Glare detection system based on binocular vision and brightness fusion
The glare detection system, which integrates binocular vision and brightness, achieves synchronous acquisition and deep fusion of multi-dimensional parameters, solving the problems of insufficient temporal efficiency, parameter correlation and adaptability in traditional glare detection methods, and improving the accuracy and real-time performance of detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional glare detection methods have shortcomings in data acquisition timing and efficiency, parameter acquisition dimensions and correlation, adaptability and automation. They are difficult to achieve synchronous acquisition and deep fusion of multi-dimensional parameters, resulting in inaccurate detection results and poor real-time performance.
A glare detection system based on binocular vision and brightness fusion is adopted. Exposure is synchronously triggered by binocular camera components, and the timing synchronization calibration of the brightness channel and vision channel is performed. Exposure parameters are dynamically adjusted by combining spatial coordinate mapping and brightness feedback. Multi-parameter correlation calibration and deep fusion are performed to establish a pixel-level coupled calculation model. The detection results are optimized by parallax consistency verification and historical data correction.
It achieves synchronous integrated acquisition and high-precision correlation calibration of multi-dimensional optical parameters, improving the timeliness, accuracy and robustness of detection, and can accurately extract glare features and perform real-time adaptive optimization under complex lighting conditions.
Smart Images
Figure CN121954428A_ABST
Abstract
Description
A Glare Detection System Based on Binocular Vision and Brightness Fusion Technical Field
[0001] This invention belongs to the field of optical measurement and detection technology, specifically relating to a glare detection system based on binocular vision and brightness fusion. Background Technology
[0002] Glare detection has significant applications in various fields, including lighting engineering, visual comfort evaluation, driving safety monitoring, and quality control of display devices. Its core function is to objectively and quantitatively assess the discomfort or decreased visual ability caused by excessive light from light sources or reflective surfaces. Accurate and efficient glare detection technology is crucial for optimizing lighting design, improving the quality of the visual environment, and ensuring operational safety.
[0003] Traditional glare detection methods primarily rely on measurement modes based on single-lens imaging devices. A typical approach involves using a DSLR camera with high dynamic range imaging capabilities, manually or programmatically adjusting different exposure parameters to capture multiple sequential shots of the same target scene, obtaining multiple images with varying brightness information. Subsequently, image processing algorithms are used to synthesize these images with different exposures to broaden the system's effective brightness measurement range, thereby calculating the brightness distribution in different areas of the scene. Finally, based on a specific glare evaluation model (such as glare value, threshold increment, glare index), combined with measured brightness data, observer position and viewing direction, and light source geometry, the glare index is calculated.
[0004] However, this traditional approach based on single-lens, multi-frame image synthesis has gradually revealed several inherent limitations in practice. First, it suffers from deficiencies in the timing and efficiency of data acquisition. Because it requires continuously capturing multiple photos with different exposures of the same static scene, the entire acquisition process is time-consuming and cannot achieve instantaneous synchronous acquisition. This not only limits the system's application in dynamically changing scenes, such as glare detection of moving vehicle headlights in traffic conditions, but also makes the measurement results susceptible to instantaneous fluctuations in light within the scene or slight movements of the target object, reducing the spatiotemporal consistency of the data. Second, it has shortcomings in the dimensionality and correlation of parameter acquisition. Traditional methods mainly rely on two-dimensional image information to infer brightness, and are weak in acquiring the precise three-dimensional spatial position (distance) and true physical size of the target object. They typically require additional ranging equipment or pre-set reference objects of known size for estimation, a cumbersome process that easily introduces errors. Since key parameters such as brightness, distance, and size originate from different times or different devices, spatiotemporal alignment and physical correlation calibration between them are difficult, resulting in a weak foundation for subsequent fusion calculations.
[0005] Secondly, due to the lack of a synchronous native acquisition and deep fusion mechanism for depth information and multi-dimensional parameters, traditional methods are poorly adaptable to complex and ever-changing lighting environments. For example, the apparent brightness of light sources at different distances can change due to atmospheric or medium attenuation, which traditional methods struggle to accurately compensate for. Furthermore, when calculating glare indices involving light source size, the inability to readily obtain precise spatial dimensions often leads to approximations, affecting the accuracy of the evaluation results. In addition, the entire processing flow involves numerous steps, relies on manual intervention or complex post-calibration, and the system's automation and real-time performance need improvement.
