set as the standard deviation of the mean-weighted average, with the weighting ratio set according to the stringency of interference suppression. Then the threshold is If the feature vector of "region A" matches the feature vector of a certain entry in the trajectory table, such as "wind blowing and grass moving A", The Euclidean distance calculation result is This value is below the similarity threshold. If the calculated distance is higher than the threshold, it will not be filtered out. If the calculated distance is higher than the threshold, it will be filtered out. This method is used to filter out all background perturbation terms with similarity higher than the similarity threshold and obtain a set of effective hot target candidate regions.

[0095] Please see Figure 5 The specific steps for obtaining the abnormal trajectory target identification marker set are as follows:

[0096] S411: Based on the effective thermal target candidate region set, call the maximum heat value, minimum heat value and corresponding frame number information recorded in each region in consecutive frames, calculate the heat value difference between consecutive frames, normalize the region time span, calculate the temperature change amplitude per unit time, and obtain the heat value inter-frame change rate set.

[0097] Based on the acquired set of valid thermal target candidate regions, "Region A" is invoked in consecutive frames (the first frame). Frame to the The maximum and minimum calorific values ​​and corresponding frame number information recorded in the frame (e.g., the first frame) The maximum heat value of a frame is , No. The maximum heat value of a frame is Calculate the thermal difference between these two frames as follows: Next, the time span of the region is normalized, since the total observation time span is... Frames, while the drone's frame rate is Frames per second, i.e., total duration is Seconds, then the unit of time (set to) The temperature change within seconds is This calculation is applied to all adjacent frame pairs to obtain a rate of change sequence, for example... Finally, these rates of change are integrated to obtain a set of inter-frame rates of change of heat value.

[0098] S412: Call the heat value inter-frame change rate set, combine it with the number of pixels in the corresponding region of each image frame, and calculate the area change value between consecutive frames using the formula:

[0099] ;

[0100] The jump trend of the hot target area is obtained by calculation and compared with the occlusion jump threshold. If the jump trend of the hot target area is greater than the occlusion jump threshold, it is marked as meeting the abnormal change standard and a jump sensitive area number set is generated.

[0101] in, Indicates the first The abrupt change trend of the hot target region in the frame. Indicates the first The normalized maximum heat value of the target region in the frame is obtained by linearly normalizing the maximum heat value of the corresponding region in the image. Indicates the first The maximum heat value of the target region in the frame after normalization. Indicates the first The normalized value of the pixel area of ​​the target region in the frame is obtained by normalizing the number of pixels according to the maximum target area of ​​the entire image. Indicates the first The normalized value of the pixel area of ​​the target region in the frame. This represents the regional thermal boundary perturbation factor, derived from the ratio of the mean of the regional thermal boundary gradient change to the variance of the background thermal gradient. It is used to adjust the sensitivity of the area change term in the trend value. This represents a small positive constant to prevent the denominator from being zero. It is set to a value within the allowable lower limit of error. The occlusion jump threshold is a judgment boundary set based on the average of the historical jump trends of stable target areas in multiple frames plus three times the standard deviation. It is used to distinguish whether the area jump belongs to occlusion interference behavior.

[0102] The obtained set of inter-frame heat value change rates is retrieved, and combined with the number of pixels in "Region A" in each image frame (i.e., the area), the area change value between consecutive frames is calculated using the formula:

[0103] ;

