AI automatic distance measurement method based on infrared image model

Abstract

The invention relates to the field of infrared image processing, and discloses an AI automatic distance measurement method based on an infrared image model, which comprises the following steps: deploying a high-precision thermal infrared imager, and collecting multi-scene and multi-distance infrared images; preprocessing the original infrared image; establishing a data set containing a plurality of typical scenes according to the preprocessed image, and marking the image; an improved yolov5 model is adopted to train the data set, and accurate target detection frame information is obtained; sampling different environment temperature image samples and corresponding actual physical distance data to form a ranging training sample; inputting the accurate target detection frame information and the distance measurement training sample into a designed Gaussian radial basis function neural network model for training to obtain a physical distance measurement model, and compensating the error of the system by using the physical distance measurement model to obtain the distance between the target and the thermal infrared imager; the method has the beneficial effects that the ranging precision is improved, the environmental adaptability is enhanced, the real-time performance is excellent, and the deployment is simple and convenient.

Description

Technical Field

[0001] The present invention relates to the field of infrared image processing, and in particular to an AI automatic ranging method based on an infrared image model. Background Art

[0002] Monocular ranging technology has a wide range of applications in various fields. In industrial manufacturing, it can be used to measure the size and position of parts, enabling automated production. In the medical field, it can be used to measure a patient's height and weight, assisting in diagnosis and treatment. In the military, it can be used to determine the distance and size of targets, guiding shooting and reconnaissance missions. In consumer electronics, it supports new human-computer interaction methods such as gesture recognition and virtual reality.

[0003] Despite significant progress in existing ranging technology, several challenges remain. For example, variations in ambient lighting conditions can significantly impact measurement accuracy. Differences in surface reflectivity can lead to system errors. Furthermore, occlusion and interference in complex scenes can reduce system robustness. These factors combine to make existing ranging technology's accuracy and stability insufficient for practical applications. Therefore, algorithmic innovation and sensor technology upgrades are urgently needed to improve the system's measurement accuracy and environmental adaptability. Summary of the Invention

[0004] The purpose of the present invention is to propose an AI automatic ranging method based on an infrared image model to solve the technical problems of low ranging accuracy and poor stability of existing monocular ranging methods.

[0005] Specifically, the present invention provides an AI automatic ranging method based on an infrared image model, comprising the following steps:

[0006] S1. Deploy high-precision infrared thermal imagers to collect infrared images of multiple scenes and distances;

[0007] S2, preprocessing the original infrared image to obtain a preprocessed image;

[0008] S3. Create a data set containing a variety of typical scenes based on the preprocessed images and annotate the images;

[0009] S4. Use the improved yolov5 model to train the data set to obtain accurate target detection box information;

[0010] S5. Sampling image samples of different ambient temperatures and their corresponding actual physical distance data to form distance measurement training samples;

[0011] S6. Input the precise target detection frame information and ranging training samples into the designed Gaussian radial basis function neural network model for training to obtain a physical ranging model. The physical ranging model is used to compensate for the system error and obtain the distance between the target and the infrared thermal imager.

[0012] A storage medium stores instructions and data for implementing an AI automatic ranging method based on an infrared image model.

[0013] An AI automatic ranging method and device based on an infrared image model includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement an AI automatic ranging method based on an infrared image model.

[0014] The beneficial effects provided by the present invention are:

[0015] High-precision target detection and enhanced adaptability: By improving the YOLOv5 model and introducing the temperature-aware CBAM module, the feature expression of high-temperature targets in infrared images is effectively enhanced, significantly improving the target detection accuracy and robustness in complex scenarios (such as rain and fog, nighttime, and high temperature interference).

[0016] Dynamically optimize training efficiency: Adopting an improved IoU Loss loss function, by dynamically adjusting the gradient penalty intensity, it accelerates the convergence speed of low IoU samples (difficult samples) while suppressing the overfitting risk of high IoU samples, significantly improving model training efficiency and generalization ability.

[0017] Accurate ranging and error compensation: A physical ranging model is constructed by combining a Gaussian radial basis function neural network and a system error compensation mechanism is introduced to effectively solve the nonlinear error problem in the geometric ranging model, achieving high stability and high-precision distance estimation under various ambient temperatures.

[0018] Multi-scenario adaptability: By establishing a dataset covering typical scenarios such as urban, outdoor, and rainy and foggy scenes, and combining it with preprocessing techniques (non-uniformity compensation and dynamic range adjustment), the algorithm's adaptability to complex lighting, differences in reflective characteristics, and occlusion environments is significantly improved, meeting the diverse needs of practical applications.

[0019] Collaborative optimization of hardware and algorithms: Through the deep integration of high-precision infrared thermal imagers and AI models, an end-to-end automated ranging process is achieved, reducing the cost of manual intervention. It is suitable for real-time ranging needs in multiple fields such as industry, military, and medical care. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0021] Figure 2It is a schematic diagram of the working of the hardware device of an embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0023] Before formally explaining the present invention, the scheme of the present invention is first generally explained for easy understanding.

[0024] Please refer to Figure 1 The present invention provides an AI automatic ranging method based on an infrared image model, comprising:

[0025] S1. Deploy high-precision infrared thermal imagers to collect infrared images of multiple scenes and distances;

[0026] It should be noted that, in the present invention, an infrared thermal imager needs to be installed on the equipment to collect infrared videos containing targets at different distances and in different scenes and obtain frame images.