[0006] Therefore, the industry needs a new glare detection technology solution that can overcome the above-mentioned defects. An ideal technology should be able to achieve high-speed synchronous acquisition of key optical and geometric parameters, ensuring the spatiotemporal uniformity of the data source; it should possess high-precision automatic calibration and deep fusion capabilities for multi-source heterogeneous parameters to construct a more reliable feature representation; and ultimately improve the system's detection accuracy, robustness, and overall efficiency under different lighting conditions. Summary of the Invention
[0007] To address the aforementioned problems in the existing technology, this invention provides a glare detection system based on binocular vision and brightness fusion. The objective of this invention can be achieved through the following technical solution: A glare detection system based on binocular vision and brightness fusion includes: a binocular synchronous acquisition module, a multi-parameter correlation calibration module, a binocular-brightness fusion calculation module, and a glare result output and verification module. The binocular synchronous acquisition module is configured with a binocular camera component, which captures the original parameters of distance, size, and brightness of the target scene through a synchronous exposure mechanism triggered by the left and right lenses. It performs a time-series synchronous calibration operation between the brightness channel and the binocular vision channel, dynamically adjusts the exposure parameters based on real-time brightness feedback of the scene, and generates a target scene original parameter set containing multi-dimensional original information. The multi-parameter correlation calibration module obtains the target scene original parameter set and, based on spatial coordinate mapping, adjusts the distance and size parameters... The system performs mutual calibration of the inch parameter, performs attenuation compensation calibration of the brightness parameter in combination with the distance factor, removes invalid correlated data through a time-series alignment algorithm, and generates a standardized calibration parameter set. The binocular-brightness fusion calculation module obtains the standardized calibration parameter set, assigns dynamic weights to the distance, size, and brightness parameters using a weighted confidence fusion mechanism, establishes a spatial pixel-level coupled calculation model for deep parameter binding, and adapts to different lighting scenarios through a dynamic threshold adaptive strategy to generate comprehensive parameters that can fully reflect glare characteristics. The glare result output and verification module obtains the comprehensive parameters, calculates the glare index and spatial distribution information of the glare area of the target scene, verifies the validity of the data through a binocular parallax consistency verification mechanism, dynamically corrects the detection results by combining historical detection data of similar scenes, and dynamically adjusts the system operating parameters based on the correction results.
[0008] Specifically, the synchronous exposure triggering mechanism for the left and right lenses includes: setting up a co-source trigger control unit to generate a synchronous trigger signal and transmitting it in parallel to the exposure drive module of the binocular lens; constructing an exposure timing capture link to synchronously acquire the action timing of the start and end of exposure for both lenses; comparing the action timing differences between the two lenses and dynamically adjusting the transmission delay of the trigger signal.
[0009] Specifically, the timing synchronization calibration operation of the brightness channel and the binocular vision channel is performed by: constructing a unified timing reference unit, generating a stable calibration time axis, and using dual-channel delay calibration technology to capture the acquisition start delay and data transmission delay of the brightness channel and the binocular vision channel; based on the difference between the two types of delays, injecting a timing compensation signal into the data receiving end to correct the timestamp deviation of the data of the brightness channel and the binocular vision channel.
[0010] Specifically, the method of dynamically adjusting exposure parameters in conjunction with real-time scene brightness feedback is as follows: configure a high-sensitivity brightness sensing component to capture scene brightness change signals in real time, divide the scene brightness change range into multiple adjustment ranges based on the scene brightness change range, and formulate brightness graded response rules; according to the range to which the brightness sensing signal belongs, call the brightness graded response rules to adjust the lens exposure time and gain, and balance the brightness acquisition accuracy and parallax matching requirements.
[0011] Specifically, the distance-size parameter mutual calibration operation based on spatial coordinate mapping is performed as follows: The three-dimensional spatial coordinate set of the target scene is calculated using the principle of binocular parallax, and the three-dimensional coordinate information of key points of the target contour is extracted; a coordinate-size association mapping table is constructed, and the spatial straight-line distance is calculated using the three-dimensional coordinates of adjacent vertices, inflection points, and edge feature points to inversely deduce the actual size reference value of the target; the reference value is compared with the original size parameter to generate a calibration coefficient corresponding to each point, and the original size parameter is corrected point by point based on the target pixel distribution to strengthen the spatial correlation between the size parameter and the distance parameter.
[0012] Specifically, the distance attenuation compensation calibration of the brightness parameter is performed as follows: extract distance information from the original parameter set of the target scene, divide different distance intervals according to the distance gradient, set differentiated compensation coefficients based on the brightness attenuation characteristics corresponding to the different distance intervals, and calculate the original brightness parameter with the compensation coefficient of the corresponding interval to correct the brightness deviation caused by the distance factor.
[0013] Specifically, the method for removing invalid associated data using the time-series alignment algorithm is as follows: obtain the timestamp information of each parameter at the time of collection, establish a timestamp association matrix, and preset a time-series association threshold to define the time difference range of valid associations; compare the timestamp differences of each parameter and mark the data combinations that exceed the threshold range; remove the data combinations from the original parameter set and retain the parameters that match the time sequence.
[0014] Specifically, the weighted confidence fusion mechanism is used to assign dynamic weights to the three types of parameters. The specific method is as follows: a confidence assessment system is generated by evaluating the disparity matching confidence of the distance parameter, the spatial fitting confidence of the size parameter, and the acquisition stability of the brightness parameter; based on the three types of confidence assessment results, the weights of each parameter are calculated according to preset association rules; and a real-time weight update link is constructed to dynamically refresh the weight allocation results.
[0015] Specifically, the method for establishing a spatial pixel-level coupled calculation model for deep parameter binding is as follows: mapping the three-dimensional spatial coordinates of the target scene obtained by binocular parallax calculation to image pixels one-to-one, generating a pixel-spatial coordinate bidirectional association table, obtaining the spatial location of image pixels, extracting the brightness parameters and size information corresponding to the image pixels, and establishing a pixel-level multi-dimensional feature set containing spatial coordinates, brightness values, and actual size; deeply binding multiple types of parameters of the same pixel through spatial location correlation to form pixel-level feature units, and dividing them into a global glare feature matrix based on spatial regions.