[0104] Perform the calculation, where, Indicates the first The abrupt change trend of the hot target region in the frame. Indicates the first The normalized maximum heat value of the target region in the frame is obtained by linearly normalizing the maximum heat value of the corresponding region in the image, for example, by adjusting the heat value range. Mapped to , Indicates the first The maximum heat value of the target region in the frame after normalization. Indicates the first The normalized value of the pixel area of ​​the target region in the frame is derived from the number of pixels divided by the maximum possible target area in the entire image (e.g., (Pixels) normalized Indicates the first The normalized value of the pixel area of ​​the target region in the frame. This represents the regional thermal boundary perturbation factor, whose value is derived from the ratio of the mean of the regional thermal boundary gradient change to the variance of the background thermal gradient. It is used to adjust the sensitivity of the area change term in the trend value. This represents a small positive constant to prevent the denominator from being zero, and is set as follows: The logic of this formula lies in the fact that the numerator combines the logarithmic ratio of calorific value change and the linear difference of area change. The logarithmic ratio smooths out drastic fluctuations in calorific value, while the area change directly reflects the target's size change. The sum of the two can comprehensively judge the target's stability. The denominator uses the square root of the area to suppress excessively high jump values ​​generated by large-area targets when the area changes slightly. The setting of the occlusion jump threshold is based on... The historical jump trend values ​​of stable target regions (such as stationary vehicle engines) in different frame segments were statistically analyzed, and their mean was calculated. The standard deviation is The judgment threshold is set as the mean plus three standard deviations, i.e. Now, let's start with the first... Frame (i.e.) Taking () as an example, let's calculate and set the first () In the frame, the maximum heat value of "Region A" is The area is Pixel, the Within the frame, due to partial occlusion, the maximum heat value is reduced to The area was reduced to Pixels, after normalization , , , Regional thermal boundary perturbation factor Calculated based on the image characteristics of the current frame Substitute into the formula:

[0105] ;

[0106] The advantage of this formula lies in its ability to effectively distinguish between drastic changes caused by actual occlusion and normal fluctuations caused by the target itself or the environment, by comprehensively considering the relative changes in calorific value and area, and by using logarithms and square roots for scale adjustment. The calculation results... Less than the occlusion transition threshold Therefore, this change is marked as not meeting the abnormal change criteria. If another frame calculates... Value greater than If it meets the abnormal change criteria, then its frame number and region number are recorded to generate a set of change-sensitive region numbers.

[0107] S413: Based on the set of sensitive jump regions, retrieve the thermal contour line closure status of the corresponding region in the image, perform boundary integrity judgment on each thermal contour line, extract the curve closure rate as a structural stability index, mark the unstable region of the structural boundary and assign it a status code, and establish an abnormal trajectory target recognition tag set.

[0108] Based on the generated set of transition-sensitive region numbers, it is set that "Region A" is included in the number... The system then retrieves the recorded frame number. The closed state of the thermal contour lines in "Region A" in the frame image is determined by generating multiple thermal contour lines from its highest thermal value to the boundary thermal value within "Region A". For each thermal contour line, a boundary integrity determination is performed. This determination process specifically involves calculating the ratio between the actual length of the curve and the perimeter of the convex hull of the area it encloses. A completely closed and regularly shaped contour has a ratio close to... The proportion of contours with broken boundaries due to occlusion is significantly smaller than that of contours with broken boundaries due to occlusion. For example, in the first In the frame, the core thermal contour closure rate of "Region A" is It exhibits structural stability, but during the transition... In a frame, due to the intrusion of a cryogenic obstruction, the core thermal contour lines break into two segments, causing the calculated closure rate to plummet. This closure ratio is extracted as a structural stability indicator, because Below the preset structural stability lower limit (e.g.) (This value was derived through extensive statistical analysis of unobstructed human targets), and the system then places "Region A" in the [missing value]. The frame is marked as structural boundary instability and assigned a specific identification status code, such as "E-02" which represents "suspected partial occlusion". Finally, all such tags are integrated to establish an abnormal trajectory target identification tag set.

[0109] Please see Figure 6 The specific steps for obtaining the search and rescue target relocation trigger signal are as follows:

[0110] S511: Based on the abnormal trajectory target recognition marker set, the region marked as the one that has changed is extracted from the thermal response matrix of the corresponding region in the current image frame. Pixels with a heat value greater than the median heat value are searched within the region. The centroid coordinates of the hot spot are calculated and a weighted value is assigned to the heat value of the hot spot. A two-dimensional weighted average is performed to obtain the focal point of the region's thermal intensity, and the set of weighted centroid coordinates of the thermal response is obtained.

