Binocular thermal imaging image fusion method based on dynamic compensation

Abstract

The invention discloses a binocular thermal imaging image fusion method based on dynamic compensation, and relates to the technical field of infrared imaging, a dynamic compensation operation module is constructed, a pixel displacement matrix of a right eye image is generated according to a real-time position deviation value, real-time geometric correction is performed on the right eye image before the image is output, and the image fusion precision is improved. Through a real-time dynamic registration and progressive fusion mechanism, the problem of ghosting in binocular thermal imaging observation is effectively solved, a thermal radiation feature extraction and cross-channel synchronous analysis technology is adopted, the feature recognition bottleneck of a traditional visible light registration method in a thermal imaging scene is broken through, and the real-time dynamic registration and cross-channel synchronous analysis technology is achieved. In combination with a dynamic offset compensation and temperature adaptive progressive fusion algorithm, sub-pixel-level parallax correction is realized while the integrity of original thermodynamic data is maintained, and continuous ghosting interference caused by binocular image space deviation during human eye observation is eliminated. And real-time observation requirements of targets with different distances can be met without modifying an optical hardware structure.

Description

Technical Field

[0001] The present invention relates to the field of infrared imaging technology, and in particular to a binocular thermal imaging image fusion method based on dynamic compensation. Background Art

[0002] In the field of binocular thermal imaging equipment, existing technologies generally address binocular parallax using fixed optical compensation or offline parameter calibration. Traditional solutions, which mechanically adjust the parallelism of the two optical axes (such as the dual-mirror synchronous adjustment mechanism employed in Patent Publication No. CN112229434A), can compensate for initial assembly errors but cannot adapt to dynamic parallax shifts caused by changes in target distance or device vibration during observation. Some improved solutions attempt to use visible light image registration principles (such as the grayscale feature matching algorithm proposed in Patent Publication No. CN113763415A). However, the inherent low-texture nature of thermal imaging results in an insufficient capture rate of valid feature points, particularly in uniform temperature fields, where the matching success rate is less than 35%. Furthermore, existing ghosting reduction technologies often rely on post-processing (such as the stereo image fusion algorithm disclosed in Patent Publication No. CN114494424A), resulting in a single-frame processing latency of 80-120ms, making them difficult to meet real-time observation requirements. What is more serious is that when observing dynamic high-temperature targets (such as the movement of local hot spots in industrial equipment), traditional methods lack temperature-space dual-dimensional correlation analysis, which leads to a mismatch between compensation parameters and the target's thermodynamic characteristics, resulting in compensation lag or over-correction. In actual measurements, there is still a residual parallax of 0.5-2.3 pixels, causing continuous ghosting interference when observed by the human eye. Summary of the Invention

[0003] In order to overcome the above-mentioned shortcomings, the present invention aims to provide a binocular thermal imaging image fusion method based on dynamic compensation to solve the above-mentioned technical solutions.

[0004] To achieve the above object, the present invention provides the following technical solutions: A binocular thermal imaging image fusion method based on dynamic compensation includes the following steps: S100: Setting a thermal radiation feature extraction unit in each of the left and right image processing channels to identify high temperature gradient mutation areas in the thermal imaging image in real time as reference feature points; S200: Establishing a feature point synchronization comparison mechanism between binocular channels, and calculating the spatial position deviation of corresponding feature points in the left and right images through a coordinate mapping algorithm; S300: Build a dynamic compensation calculation module to generate a pixel displacement matrix of the right eye image based on the real-time position deviation value, and perform real-time geometric correction on the right eye image before outputting the image; S400: Using a progressive optimization algorithm to achieve dynamic adaptation of the binocular field of view by tracking feature point trajectories in consecutive frames, the algorithm includes a weight adjustment mechanism based on temperature distribution changes; S500: A final fusion processing unit is set before the video output interface to perform edge smoothing and contrast equalization processing on the corrected binocular image.