[0016] Specifically, the dynamic threshold adaptive strategy adapts to different lighting scenarios in the following ways: a light intensity detection component is set up to capture scene light fluctuation signals in real time, and instantaneous light interference is eliminated through signal filtering; a multi-scene threshold library is preset to store the fusion threshold range corresponding to different light intensity scenes; based on the light detection results, the threshold range in the multi-scene threshold library is called to adjust the threshold parameters of the fusion operation to adapt to the dynamic changes in scene light.
[0017] Specifically, the binocular disparity consistency verification mechanism verifies the validity of the data by: extracting image data acquired by the binocular lens, calculating the disparity data of all pixels, and generating a disparity map; establishing disparity consistency evaluation rules and setting the disparity fluctuation range in combination with the spatial structure characteristics of the target scene; comparing the disparity data with the disparity fluctuation range, marking abnormal pixels that exceed the range, and initiating a data re-acquisition process when the proportion of abnormal pixels exceeds a preset standard.
[0018] Specifically, the method for dynamically correcting the detection results by combining historical detection data of similar scenarios is as follows: classify and store detection data and correction records of different scenarios to establish a historical data storage system; use data association retrieval technology to retrieve historical data of similar scenarios based on the characteristics of the current scenario, extract correction patterns from the historical data, generate the correction direction and magnitude of the current detection results, and optimize and adjust the detection results.
[0019] The beneficial effects of this invention are as follows: First, it achieves synchronous and integrated acquisition and high-precision correlation calibration of multi-dimensional optical parameters, improving the timeliness and consistency of the data source. Traditional methods rely on multiple shots taken at different times with a single lens, resulting in asynchronous parameter acquisition and a cumbersome process. This invention uses hardware-synchronized exposure triggering of the left and right lenses of a binocular camera and performs temporal synchronous calibration of the brightness and visual channels, instantly capturing the original information of distance, size, and brightness of the target scene. Furthermore, it performs mutual calibration of distance and size through spatial coordinate mapping and performs distance-based attenuation compensation for brightness, fundamentally ensuring the spatiotemporal alignment and physical correlation of multi-source data, laying a reliable foundation for subsequent accurate fusion calculations.
[0020] Secondly, by employing a confidence-based dynamic fusion mechanism and pixel-level deep binding, a more comprehensive and accurate glare feature representation model is constructed, significantly improving detection accuracy. Traditional methods have single parameter dimensions and lack effective fusion mechanisms. This invention innovatively adopts a weighted confidence fusion mechanism, dynamically assigning weights to three types of parameters: distance, size, and brightness, and establishing a spatial pixel-level coupled calculation model. This deeply binds multi-dimensional parameters at the same spatial location, generating a pixel-level multi-dimensional feature set. This deep fusion approach can more comprehensively characterize the optical and spatial properties of glare. Combined with a dynamic threshold adaptive strategy, the system can automatically adapt to different lighting scenarios, thus achieving accurate and stable glare feature extraction and quantification under various complex environmental conditions.
[0021] Third, a closed-loop optimization system incorporating real-time verification and historical feedback was established, enhancing the robustness of the detection process and the system's adaptability. Traditional methods are mostly open-loop measurements, lacking effective verification and optimization mechanisms. This invention, through a binocular parallax consistency verification mechanism, can verify data validity before output and trigger resampling in case of anomalies, ensuring the quality of the original data. More importantly, the system can dynamically correct the current results by combining historical detection data from similar scenarios, and dynamically adjust the operating parameters of each module (such as acquisition strategies, calibration rules, and fusion thresholds) based on the correction feedback. This closed-loop workflow of "detection-verification-correction-optimization" enables the system to continuously learn and self-optimize, effectively improving long-term operational reliability and adaptability to different application scenarios. Attached Figure Description
[0022] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.
[0023] Figure 1 is a flowchart of a glare detection system based on binocular vision and brightness fusion according to the present invention; Figure 2 is a timing diagram of a glare detection system based on binocular vision and brightness fusion according to the present invention. Detailed Implementation
[0024] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.
[0025] Please refer to Figure 1-2. A glare detection system based on binocular vision and brightness fusion includes: a binocular synchronous acquisition module, a multi-parameter correlation calibration module, a binocular-brightness fusion calculation module, and a glare result output and verification module. The binocular synchronous acquisition module is equipped with a binocular camera component. It captures the original parameters of distance, size, and brightness of the target scene by synchronously triggering an exposure mechanism through the left and right lenses. It performs a time-series synchronous calibration operation between the brightness channel and the binocular vision channel, and dynamically adjusts the exposure parameters in combination with real-time brightness feedback of the scene to generate a target scene original parameter set containing multi-dimensional original information. The multi-parameter correlation calibration module obtains the target scene original parameter set, performs mutual calibration operation on the distance and size parameters based on the spatial coordinate mapping relationship, and performs attenuation compensation on the brightness parameters in combination with the distance factor. The calibration process involves eliminating invalid correlated data using a time-series alignment algorithm to generate a standardized calibration parameter set. The binocular-brightness fusion calculation module acquires this standardized calibration parameter set and uses a weighted confidence fusion mechanism to assign dynamic weights to three types of parameters: distance, size, and brightness. A spatial pixel-level coupled calculation model is established for deep parameter binding, and a dynamic threshold adaptive strategy is used to adapt to different lighting scenarios, generating comprehensive parameters that fully reflect glare characteristics. The glare result output and verification module acquires these comprehensive parameters, calculates the glare index and spatial distribution information of the glare area in the target scene, verifies the data validity using a binocular parallax consistency verification mechanism, dynamically corrects the detection results by combining historical detection data from similar scenes, and dynamically adjusts the system operating parameters based on the correction results.