[0111] Based on the established abnormal trajectory target identification tag set, it is marked as the first... For the frame where a jump occurs in "region A", the thermal response matrix corresponding to "region A" in the current frame is extracted. This matrix contains the heat value data of all pixels within "region A". Then, pixels with heat values ​​greater than the median heat value within the region are searched and considered as hotspots. For example, if "region A" is in the [frame transition], the heat response matrix of "region A" is extracted. This matrix contains the heat value data of all pixels within "region A". The median heat value of the frame is Then all calorific values ​​higher than All pixels were extracted. Then, the centroid coordinates of these hotspots were calculated, and a weighted value was assigned to the heat value of each hotspot. This weighted value is proportional to the heat value itself; that is, points with higher heat values ​​have a greater weight in the centroid calculation. A two-dimensional weighted average was then applied to the coordinates of all hotspots. , ,in These are hotspot coordinates. Based on its calorific value, the focal point of thermal intensity in that region is ultimately calculated, and the calculated coordinates are set as follows: This set of coordinates is the weighted center coordinate set of the thermal response.

[0112] S512: Call the weighted centroid coordinate set of thermal response, construct a square region window with a side length twice that of the original target outline with each centroid coordinate as the center, and extract the infrared thermal imaging image, visible light image and heat value data frame corresponding to the window, construct the intra-frame target rescan information structure indexed by spatial thermal density, and obtain the region rescan window data block set.

[0113] The obtained weighted center coordinate set of the thermal response is called, with "Region A" in the first position. Centroid coordinates calculated from the frame Construct a square region window centered on "Region A" with its side length set to the value of "Region A" before any transition occurs (e.g., the first transition). The outline of the frame is twice the side length of the bounding rectangle. If the original bounding rectangle size is For pixels, take the longer side. Twice the number of pixels, that is Pixels are used as the side length of the new window to construct a A window of pixels, and using this window to capture the 1st pixel. The infrared thermal imaging image corresponding to the frame, the visible light image simultaneously captured by the drone, and the original thermal value data frame are integrated into a single data structure, and the spatial thermal density, i.e., the barycentric coordinates, are used as the data source. As an index, an intra-frame target rescan information structure is constructed. This process is applied to all targets marked as abnormal to obtain the set of region rescan window data blocks.

[0114] S513: Based on the area rescan window data block set, write the corresponding infrared thermal image, visible light image and centroid positioning coordinates into the search and rescue intermediate interaction frame structure, and send them synchronously to the front-end rescue node as the target information to be confirmed. Encode the target data and register the status field, and establish a search and rescue target relocation trigger signal.

[0115] Based on the acquired set of region rescan window data blocks, it will be compared with "Region A" in the [missing information]. Frame-related Pixel infrared thermal image, corresponding visible light image, and centroid positioning coordinates Together, they are written into a search and rescue intermediate interaction frame structure, which is a standardized data packet. Subsequently, the data packet is synchronously sent to the ground command center or handheld terminal and other front-end rescue nodes via wireless data link as a target highlighting information awaiting manual confirmation. At the same time, the system encodes the target data of "Area A", for example, assigning an ID "T-001", and registers its status field as "Pending-Confirmation" in the database, establishing a search and rescue target relocation trigger signal.