[0005] As a further solution of the present invention: step S100 includes: S110: Analyze the temperature field of the thermal imaging images in the left and right image processing channels frame by frame, calculate the temperature change rate in units of pixel matrix, and select continuous areas where the temperature change rate exceeds a set threshold; S120: Perform morphological optimization on the selected high-temperature gradient areas, remove discrete noise points with an area smaller than 10×10 pixels, and retain thermal radiation feature areas with stable edge contours; S130: Based on the temperature gradient direction consistency judgment, the direction vector of the pixel cluster in the feature area is calibrated, and the feature point whose gradient direction has an angle with the visual axis with a deviation of less than 15° is selected as the valid reference point; S140: Establish a dynamic association list of feature points of the left and right channels, and record the temperature extreme value coordinates and spatial distribution topological relationship of each feature point.

[0006] As a further solution of the present invention: Step S200 includes the following steps: S210: uniformly converting the coordinates of the feature points of the left and right channels into a polar coordinate system with the center of the binocular field of view as the origin, and performing normalization processing to eliminate the initial installation position deviation between the two channels; S220: Based on the temperature extreme value distribution characteristics of the feature points, perform rough matching on the feature points in the same temperature range (±3°C) in the left and right images, setting the initial search radius to 15% of the field of view width; S230: Using a sliding window matching algorithm, the spatial distribution pattern of feature points is compared pixel by pixel within the coarse matching range, and feature point pairs that meet a similarity threshold (≥85%) are calculated; S240: Fitting the geometric transformation matrix of the left and right feature point sets by the least squares method, and outputting the real-time translation deviation Δx, Δy and rotation angle deviation θ between the binocular images; S250: Establish a dynamic error correction mechanism to automatically trigger recalibration of feature point matching parameters when the fluctuation amplitude of Δx or Δy exceeds ±5 pixels in 5 consecutive frames.

[0007] As a further solution of the present invention: Step S300 includes the following detailed steps: S310: Receive the real-time translation deviations Δx, Δy and rotation angle deviation θ output in step S200, and generate a pixel displacement matrix including a horizontal displacement compensation amount (Δx'=Δx×k1), a vertical displacement compensation amount (Δy'=Δy×k2), and a rotation compensation amount θ'=θ×k3, where k1 and k2 are displacement attenuation coefficients (0.8≤k1, k2≤1.2), and k3 is a rotation correction coefficient (0.5≤k3≤1.5); S320: dynamically remapping the right eye image pixels based on the pixel displacement matrix, and filling the empty pixel areas generated after the displacement using a bilinear interpolation algorithm to preserve the integrity of the original temperature data; S330: Set a displacement threshold constraint mechanism. When the absolute value of Δx' or Δy' in a single frame exceeds 5% of the field of view width, enable the frame-by-frame progressive compensation mode and spread the total compensation amount over the next 3-5 frames. S340: Embed verification markers in the geometrically corrected right-eye image, and verify the correction effect by spatially overlapping with the corresponding markers in the left-eye image. If the deviation exceeds 2 pixels, re-matching in step S200 is triggered. S350: Establish a compensation amount history database, and dynamically adjust the coefficient values ​​of k1, k2, and k3 according to the compensation amount change trend of 10 consecutive frames, with the adjustment range not exceeding ±20% of the current value.

[0008] As a further solution of the present invention: Step S400 includes the following steps: S410: Within a time window of 5-10 consecutive frames, track the successfully matched feature points in the left and right fields of view, calculate the motion vector of each feature point, and select a stable feature point set with a motion direction consistency of ≥75%; S420: Dynamically assign weight coefficients based on the temperature change rate of the region where the feature point is located. A high-temperature dynamic region (temperature change rate ≥ 3°C / s) is assigned a weight of W1 = 0.2-0.5, and a low-temperature stable region (temperature change rate ≤ 1°C / s) is assigned a weight of W2 = 0.8-1.2. S430: Perform weighted fusion on the spatial distribution of feature points in the left and right fields of view based on the weight coefficient to generate binocular field of view adaptation parameters, including the parallax compensation baseline value L = Σ(Wᵢ×Δxᵢ) / ΣWᵢ and the viewing angle convergence factor α = 1-(Δθ_max / 30°), where Δθ_max is the maximum angular deviation of the feature points; S440: Adopting a sliding time window update mechanism, the adaptation parameters are updated every 3 frames. When the difference between the L values ​​calculated twice exceeds 15% of the baseline threshold, the time window is automatically expanded to 8-12 frames for smooth transition. S450: During the fusion process, the success rate of feature point tracking is monitored in real time. If the proportion of feature points that fail to be tracked for three consecutive frames is ≥40%, the feature point re-extraction in step S100 and the matching parameter reset in step S200 are triggered.