[0026] Specifically, the synchronous exposure triggering mechanism for the left and right lenses includes: setting up a co-source trigger control unit to generate a synchronous trigger signal and transmitting it in parallel to the exposure drive module of the binocular lens; constructing an exposure timing capture link to synchronously acquire the action timing of the start and end of exposure for both lenses; comparing the action timing differences between the two lenses and dynamically adjusting the transmission delay of the trigger signal.
[0027] In this embodiment, the binocular camera component is equipped with an FPGA chip as a co-source trigger control unit, generating a 10MHz synchronous square wave trigger signal. This frequency selection is based on industry-standard parameters for binocular vision acquisition—a 10MHz square wave satisfies the real-time requirements of parameter acquisition (single-cycle response time of 100ns) while avoiding transmission interference caused by high-frequency signals, conforming to the frequency specifications for synchronous trigger signals in GB / T38265-2019 "Technical Requirements for Machine Vision Camera Interfaces". The signals are transmitted in parallel to the exposure drive modules of the left and right lenses via an LVDS differential transmission link. The differential transmission design effectively resists electromagnetic interference, ensuring the consistency of trigger signal transmission. An exposure timing capture link composed of a high-precision time-to-digital converter (TDC) is constructed, simultaneously acquiring the rising edge of the exposure start signal and the falling edge of the exposure end signal of both lenses at a sampling resolution of 500ps, and storing the acquired timing data in an on-chip buffer.
[0028] The timing differences between the two lenses are compared by the internal logic units of the FPGA, and the start-up delay difference and end-of-life delay difference are calculated. The timing synchronization threshold is set to 1ns. When the start-up delay of the left lens compared to the right lens is detected to be 0.8ns, a transmission delay compensation of 0.8ns is injected into the exposure drive module of the right lens through the trigger signal delay adjustment module. If the end-of-life delay difference exceeds the threshold, the transmission delay of the trigger signal is dynamically fine-tuned according to the difference.
[0029] The entire process is executed in real time through hardware logic without software intervention, ensuring that the start and end times of the exposure actions of the two lenses are completely synchronized. This provides spatiotemporal consistency for the subsequent synchronous acquisition of distance, size, and brightness parameters, avoiding multi-parameter correlation errors caused by trigger timing deviations.
[0030] Specifically, the timing synchronization calibration operation of the brightness channel and the binocular vision channel is performed by: constructing a unified timing reference unit, generating a stable calibration time axis, and using dual-channel delay calibration technology to capture the acquisition start delay and data transmission delay of the brightness channel and the binocular vision channel; based on the difference between the two types of delays, injecting a timing compensation signal into the data receiving end to correct the timestamp deviation of the data of the brightness channel and the binocular vision channel.
[0031] Specifically, the method of dynamically adjusting exposure parameters in conjunction with real-time scene brightness feedback is as follows: configure a high-sensitivity brightness sensing component to capture scene brightness change signals in real time, divide the scene brightness change range into multiple adjustment ranges based on the scene brightness change range, and formulate brightness graded response rules; according to the range to which the brightness sensing signal belongs, call the brightness graded response rules to adjust the lens exposure time and gain, and balance the brightness acquisition accuracy and parallax matching requirements.
[0032] Specifically, the distance-size parameter mutual calibration operation based on spatial coordinate mapping is performed as follows: The three-dimensional spatial coordinate set of the target scene is calculated using the principle of binocular parallax, and the three-dimensional coordinate information of key points of the target contour is extracted; a coordinate-size association mapping table is constructed, and the spatial straight-line distance is calculated using the three-dimensional coordinates of adjacent vertices, inflection points, and edge feature points to inversely deduce the actual size reference value of the target; the reference value is compared with the original size parameter to generate a calibration coefficient corresponding to each point, and the original size parameter is corrected point by point based on the target pixel distribution to strengthen the spatial correlation between the size parameter and the distance parameter.
[0033] In this embodiment, the glare detection scenario is applied to road lighting. The target is rectangular traffic signs on both sides of the road (pre-defined as regular geometric outlines). The binocular camera assembly is installed on the detection pole beside the road. The camera baseline distance is B. The intrinsic parameters (focal length f, pixel equivalent k) have been pre-calibrated using the checkerboard calibration method.
[0034] First, the binocular camera acquires the left and right views of the signboard through a synchronously triggered exposure mechanism. Based on the principle of binocular parallax, a stereo matching algorithm is performed on the left and right views to obtain the parallax map D(x,y). The three-dimensional coordinate set (X(x,y),Y(x,y),Z(x,y)) of the target scene is calculated using the three-dimensional coordinate calculation formulas X(x,y)=(x-cx)×Z(x,y) / f, Y(x,y)=(y-cy)×Z(x,y) / f, and Z(x,y)=B×f / [k×D(x,y)]. Here, cx and cy are the coordinates of the camera principal point, x and y are the pixel coordinates, and Z(x,y) is the distance from each pixel of the signboard to the camera.