[0116] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting visual targets using UAV thermal imaging for search and rescue missions, characterized in that, Includes the following steps: S1: Acquire multiple frames of thermal imaging images collected by the UAV in the search and rescue mission area. Based on the degree of thermal difference deviation from the background thermal equilibrium baseline temperature difference, mark the image sub-regions with non-background features and generate a set of high-thermal candidate positioning regions. S2: Based on the set of high-heat candidate positioning regions, determine whether the interval heat value meets the fitting condition. If it does, perform linear gradient fitting processing on the pixel heat value in the interval to reconstruct the suspected human body contour region with continuous boundary in the region and generate the suspected target thermal morphology region. S3: Based on the spatial position of each contour region in the suspected target thermal morphology region, obtain the coordinate change trajectory in continuous image frames, and compare it with the path record in the constructed thermal background disturbance trajectory table to filter out background disturbance hotspot regions and generate a set of effective thermal target candidate regions. S4: Call the set of effective thermal target candidate regions, calculate the rate of change of thermal value and the rate of change of pixel area of ​​the region between frames, determine whether there is local occlusion or target switching, obtain the thermal contour closure of the corresponding region in the frame as an auxiliary indicator, and generate an abnormal trajectory target identification marker set. S5: Based on the abnormal trajectory target identification marker set, the hot target area that has changed is marked as the hot target area. The position and thermal response distribution of the area in the current frame are called. The hot spot dense center point is extracted and weighted center positioning processing is performed. The in-frame area rescan window is constructed. The normal image of the area, the thermal imaging image and the area positioning information are sent to the rescue personnel. The search and rescue target relocation trigger signal is generated. The specific steps for obtaining the search and rescue target relocation trigger signal are as follows: S511: Based on the regions marked as having abrupt changes in the abnormal trajectory target identification marker set, extract the thermal response matrix of the corresponding region in the current image frame, and search for pixels with heat values ​​greater than the median heat value within the region. Calculate the centroid coordinates of the hotspots and assign weighted values ​​to the heat values ​​of the hotspots. Perform a two-dimensional weighted average to obtain the focal point of the region's thermal intensity, and obtain the set of thermal response weighted centroid coordinates. S512: Call the thermal response weighted centroid coordinate set, construct a square region window with a side length twice that of the original target outline with each centroid coordinate as the center, and extract the infrared thermal imaging image, visible light image and heat value data frame corresponding to the window, construct the intra-frame target rescan information structure indexed by spatial thermal density, and obtain the region rescan window data block set. S513: Based on the set of data blocks in the rescan window of the region, write the corresponding infrared thermal image, visible light image and centroid positioning coordinates into the search and rescue intermediate interactive frame structure, and send them synchronously to the front-end rescue node as the target information to be confirmed. Encode the target data and register the status field, and establish a search and rescue target relocation trigger signal. The search and rescue target relocation trigger signal includes the thermal response centroid coordinate point, regional rescan window parameters, trigger condition judgment value, and target detection module activation command.

2. The UAV thermal imaging visual target detection method for search and rescue missions according to claim 1, characterized in that, The high-heat candidate location region set includes multi-window significant thermal difference blocks, regional thermal response standard deviation, thermal difference deviation from background reference value mapping matrix, and boundary coordinates of abnormal thermal difference regions. The suspected target thermal morphology region includes continuous image blocks of thermal value gradient, thermal contour distribution after linear fitting, bilinear interpolation completed pixel layer, and potential target morphology closed boundary in the fitted block. The effective thermal target candidate region set includes thermally stable zone trajectory set, spatial target location set after excluding disturbances, thermal response continuous feature data, and non-background matching identifier coordinate points. The abnormal trajectory target identification marker set includes thermal value jump identification marker, occlusion behavior inference marker, area abrupt change region number, and morphology closure judgment marker.

3. The UAV thermal imaging visual target detection method for search and rescue missions according to claim 2, characterized in that, The specific steps for obtaining the high-heat candidate location region set are as follows: S111: Acquire multiple frames of thermal imaging images collected by the UAV in the search and rescue mission area. Divide each frame of the image into a fixed-size window area using a regular grid. Extract the pixel difference between the maximum thermal value pixel in each window and the minimum thermal value pixel in the corresponding surrounding window to obtain the thermal response difference in the window area and obtain the set of thermal difference responses between windows. S112: Based on the set of thermal difference responses between windows, the range and variance of the pixel difference sequence in each image window are calculated to establish a difference statistical index. The statistical value in the window area is offset and compared with the reference temperature difference of the overall background of the thermal imaging image. The thermal difference response deviation value of the image window area is calculated and obtained. S113: Based on the thermal response deviation value, compare it with the thermal response offset threshold, filter the window numbers that exceed the thermal response offset threshold, extract the boundary coordinate information of the corresponding image region, perform region numbering and structure classification processing, and uniformly mark the region set as non-background response region to obtain the high-heat candidate positioning region set.