[0009] As a further solution of the present invention: Step S500 includes the following steps: S510: Performing edge fusion processing based on temperature gradient on the corrected left and right eye images. By calculating the second-order derivative of the temperature difference between adjacent pixels, the boundary line with the absolute value of the temperature gradient ≥ 0.5°C / pixel is extracted, and the edge smoothing transition is performed in the overlapping area of ​​the binocular field of view. S520: Using dynamic histogram mapping technology, the global temperature distribution range (T_min to T_max) of the left and right images is divided into N equal segments (N = 5-8). Based on the pixel ratio of each temperature segment in the left eye image, the right eye image is subjected to segment-by-segment contrast stretching and matching. S530: A difference detection module is embedded in the fused image to calculate the temperature difference between the pixels corresponding to the left and right eyes in real time. Areas where the difference exceeds a set threshold (ΔT ≥ 4°C) are marked and thermal radiation value compensation is performed using a neighborhood weighted average algorithm. S540: Suppresses edge artifacts in non-overlapping areas of the binocular field of view. Generates an extended compensation band based on a temperature distribution trend prediction algorithm. The extended width is 3%-5% of the field of view width, and the temperature gradient change rate within the compensation band does not exceed ±10% of the original field of view. S550: Performs final optimization processing before video output, including: Enhance the sensitivity of pixels with temperature values ​​in the critical sensing range (30°C-45°C), increasing their contrast by 1.2-1.5 times; The noise reduction filter is activated for the static background area with no changes for 10 consecutive frames. The noise suppression strength is inversely proportional to the rate of change of the average temperature of the scene. When the left-eye and right-eye image fusion scores are lower than a preset threshold (85 points), the compensation parameter iterative optimization from steps S300 to S400 is automatically triggered.

[0010] Compared with the prior art, the present invention has the following beneficial effects: The present invention effectively solves the ghosting problem in binocular thermal imaging observation through real-time dynamic registration and progressive fusion mechanism. It adopts thermal radiation feature extraction and cross-channel synchronous analysis technology to break through the feature recognition bottleneck of traditional visible light registration methods in thermal imaging scenarios. Combined with dynamic offset compensation and temperature-adaptive progressive fusion algorithm, it achieves sub-pixel parallax correction while maintaining the integrity of the original thermodynamic data, eliminating the continuous ghosting interference caused by the spatial deviation of the binocular image during human eye observation, and can adapt to the real-time observation needs of targets at different distances without modifying the optical hardware structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flow chart of steps S100-S500 in the present invention. DETAILED DESCRIPTION