[0035] Subsequently, an edge detection algorithm was used to extract key points of the sign outline, including four vertices P1, P2, P3, and P4, and the midpoint feature points P5, P6, P7, and P8 of the four edges. The three-dimensional coordinates (Pi, (Xi, Yi, Zi)) (i = 1~8) of each key point were recorded, and a coordinate-size association mapping table T = [Pi, (Xi, Yi, Zi)] was constructed. The distances L between adjacent vertices P1-P2 and P2-P3 were calculated using the spatial straight-line distance formula. 12 L 23 The actual width reference value W_ref=L of the sign was calculated by reverse calculation. 12 Actual height reference value H_ref=L 23 .
[0036] Finally, the original size parameters W_ori and H_ori of the sign are retrieved, and the differences between W_ref and W_ori, and H_ref and H_ori are compared to generate point-by-point corresponding width calibration coefficients K(x,y)=W_ref / W_ori and height calibration coefficients K′(x,y)=H_ref / H_ori. Based on the pixel distribution of the sign, point-by-point corrections are performed on the original size parameters W_ori(x,y) and H_ori(x,y) to obtain the calibrated size parameters W_cal(x,y)=W_ori(x,y)×K(x,y) and H_cal(x,y)=H_ori(x,y)×K′(x,y), strengthening the spatial correlation between the size parameters and the distance parameter Z(x,y) and ensuring the matching accuracy of target size and distance in subsequent glare detection.
[0037] Specifically, the distance attenuation compensation calibration of the brightness parameter is performed as follows: extract distance information from the original parameter set of the target scene, divide different distance intervals according to the distance gradient, set differentiated compensation coefficients based on the brightness attenuation characteristics corresponding to the different distance intervals, and calculate the original brightness parameter with the compensation coefficient of the corresponding interval to correct the brightness deviation caused by the distance factor.
[0038] In this embodiment, taking the urban road nighttime glare detection scenario as an example, the binocular camera assembly is installed on the roadside detection pole, targeting the headlight glare of oncoming vehicles and the street light sources on both sides of the road, with the detection range covering a road area of 0-150m.
[0039] First, the distance information Z(x,y) of each light source pixel is extracted from the original parameter set of the target scene. This distance information is obtained by binocular parallax calculation and is consistent with the three-dimensional spatial coordinates Z(x,y) in weight 5. Three intervals are divided according to the distance gradient: 0-50m (near distance interval), 50-100m (medium distance interval), and 100-150m (far distance interval). This interval division adapts to the actual distance distribution of road detection.
[0040] Based on atmospheric attenuation characteristics and measured road illumination data, differentiated compensation coefficients were set: k1=1.05 for the near distance range, k2=1.15 for the medium distance range, and k3=1.3 for the far distance range. The greater the distance, the larger the compensation coefficient, which conforms to the physical law that brightness decreases with distance.
[0041] Finally, the original brightness parameters L_ori(x,y) of each light source pixel are retrieved, and the corresponding compensation coefficient k (k1, k2, or k3) is matched according to its distance range. The calibrated brightness parameters are obtained by calculating L_cal(x,y) = L_ori(x,y) × k. After correction, the brightness data of the street light at 150m is comparable to that of the same street light at 50m, avoiding misjudgment of the glare index due to distance attenuation and ensuring the consistency of brightness parameters at different spatial locations.
[0042] Specifically, the method for removing invalid associated data using the time-series alignment algorithm is as follows: obtain the timestamp information of each parameter at the time of collection, establish a timestamp association matrix, and preset a time-series association threshold to define the time difference range of valid associations; compare the timestamp differences of each parameter and mark the data combinations that exceed the threshold range; remove the data combinations from the original parameter set and retain the parameters that match the time sequence.
[0043] Specifically, the weighted confidence fusion mechanism is used to assign dynamic weights to the three types of parameters. The specific method is as follows: a confidence assessment system is generated by evaluating the disparity matching confidence of the distance parameter, the spatial fitting confidence of the size parameter, and the acquisition stability of the brightness parameter; based on the three types of confidence assessment results, the weights of each parameter are calculated according to preset association rules; and a real-time weight update link is constructed to dynamically refresh the weight allocation results.
[0044] Specifically, the method for establishing a spatial pixel-level coupled calculation model for deep parameter binding is as follows: The target scene's three-dimensional spatial coordinates obtained from binocular parallax calculation are mapped one-to-one with image pixels to generate a pixel-spatial coordinate bidirectional association table. The spatial location of the image pixels is obtained, and the brightness parameters and size information corresponding to the image pixels are extracted to establish a pixel-level multi-dimensional feature set containing spatial coordinates, brightness values, and actual dimensions. Multiple parameters of the same pixel are deeply bound through spatial location correlation to form pixel-level feature units, which are then divided into a global glare feature matrix based on spatial regions. The model employs a four-layer structured design to ensure the accuracy and correlation of parameter binding: Coordinate mapping layer: Its core function is to establish a bidirectional association between the target scene's three-dimensional spatial coordinates and image pixels. By generating a pixel-spatial coordinate bidirectional association table through the one-to-one correspondence between the pixel's row and column indices and the three-dimensional coordinates, the spatial location of each pixel is accurately anchored, providing a spatial reference for subsequent multi-parameter binding.