4. The UAV thermal imaging visual target detection method for search and rescue missions according to claim 3, characterized in that, The specific steps for obtaining the suspected target thermal morphology region are as follows: S211: Based on the set of high-heat candidate positioning regions, extract the maximum and minimum heat value pixels in each region, construct heat value interval boundaries, and perform interval judgment on each interval boundary according to the set human body surface temperature reference range. Select image regions whose heat value intervals fall within the human body surface temperature reference range to obtain a target region set that conforms to the heat reference interval. S212: Call the target region set that conforms to the thermal reference interval, perform coordinate mapping and numerical extraction on the effective pixels with recorded thermal values ​​inside each image region, construct a set of two-dimensional coordinate and thermal value pairs, calculate the thermal value gradient difference between adjacent pixels in the horizontal and vertical directions respectively, estimate the continuous trend of pixel response in the region according to the linear increment rule, construct the thermal value distribution function surface, and generate thermal gradient fitting distribution information. S213: Based on the thermal gradient fitting distribution information, identify pixel thermal value void regions in the fitted surface caused by missing, occluded or noise suppression, and perform bilinear interpolation on the neighborhood around the void to fill the thermal value missing. In the filled image area, extract the closed boundary curve according to the thermal value change gradient direction, construct the contour structure map, and obtain the suspected target thermal morphology region.

5. The UAV thermal imaging visual target detection method for search and rescue missions according to claim 4, characterized in that, The specific steps for obtaining the effective thermal target candidate region set are as follows: S311: Based on the geometric boundary coordinates of each contour structure in the suspected target thermal morphology region, extract the positioning point coordinates of each target in consecutive image frames, construct the spatial position change trajectory within the frame sequence, record the pixel displacement direction and amplitude between adjacent frames, calculate the multi-frame path vector, and obtain the contour trajectory direction parameter set. S312: Call the contour trajectory direction parameter set, combine it with the corresponding frame number, count the number of times each coordinate point appears, calculate the number of frames in which the target exists in the image sequence, extract the thermal response value change sequence within the corresponding time period, calculate the average temperature change rate according to the temperature increase and decrease trend between frames, construct continuous thermal dynamic features bound to each target contour, and obtain target thermal time change information. S313: Based on the target thermal time change information, the three feature parameters of trajectory direction vector, duration frame number and temperature change rate are normalized to construct a three-dimensional feature vector. The three-dimensional feature vector is then matched with the perturbation path set in the pre-stored thermal background perturbation trajectory table using Euclidean distance. Background perturbation items with similarity higher than the similarity threshold are filtered out to obtain a set of effective thermal target candidate regions.

6. The UAV thermal imaging visual target detection method for search and rescue missions according to claim 5, characterized in that, The process of normalizing the three feature parameters of trajectory direction vector, number of consecutive frames, and temperature change rate is specifically as follows: the trajectory direction vector in each image frame is represented as a two-dimensional coordinate difference sequence, the number of consecutive frames is normalized proportionally according to the total number of frames, and the temperature change rate is normalized based on the average change of pixel heat value per unit time. The similarity threshold is a distance boundary value calculated during the training phase based on the mean and standard deviation of the Euclidean distance of historical perturbation paths in the thermal background perturbation trajectory table. The distance boundary value is the standard deviation of the mean weighted by a set ratio.

7. The UAV thermal imaging visual target detection method for search and rescue missions according to claim 6, characterized in that, The specific steps for obtaining the abnormal trajectory target identification marker set are as follows: S411: Based on the set of effective thermal target candidate regions, call the maximum heat value, minimum heat value and corresponding frame number information recorded in each region in consecutive frames, calculate the heat value difference between consecutive frames, normalize the region time span, calculate the temperature change amplitude per unit time, and obtain the heat value inter-frame change rate set. S412: Call the set of heat value inter-frame change rate, combine it with the number of pixels in the corresponding region in each image frame, count the area change value between consecutive frames, calculate the jump trend of the heat target region, and compare it with the occlusion jump threshold. If the jump trend of the heat target region is greater than the occlusion jump threshold, mark it as meeting the abnormal change standard and generate a jump sensitive region number set. S413: Based on the set of abrupt change sensitive regions, retrieve the thermal contour line closure status of the corresponding region in the image, perform boundary integrity judgment on each thermal contour line, extract the curve closure rate as a structural stability index, mark the unstable region of the structural boundary and assign it an identification status code, and establish an abnormal trajectory target identification tag set.

Images