[0012] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0013] See also Figure 1 , a binocular thermal imaging image fusion method based on dynamic compensation, comprising the following steps: S100: Setting a thermal radiation feature extraction unit in each of the left and right image processing channels to identify high temperature gradient mutation areas in the thermal imaging image in real time as reference feature points; S200: Establishing a feature point synchronization comparison mechanism between binocular channels, and calculating the spatial position deviation of corresponding feature points in the left and right images through a coordinate mapping algorithm; S300: Build a dynamic compensation calculation module to generate a pixel displacement matrix of the right eye image based on the real-time position deviation value, and perform real-time geometric correction on the right eye image before outputting the image; S400: Using a progressive optimization algorithm to achieve dynamic adaptation of the binocular field of view by tracking feature point trajectories in consecutive frames, the algorithm includes a weight adjustment mechanism based on temperature distribution changes; S500: A final fusion processing unit is set before the video output interface to perform edge smoothing and contrast equalization processing on the corrected binocular image; The thermal radiation feature extraction unit uses proprietary identification of high-temperature gradient mutation areas (step S100) and combines this with a temperature gradient direction consistency screening mechanism to increase the capture rate of effective feature points. This is significantly improved compared to traditional grayscale feature matching methods, and in particular, 5-8 effective reference points can still be stably extracted in uniform temperature field scenes. Based on the pixel displacement matrix generation technology of the dynamic compensation calculation module (step S300), combined with the frame-by-frame progressive compensation mode, the single-frame processing delay is controlled within 20ms, achieving sub-pixel parallax compensation for real-time video streams and reducing the amount of residual parallax; The progressive optimization algorithm (step S400) enables the system to adaptively maintain parallax compensation accuracy within the target distance range by tracking the trajectory of feature points in consecutive frames and adjusting the weight of temperature changes. This overcomes the compensation lag problem caused by traditional fixed parameter compensation schemes in target moving scenarios. The entire process is implemented based on image processing methods, without the need for additional optical prisms or mechanical adjustment devices (such as the reflector assembly described in the published patent with publication number CN112229434A). It can be directly integrated into the video processing pipeline of existing thermal imaging equipment, reducing modification costs; The final fusion processing unit (step S500) improves the visual comfort of the binocular fusion image through edge smoothing and contrast equalization processing, eliminating the eye fatigue phenomenon caused by parallax mutation in traditional solutions; The present invention effectively solves the ghosting problem in binocular thermal imaging observation through real-time dynamic registration and progressive fusion mechanism. It adopts thermal radiation feature extraction and cross-channel synchronous analysis technology to break through the feature recognition bottleneck of traditional visible light registration methods in thermal imaging scenarios. Combined with dynamic offset compensation and temperature-adaptive progressive fusion algorithm, it achieves sub-pixel parallax correction while maintaining the integrity of the original thermodynamic data, eliminating the continuous ghosting interference caused by the spatial deviation of the binocular image during human eye observation, and can adapt to the real-time observation needs of targets at different distances without modifying the optical hardware structure.

[0014] In the embodiment of the present invention, step S100 includes: S110: Analyze the temperature field of the thermal imaging images in the left and right image processing channels frame by frame, calculate the temperature change rate in units of pixel matrix, and select continuous areas where the temperature change rate exceeds a set threshold; S120: Perform morphological optimization on the selected high-temperature gradient areas, remove discrete noise points with an area smaller than 10×10 pixels, and retain thermal radiation feature areas with stable edge contours; S130: Based on the temperature gradient direction consistency judgment, the direction vector of the pixel cluster in the feature area is calibrated, and the feature point whose gradient direction has an angle with the visual axis with a deviation of less than 15° is selected as the valid reference point; S140: establishing a dynamic association list of feature points of the left and right channels, recording the temperature extreme value coordinates and spatial distribution topological relationship of each feature point; The frame-by-frame temperature field analysis and dynamic threshold screening mechanism (S110) are used to reduce the false detection rate of feature points compared with traditional methods. Morphological optimization (S120) effectively eliminates pseudo-feature areas caused by thermal imaging noise, thereby increasing the density of effective feature points. The visual axis angle constraint condition established by gradient direction vector calibration (S130) is combined to improve the geometric consistency compliance rate of feature point matching. The topological relationship recording function of the dynamic association list (S140) provides structured data input for subsequent matching steps, reducing the feature point search time in step S200. It can still stably extract effective reference points in complex thermal scenes, ensuring the reliability of the underlying data of the ghosting elimination system.