[0045] Feature extraction layer: Based on the spatial localization results of the coordinate mapping layer, it extracts multi-dimensional raw parameters corresponding to each pixel in batches. It focuses on capturing brightness parameters (light intensity signals after scene illumination reflection) and size information (local physical size of the target entity corresponding to the pixel), and integrates spatial coordinate data to form a standardized pixel-level multi-dimensional feature set.
[0046] Parameter binding layer: Using spatial location correlation as the core link, it deeply couples the 3D spatial coordinates, brightness parameters, and size information of the same pixel. Through spatial location consistency verification, it ensures that the three types of parameters belong to a unique pixel, avoids cross-pixel parameter confusion, and forms a pixel-level feature unit containing complete spatial-attribute information.
[0047] Matrix Integration Layer: Based on the spatial region division rules of the target scene (such as partitioning by image rows / columns or by target entity contours), scattered pixel-level feature units are categorized and integrated. A global glare feature matrix is formed through regional arrangement, so that the matrix row / column indices directly correspond to the spatial location of the scene, facilitating the rapid calculation of glare index and regional distribution in the subsequent process.
[0048] The three-dimensional spatial coordinates, generated by binocular parallax calculation, include positional information in three dimensions: X (horizontal direction), Y (vertical direction), and Z (distance direction), serving as the core basis for pixel spatial positioning. The image pixel refers to the smallest unit of an image captured by a binocular camera, uniquely identified by row and column indices, and is the basic carrier for parameter binding. The brightness parameter corresponds to the light radiation intensity signal received by the pixel, reflecting the light intensity at the corresponding location in the target scene, and is one of the core attributes of glare characteristics. Size information refers to the local physical size of the target entity corresponding to the pixel, obtained by inverse calculation based on the spatial coordinate spacing, and, together with the distance parameter, reflects the spatial morphology of the target entity. The pixel-level multi-dimensional feature set is a complete information set for a single pixel, containing three core data types: three-dimensional spatial coordinates, brightness parameters, and size information, and is the smallest data unit for parameter binding. The pixel-level feature unit is a standardized data unit coupled through the parameter binding layer, ensuring the spatial consistency of the three types of parameters, and is the basic module for constructing the global glare feature matrix. The global glare feature matrix is the result of scene-level data integration, a set of feature units arranged by spatial region, achieving spatialized and structured storage of multiple parameters.
[0049] The overall operation process is as follows: First, the coordinate mapping layer completes the bidirectional association between 3D spatial coordinates and image pixels, generating an association table and realizing the spatial positioning of pixels. Second, the feature extraction layer extracts the brightness parameters and size information of each pixel based on the positioning results, and combines them with spatial coordinates to form a multi-dimensional feature set. Next, the parameter binding layer uses spatial location as a link to deeply couple the three types of parameters of the same pixel to generate pixel-level feature units. Finally, the matrix integration layer integrates the feature units into a global glare feature matrix according to the spatial region division rules, completing the entire process of "coordinate mapping-feature extraction-parameter binding-matrix integration", providing structured and highly correlated data support for subsequent glare calculation.
[0050] Specifically, the dynamic threshold adaptive strategy adapts to different lighting scenarios in the following ways: a light intensity detection component is set up to capture scene light fluctuation signals in real time, and instantaneous light interference is eliminated through signal filtering; a multi-scene threshold library is preset to store the fusion threshold range corresponding to different light intensity scenes; based on the light detection results, the threshold range in the multi-scene threshold library is called to adjust the threshold parameters of the fusion operation to adapt to the dynamic changes in scene light.
[0051] Specifically, the binocular disparity consistency verification mechanism verifies the validity of the data by: extracting image data acquired by the binocular lens, calculating the disparity data of all pixels, and generating a disparity map; establishing disparity consistency evaluation rules and setting the disparity fluctuation range in combination with the spatial structure characteristics of the target scene; comparing the disparity data with the disparity fluctuation range, marking abnormal pixels that exceed the range, and initiating a data re-sampling process when the proportion of abnormal pixels exceeds a preset standard.
[0052] This embodiment is applied to urban road glare detection scenarios. The binocular camera assembly is installed on a roadside pole, and the detection targets are oncoming vehicle headlights, streetlights, and other light sources, as well as surrounding environmental objects. The camera parameters are pre-calibrated using a checkerboard calibration method: the baseline distance is set to 12cm (adapting to a road detection range of 10-150m, referring to "GB / T38265-2019 Technical Requirements for Machine Vision Camera Interfaces"), the focal length is calibrated to 8mm, the pixel equivalent is 0.01mm / pixel, and the intrinsic parameters (principal point coordinates, distortion coefficients) are stored in the system parameter library, providing a basis for disparity calculation.
[0053] First, extract the left and right image data acquired synchronously by the binocular lenses, and process the images using a block matching stereo matching algorithm (the industry's mainstream binocular disparity calculation scheme): divide the left image into image blocks of fixed size, and find the matching block with the highest gray-level similarity within the corresponding search range of the right image. Calculate the disparity data of each pixel by using the column coordinate difference of the matching blocks, combined with the calibrated focal length, baseline distance, and pixel equivalent. After traversing all pixels in the domain, a complete disparity map is generated. The range of disparity data is positively correlated with the detection distance (approximately 96 pixels of disparity at 10m and approximately 6.4 pixels of disparity at 150m).