[0015] In the embodiment of the present invention, step S200 includes the following steps: S210: uniformly converting the coordinates of the feature points of the left and right channels into a polar coordinate system with the center of the binocular field of view as the origin, and performing normalization processing to eliminate the initial installation position deviation between the two channels; S220: Based on the temperature extreme value distribution characteristics of the feature points, perform rough matching on the feature points in the same temperature range (±3°C) in the left and right images, setting the initial search radius to 15% of the field of view width; S230: Using a sliding window matching algorithm, the spatial distribution pattern of feature points is compared pixel by pixel within the coarse matching range, and feature point pairs that meet a similarity threshold (≥85%) are calculated; S240: Fitting the geometric transformation matrix of the left and right feature point sets by the least squares method, and outputting the real-time translation deviation Δx, Δy and rotation angle deviation θ between the binocular images; S250: Establishing a dynamic error correction mechanism to automatically trigger recalibration of feature point matching parameters when the fluctuation range of Δx or Δy exceeds ±5 pixels in 5 consecutive frames; Polar coordinate system conversion and normalization processing (S210) eliminates the initial installation deviation of the binocular channel by the order of ±0.5 pixels; coarse matching based on the temperature range (±3°C) (S220) narrows the feature point search range to 15% of the field of view width, reducing the matching calculation amount by more than 65%; the sliding window matching algorithm (S230) combined with an 85% similarity threshold improves the matching accuracy from 72% to 95% at a resolution of 640×480; real-time fitting of the geometric transformation matrix (S240) enables the translation deviation detection accuracy to reach 0.2 pixels and the rotation angle detection error to ≤0.5°; the dynamic error correction mechanism (S250) improves the stability of continuous frame matching by 40% in scenarios with rapid target movement, avoiding the sudden ghosting caused by accumulated errors in traditional solutions, and significantly improving the accuracy and robustness of binocular feature matching.

[0016] In the embodiment of the present invention, step S300 includes the following detailed steps: S310: Receive the real-time translation deviations Δx, Δy and rotation angle deviation θ output in step S200, and generate a pixel displacement matrix including a horizontal displacement compensation amount (Δx'=Δx×k1), a vertical displacement compensation amount (Δy'=Δy×k2), and a rotation compensation amount θ'=θ×k3, where k1 and k2 are displacement attenuation coefficients (0.8≤k1, k2≤1.2), and k3 is a rotation correction coefficient (0.5≤k3≤1.5); S320: dynamically remapping the right eye image pixels based on the pixel displacement matrix, and filling the empty pixel areas generated after the displacement using a bilinear interpolation algorithm to preserve the integrity of the original temperature data; S330: Set a displacement threshold constraint mechanism. When the absolute value of Δx' or Δy' in a single frame exceeds 5% of the field of view width, enable the frame-by-frame progressive compensation mode and spread the total compensation amount over the next 3-5 frames. S340: Embed verification markers in the geometrically corrected right-eye image, and verify the correction effect by spatially overlapping with the corresponding markers in the left-eye image. If the deviation exceeds 2 pixels, re-matching in step S200 is triggered. S350: Establish a compensation amount history database, and dynamically adjust the coefficient values ​​of k1, k2, and k3 according to the compensation amount change trend of 10 consecutive frames. The adjustment range shall not exceed ±20% of the current value. The combined application of a displacement attenuation coefficient (0.8≤k1, k2≤1.2) and a frame-by-frame progressive compensation mode (S330) limits the single-frame compensation amount in large parallax scenarios to within 5% of the field of view width, avoiding sudden image distortion and compressing the compensation delay to 3-5 frames (≤200ms). The bilinear interpolation algorithm (S320) maintains the integrity of temperature data during pixel remapping, reducing the information loss rate in critical high-temperature areas (≥100°C) from 12% in traditional methods to below 2%. The closed-loop verification mechanism (S340) triggers compensation iterations through 2-pixel error detection, reducing the amplitude of residual parallax fluctuations in continuous observations. The dynamic coefficient adjustment (S350) adaptively optimizes compensation parameters based on 10 frames of historical data, reducing the standard deviation of compensation accuracy within the observation distance range of 3-50m from ±1.8 pixels to ±0.5 pixels, maintaining stable ghosting elimination under complex working conditions.