[0054] Subsequently, a disparity consistency assessment rule was established: Considering the spatial structural characteristics of road scenes (targets such as road surfaces, vehicles, and streetlights all possess continuous spatial forms, and disparity should exhibit a gradual distribution), a disparity fluctuation range was set. Using adjacent 8×8 pixel blocks as units, the rule stipulated that the difference between the maximum and minimum disparity values within the same pixel block should not exceed 3 pixels, and the disparity change between adjacent pixel blocks should not exceed 2 pixels. This range was derived from statistical analysis of 1000 sets of measured disparity data from road scenes, effectively filtering out isolated outliers caused by noise.
[0055] Finally, the disparity data of each pixel in the disparity map is compared with the preset fluctuation range: pixels exceeding the difference threshold within a single pixel block or the change threshold between adjacent blocks are marked as abnormal pixels. The proportion of abnormal pixels to the total number of pixels in the disparity map is calculated, with a preset standard of 5% (this threshold was determined through testing with multiple sets of noisy images; a proportion below this will not affect the accuracy of subsequent glare calculations). When the statistical proportion exceeds 5%, it is determined that the current acquired data is insufficient in effectiveness due to sudden changes in lighting, lens contamination, or target occlusion. The system automatically initiates the data re-acquisition process, re-triggers the synchronous exposure of the binocular lenses to acquire images, ensuring the reliability of the disparity data input into subsequent processing stages.
[0056] Specifically, the method for dynamically correcting detection results by combining historical detection data from similar scenarios involves: classifying and storing detection data and correction records for different scenarios to establish a historical data storage system; employing data association retrieval technology to retrieve historical data from similar scenarios based on the features of the current scenario, extracting correction patterns from the historical data, generating the correction direction and magnitude of the current detection results, and optimizing and adjusting the detection results. The data association retrieval technology is a targeted data mining technique based on scene feature similarity matching. Its core is to accurately filter reusable historical correction data by quantifying the feature correlation between the current scenario and historical scenarios. Key operations include: first, extracting the core feature dimensions of the current scenario (such as ambient light intensity, target type, distance range, and original parameter distribution characteristics), converting these features into standardized retrieval vectors; then, traversing the historical data storage system, calculating the matching degree between the retrieval vector and the feature vectors of each historical scenario using algorithms such as cosine similarity and Euclidean distance; filtering similar scenario data with matching degrees higher than a preset threshold, associating their corresponding detection results and correction records, and extracting the mapping pattern of "scene features - deviation type - correction magnitude" to provide data support for the correction direction and quantitative adjustment of the current detection results.
[0057] Specifically, the dynamic adjustment of system operating parameters based on the correction results is carried out in the following manner: a feedback classification mapping rule is established, and the correction results are divided into acquisition, calibration, and fusion parameters according to the parameter functional attributes; after each preceding module receives the corresponding category of correction parameters, it compares the differences between the original operating parameters and the correction parameters; based on the difference analysis results, the corresponding acquisition strategy, calibration rule, or fusion parameters in the module are updated to form a "detection-correction-optimization" cyclic optimization link.
[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations 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 scope of the present invention.
Claims
1. A glare detection system based on binocular vision and brightness fusion, characterized in that, include: Binocular synchronous acquisition module, multi-parameter correlation calibration module, binocular-brightness fusion calculation module, glare result output and verification module; The binocular synchronous acquisition module is equipped with a binocular camera component. It captures the original parameters of distance, size and brightness of the target scene by synchronously triggering the exposure mechanism through the left and right lenses. It performs a time-series synchronous calibration operation between the brightness channel and the binocular vision channel. It dynamically adjusts the exposure parameters in combination with the real-time brightness feedback of the scene to generate a target scene original parameter set containing multi-dimensional original information. The multi-parameter correlation calibration module acquires the original parameter set of the target scene, performs mutual calibration operations on distance and size parameters based on spatial coordinate mapping, performs attenuation compensation calibration on brightness parameters in combination with distance factors, and removes invalid correlation data through a time-series alignment algorithm to generate a standardized calibration parameter set. The binocular-brightness fusion calculation module acquires the standardized calibration parameter set, assigns dynamic weights to distance, size, and brightness parameters using a weighted confidence fusion mechanism, establishes a spatial pixel-level coupled calculation model for deep parameter binding, and adapts to different lighting scenarios through a dynamic threshold adaptive strategy to generate comprehensive parameters that can fully reflect glare characteristics. The glare result output and verification module acquires the comprehensive parameters, calculates the glare index and spatial distribution information of the glare area of the target scene, verifies the validity of the data through a binocular parallax consistency verification mechanism, dynamically corrects the detection results by combining historical detection data of similar scenes, and dynamically adjusts the system operating parameters based on the correction results.
2. The system according to claim 1, characterized in that, The synchronous exposure triggering mechanism for the left and right lenses includes: setting up a co-source trigger control unit to generate a synchronous trigger signal and transmitting it in parallel to the exposure drive module of the binocular lens; constructing an exposure timing capture link to synchronously acquire the action timing of the start and end of exposure for both lenses; comparing the action timing differences between the two lenses and dynamically adjusting the transmission delay of the trigger signal.