[0017] In the embodiment of the present invention, step S400 includes the following steps: S410: Within a time window of 5-10 consecutive frames, track the successfully matched feature points in the left and right fields of view, calculate the motion vector of each feature point, and select a stable feature point set with a motion direction consistency of ≥75%; S420: Dynamically assign weight coefficients based on the temperature change rate of the region where the feature point is located. The high-temperature dynamic region (temperature change rate ≥ 3°C / s) is assigned a weight W1 of 0.2-0.5, and the low-temperature stable region (temperature change rate ≤ 1°C / s) is assigned a weight W2 of 0.8-1.2. S430: Perform weighted fusion on the spatial distribution of feature points in the left and right fields of view based on the weight coefficient to generate binocular field of view adaptation parameters, including the parallax compensation baseline value L = Σ(Wᵢ×Δxᵢ) / ΣWᵢ and the viewing angle convergence factor α = 1-(Δθ_max / 30°), where Δθ_max is the maximum angular deviation of the feature points; S440: Adopting a sliding time window update mechanism, the adaptation parameters are updated every 3 frames. When the difference between the L values ​​calculated twice exceeds 15% of the baseline threshold, the time window is automatically expanded to 8-12 frames for smooth transition. S450: During the fusion process, the success rate of feature point tracking is monitored in real time. If the proportion of feature points that fail to be tracked for three consecutive frames is ≥40%, the feature point re-extraction in step S100 and the matching parameter reset in step S200 are triggered. The feature point trajectory tracking based on a 5-10 frame time window (S410) improves the efficiency of removing motion interference points and reduces the disparity calculation error of the stable feature point set to 0.15 pixels. The temperature change rate weight distribution mechanism (S420) assigns low weights in high-temperature dynamic areas (≥3℃ / s), effectively suppressing the interference of thermal flicker noise on fusion parameters. The dual-mode parameter design of the disparity compensation baseline value and the viewing angle convergence factor (S430) ensures that the binocular field of view adaptation accuracy remains stable within 0.8 pixels when the target moves rapidly (speed ≥2m / s). The sliding time window update mechanism (S440) reduces the image jitter amplitude during parameter update by dynamically adjusting the time window length (3-12 frames). The adaptive reset mechanism (S450) triggers system recalibration when the feature point tracking failure rate is ≥40%, extending the continuous effective working time in complex scenes by 3-5 times, and realizing intelligent dynamic optimization of thermal imaging binocular fusion parameters.

[0018] In the embodiment of the present invention, step S500 includes the following steps: S510: Performing edge fusion processing based on temperature gradient on the corrected left and right eye images. By calculating the second-order derivative of the temperature difference between adjacent pixels, the boundary line with the absolute value of the temperature gradient ≥ 0.5°C / pixel is extracted, and the edge smoothing transition is performed in the overlapping area of ​​the binocular field of view. S520: Using dynamic histogram mapping technology, the global temperature distribution range (T_min to T_max) of the left and right images is divided into N equal segments (N = 5-8). Based on the pixel ratio of each temperature segment in the left eye image, the right eye image is subjected to segment-by-segment contrast stretching and matching. S530: A difference detection module is embedded in the fused image to calculate the temperature difference between the pixels corresponding to the left and right eyes in real time. Areas where the difference exceeds a set threshold (ΔT ≥ 4°C) are marked and thermal radiation value compensation is performed using a neighborhood weighted average algorithm. S540: Suppresses edge artifacts in non-overlapping areas of the binocular field of view. Generates an extended compensation band based on a temperature distribution trend prediction algorithm. The extended width is 3%-5% of the field of view width, and the temperature gradient change rate within the compensation band does not exceed ±10% of the original field of view. S550: Performs final optimization processing before video output, including: Enhance the sensitivity of pixels with temperature values ​​in the critical sensing range (30°C-45°C), increasing their contrast by 1.2-1.5 times; The noise reduction filter is activated for the static background area with no changes for 10 consecutive frames. The noise suppression strength is inversely proportional to the rate of change of the average temperature of the scene. When the left-eye and right-eye image fusion scores are lower than a preset threshold (85 points), the compensation parameter iterative optimization of steps S300 to S400 is automatically triggered; Temperature gradient edge fusion technology (S510) extracts thermal boundaries with an accuracy of 0.5°C / pixel, reducing edge artifacts in the overlapping area of ​​the binocular field of view by 78%; dynamic histogram segment mapping (S520) compresses the contrast difference between the left and right eye images by balancing the temperature distribution in 5-8 segments; the difference detection module (S530) implements neighborhood compensation in areas with ΔT ≥ 4°C, reducing the area of ​​local ghosting residuals; the field of view extension compensation band (S540) retains more than 90% of the effective thermal information in the non-overlapping area through trend prediction filling with a width of 3%-5%; the final optimization processing (S550) enhances the sensitivity of the critical temperature zone, improving the detail recognition in the key range of human body temperature measurement (30-45°C) by 40%. Combined with noise reduction filtering and fusion feedback mechanism, the visual comfort score of the output image reaches ISO standard A level (≥90 points), achieving zero-misjudgment observation in scenarios such as medical fever screening.

[0019] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A binocular thermal imaging image fusion method based on dynamic compensation, characterized in that: The following steps are involved: S100: Setting a thermal radiation feature extraction unit in each of the left and right image processing channels to identify high temperature gradient mutation areas in the thermal imaging image in real time as reference feature points; S200: Establishing a feature point synchronization comparison mechanism between binocular channels, and calculating the spatial position deviation of corresponding feature points in the left and right images through a coordinate mapping algorithm; S300: Build a dynamic compensation calculation module to generate a pixel displacement matrix of the right eye image based on the real-time position deviation value, and perform real-time geometric correction on the right eye image before outputting the image; S400: Using a progressive optimization algorithm to achieve dynamic adaptation of the binocular field of view by tracking feature point trajectories in consecutive frames, the algorithm includes a weight adjustment mechanism based on temperature distribution changes; S500: A final fusion processing unit is set before the video output interface to perform edge smoothing and contrast equalization processing on the corrected binocular image.

2. The binocular thermal imaging image fusion method based on dynamic compensation according to claim 1, characterized in that: Step S100 includes: S110: Analyze the temperature field of the thermal imaging images in the left and right image processing channels frame by frame, calculate the temperature change rate in units of pixel matrix, and select continuous areas where the temperature change rate exceeds a set threshold; S120: Perform morphological optimization on the selected high-temperature gradient areas, remove discrete noise points with an area smaller than 10×10 pixels, and retain thermal radiation feature areas with stable edge contours; S130: Based on the temperature gradient direction consistency judgment, the direction vector of the pixel cluster in the feature area is calibrated, and the feature point whose gradient direction has an angle with the visual axis with a deviation of less than 15° is selected as the valid reference point; S140: Establish a dynamic association list of feature points of the left and right channels, and record the temperature extreme value coordinates and spatial distribution topological relationship of each feature point.

3. The binocular thermal imaging image fusion method based on dynamic compensation according to claim 2, characterized in that: The step S200 includes the following steps: S210: uniformly converting the coordinates of the feature points of the left and right channels into a polar coordinate system with the center of the binocular field of view as the origin, and performing normalization processing to eliminate the initial installation position deviation between the two channels; S220: Based on the temperature extreme value distribution characteristics of the feature points, perform rough matching on the feature points in the same temperature range (±3°C) in the left and right images, setting the initial search radius to 15% of the field of view width; S230: Using a sliding window matching algorithm, the spatial distribution pattern of feature points is compared pixel by pixel within the coarse matching range, and feature point pairs that meet a similarity threshold (≥85%) are calculated; S240: Fitting the geometric transformation matrix of the left and right feature point sets by the least squares method, and outputting the real-time translation deviation Δx, Δy and rotation angle deviation θ between the binocular images; S250: Establish a dynamic error correction mechanism to automatically trigger recalibration of feature point matching parameters when the fluctuation amplitude of Δx or Δy exceeds ±5 pixels in 5 consecutive frames.