3. The system according to claim 1, characterized in that, The specific method for performing the timing synchronization calibration operation of the brightness channel and the binocular vision channel is as follows: construct a unified timing reference unit, generate a stable calibration time axis, and use dual-channel delay calibration technology to capture the acquisition start delay and data transmission delay of the brightness channel and the binocular vision channel; Based on the two types of time delay differences, a timing compensation signal is injected into the data receiving end to correct the timestamp deviation between the brightness channel and the binocular vision channel data.
4. The system according to claim 1, characterized in that, The method of dynamically adjusting exposure parameters by combining real-time scene brightness feedback is as follows: configure a high-sensitivity brightness sensing component to capture scene brightness change signals in real time, divide the scene brightness change range into multiple adjustment ranges based on the scene brightness change range, and formulate brightness graded response rules; according to the range to which the brightness sensing signal belongs, call the brightness graded response rules to adjust the lens exposure time and gain, and balance the brightness acquisition accuracy and parallax matching requirements.
5. The system according to claim 1, characterized in that, The distance-size parameter mutual calibration operation based on spatial coordinate mapping is specifically performed as follows: the three-dimensional spatial coordinate set of the target scene is calculated using the principle of binocular parallax, and the three-dimensional coordinate information of key points of the target contour is extracted; a coordinate-size association mapping table is constructed, and the spatial straight-line distance is calculated by using the three-dimensional coordinates of adjacent vertices, inflection points and edge feature points, and the actual size reference value of the target is inferred. By comparing the reference value with the original size parameters, calibration coefficients corresponding to each point are generated. The original size parameters are then corrected point by point based on the target pixel distribution, thereby strengthening the spatial correlation between the size parameters and the distance parameters.
6. The system according to claim 1, characterized in that, The distance attenuation compensation calibration of the brightness parameter is specifically performed as follows: extract distance information from the original parameter set of the target scene, divide different distance intervals according to the distance gradient, set differentiated compensation coefficients based on the brightness attenuation characteristics corresponding to the different distance intervals, and calculate the original brightness parameter with the compensation coefficient of the corresponding interval to correct the brightness deviation caused by the distance factor.
7. The system according to claim 1, characterized in that, The method for removing invalid associated data using the time-series alignment algorithm is as follows: obtain the timestamp information of each parameter at the time of collection, establish a timestamp association matrix, and preset a time-series association threshold to define the time difference range of valid associations; Compare the timestamp differences of each parameter and mark the data combinations that exceed the threshold range; remove the data combinations from the original parameter set and retain the parameters that match the time sequence.
8. The system according to claim 1, characterized in that, The weighted confidence fusion mechanism is used to assign dynamic weights to the three types of parameters. Specifically, a confidence assessment system is generated by evaluating the disparity matching confidence of the distance parameter, the spatial fitting confidence of the size parameter, and the acquisition stability of the brightness parameter. Based on the three types of confidence assessment results, the weights of each parameter are calculated according to preset association rules. Build a real-time weight update link to dynamically refresh the weight allocation results.
9. The system according to claim 1, characterized in that, The method for establishing a spatial pixel-level coupled calculation model for parameter deep binding is as follows: mapping the three-dimensional spatial coordinates of the target scene obtained by binocular parallax calculation to image pixels one by one, generating a pixel-spatial coordinate bidirectional association table, obtaining the spatial location of image pixels, extracting the brightness parameters and size information corresponding to the image pixels, and establishing a pixel-level multi-dimensional feature set containing spatial coordinates, brightness values, and actual size. By using spatial location correlation, multiple parameters of the same pixel are deeply bound together to form pixel-level feature units, and then divided into a global glare feature matrix based on spatial region.
10. The system according to claim 1, characterized in that, The dynamic threshold adaptive strategy adapts to different lighting scenarios. Specifically, it involves setting up a light intensity detection component to capture scene light fluctuation signals in real time and filtering out instantaneous light interference through signal filtering. A multi-scene threshold library is preset to store the fusion threshold ranges corresponding to different light intensity scenes; Based on the illumination detection results, the threshold range in the multi-scene threshold library is called to adjust the threshold parameters of the fusion operation to adapt to the dynamic changes in scene illumination.
11. The system according to claim 1, characterized in that, The binocular disparity consistency verification mechanism verifies the validity of the data. The specific method is as follows: extract the image data acquired by the binocular lens, calculate the disparity data of the entire pixel point, and generate a disparity map; establish disparity consistency evaluation rules, and set the disparity fluctuation range in combination with the spatial structure characteristics of the target scene. Compare the disparity data with the disparity fluctuation range, mark abnormal pixels that exceed the range, and start the data re-sampling process when the proportion of abnormal pixels exceeds the preset standard.
12. The system according to claim 1, characterized in that, The method for dynamically correcting the detection results by combining historical detection data from similar scenarios is as follows: classify and store detection data and correction records for different scenarios, and establish a historical data storage system. By employing data association retrieval technology, historical data of similar scenarios are retrieved based on the characteristics of the current scenario. Correction patterns in the historical data are extracted to generate the correction direction and magnitude of the current detection results, thereby optimizing and adjusting the detection results.