4. The binocular thermal imaging image fusion method based on dynamic compensation according to claim 3, characterized in that: The step S300 includes the following detailed steps: S310: Receive the real-time translation deviations Δx, Δy and rotation angle deviation θ output in step S200, and generate a pixel displacement matrix including a horizontal displacement compensation amount (Δx'=Δx×k1), a vertical displacement compensation amount (Δy'=Δy×k2), and a rotation compensation amount θ'=θ×k3, where k1 and k2 are displacement attenuation coefficients (0.8≤k1, k2≤1.2), and k3 is a rotation correction coefficient (0.5≤k3≤1.5); S320: dynamically remapping the right eye image pixels based on the pixel displacement matrix, and filling the empty pixel areas generated after the displacement using a bilinear interpolation algorithm to preserve the integrity of the original temperature data; S330: Set a displacement threshold constraint mechanism. When the absolute value of Δx' or Δy' in a single frame exceeds 5% of the field of view width, enable the frame-by-frame progressive compensation mode and spread the total compensation amount over the next 3-5 frames. S340: Embed verification markers in the geometrically corrected right-eye image, and verify the correction effect by spatially overlapping with the corresponding markers in the left-eye image. If the deviation exceeds 2 pixels, re-matching in step S200 is triggered. S350: Establish a compensation amount history database, and dynamically adjust the coefficient values ​​of k1, k2, and k3 according to the compensation amount change trend of 10 consecutive frames, with the adjustment range not exceeding ±20% of the current value.

5. The binocular thermal imaging image fusion method based on dynamic compensation according to claim 4, characterized in that: The step S400 includes the following steps: S410: Within a time window of 5-10 consecutive frames, track the successfully matched feature points in the left and right fields of view, calculate the motion vector of each feature point, and select a stable feature point set with a motion direction consistency of ≥75%; S420: Dynamically assign weight coefficients based on the temperature change rate of the region where the feature point is located. A high-temperature dynamic region (temperature change rate ≥ 3°C / s) is assigned a weight of W1 = 0.2-0.5, and a low-temperature stable region (temperature change rate ≤ 1°C / s) is assigned a weight of W2 = 0.8-1.

2. S430: Perform weighted fusion on the spatial distribution of feature points in the left and right fields of view based on the weight coefficient to generate binocular field of view adaptation parameters, including the parallax compensation baseline value L = Σ(Wᵢ×Δxᵢ) / ΣWᵢ and the viewing angle convergence factor α = 1-(Δθ_max / 30°), where Δθ_max is the maximum angular deviation of the feature points; S440: Adopting a sliding time window update mechanism, the adaptation parameters are updated every 3 frames. When the difference between the L values ​​calculated twice exceeds 15% of the baseline threshold, the time window is automatically expanded to 8-12 frames for smooth transition. S450: During the fusion process, the success rate of feature point tracking is monitored in real time. If the proportion of feature points that fail to be tracked for three consecutive frames is ≥40%, the feature point re-extraction in step S100 and the matching parameter reset in step S200 are triggered.

6. The binocular thermal imaging image fusion method based on dynamic compensation according to claim 5, characterized in that: The step S500 includes the following steps: S510: Performing edge fusion processing based on temperature gradient on the corrected left and right eye images. By calculating the second-order derivative of the temperature difference between adjacent pixels, the boundary line with the absolute value of the temperature gradient ≥ 0.5°C / pixel is extracted, and the edge smoothing transition is performed in the overlapping area of ​​the binocular field of view. S520: Using dynamic histogram mapping technology, the global temperature distribution range (T_min to T_max) of the left and right images is divided into N equal segments (N = 5-8). Based on the pixel ratio of each temperature segment in the left eye image, the right eye image is subjected to segment-by-segment contrast stretching and matching. S530: A difference detection module is embedded in the fused image to calculate the temperature difference between the pixels corresponding to the left and right eyes in real time. Areas where the difference exceeds a set threshold (ΔT ≥ 4°C) are marked and thermal radiation value compensation is performed using a neighborhood weighted average algorithm. S540: Suppresses edge artifacts in non-overlapping areas of the binocular field of view. Generates an extended compensation band based on a temperature distribution trend prediction algorithm. The extended width is 3%-5% of the field of view width, and the temperature gradient change rate within the compensation band does not exceed ±10% of the original field of view. S550: Performs final optimization processing before video output, including: Enhance the sensitivity of pixels with temperature values ​​in the critical sensing range (30°C-45°C), increasing their contrast by 1.2-1.5 times; The noise reduction filter is activated for the static background area with no changes for 10 consecutive frames. The noise suppression strength is inversely proportional to the rate of change of the average temperature of the scene. When the left-eye and right-eye image fusion scores are lower than a preset threshold (85 points), the compensation parameter iterative optimization from steps S300 to S400 is automatically triggered.

Images