[0027] S2, preprocessing the original infrared image to obtain a preprocessed image;

[0028] Specifically, the image preprocessing process includes: non-uniformity correction to eliminate the response differences of each pixel of the detector, image noise reduction and enhancement to reduce the impact of image noise and improve image details, dynamic range compression to enhance image contrast, and normalization, rotation and cropping of the acquired data to enrich the image data set.

[0029] S3. Create a data set containing a variety of typical scenes based on the preprocessed images and annotate the images;

[0030] It should be noted that the typical scenarios in step S3 include: urban, outdoor, rainy and foggy, nighttime and high temperature interference.

[0031] S4. Use the improved yolov5 model to train the data set to obtain accurate target detection box information;

[0032] In the present invention, an improved yolov5 target detection network is used to train the acquired training data set, detect targets in the image, and output the target's bounding box information in real time according to the detection results.

[0033] Temperature-aware CBAM is introduced after the C3 module of the Yolov5 network backbone to enhance high-level semantic features. The model can adaptively enhance the feature expression of high-temperature targets in infrared images and significantly improve detection.

[0034] The mathematical expression of the temperature sensing CBAM module is as follows:

[0035] Assume the input feature map is Global feature compression is achieved through global average pooling GAP and global maximum pooling GMP, as shown in formula (1):

[0036]

[0037] The weights are generated through the MLP containing the bottleneck structure, as shown in formula (2):

[0038] M c (F)=σ(W1(W0(GAP(F))))+W1(W0(GMP(F)))(2)

[0039] The loss function in the model training phase is optimized. Based on the existing IoU Loss, the proposed α-Iou Loss is introduced to dynamically adjust the penalty intensity of bounding box regression by introducing the exponential parameter α to improve the accuracy and robustness of the target detection model. The optimized loss function is shown in formula (3):

[0040]

[0041] When α=1, it degenerates into standard IoU Loss.

[0042] When α>1, a stronger gradient penalty is imposed on low IoU samples (difficult samples) to accelerate convergence.

[0043] When 0<α<1, reduce the gradient of high IoU samples to avoid overfitting.

[0044] S5. Sampling image samples of different ambient temperatures and their corresponding actual physical distance data to form distance measurement training samples;

[0045] It should be noted that the infrared thermal imager will be used to collect and measure the distance between the infrared thermal imager and the target. Multiple sets of samples at different angles will be collected at each distance point, and a total of 100 sets of images and data on the actual physical distance of the target in the corresponding image will be sampled.

[0046] S6. Input the precise target detection frame information and ranging training samples into the designed Gaussian radial basis function neural network model for training to obtain a physical ranging model. The physical ranging model is used to compensate for the system error and obtain the distance between the target and the infrared thermal imager.

[0047] It should be noted that the Gaussian radial basis function neural network design includes: using a deep learning-based target detection algorithm to obtain the target's border information (bounding box center coordinates (u, v), width w, height h)), inputting it into the radial basis function neural network, and training the radial basis function neural network through supervised learning to select the center method, thereby constructing a ranging model.

[0048] Furthermore, the ranging algorithm in this paper mainly utilizes the ranging idea of ​​data regression modeling, uses the hidden layer of the radial basis function neural network to solve the error problem of the geometric ranging model, and adopts a ranging model based on the radial basis function neural network.

[0049] The value of the radial basis function depends only on the distance from the origin, and the function value is a non-negative real number. The radial basis function uses a Gaussian kernel function, as shown in formula (4):

[0050]

[0051] Among them, x represents the input, r is the center point of the Gaussian kernel function, σ is the variance of the Gaussian kernel function, is the norm of xr, and ||xr|| represents the distance from x to r.

[0052] The final output expression of the radial basis function neural network is shown in formula (5):

[0053]

[0054] Among them, x is the output of the radial basis function neural network, y is the output of the radial basis function neural network, r l is the center of the l hidden layer neurons of the Gaussian kernel function (determined by the least squares method), L is the total number of centers, w l is the weight between the lth center point and the output layer, and σ is the width parameter.

[0055] The distance estimate d is the weighted sum of the hidden layer outputs, as shown in formula (6):

[0056]

[0057] The mean square error is used to optimize the weights, as shown in formula (7):

[0058]

[0059] Where: λ is the regularization coefficient to prevent overfitting.

[0060] See Figure 2 , Figure 24 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: an AI automatic ranging method device 401 based on an infrared image model, a processor 402 and a storage medium 403.

[0061] An AI automatic ranging method and device based on an infrared image model 401: The AI ​​automatic ranging method and device based on an infrared image model 401 implements the AI ​​automatic ranging method based on an infrared image model.

[0062] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the AI ​​automatic ranging method based on the infrared image model.

[0063] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the AI ​​automatic ranging method based on the infrared image model.

[0064] As a specific embodiment, vehicle ranging in urban traffic monitoring, especially in rainy and foggy weather, is easily affected by visibility, resulting in reduced accuracy.

[0065] Implementation steps:

[0066] Data acquisition and preprocessing: High-precision infrared thermal imagers were deployed at traffic intersections to collect infrared images of vehicles at different distances (10m to 100m) in rainy and foggy weather, obtaining a total of 500 sets of image samples.

[0067] Non-uniformity compensation is performed on the original image to eliminate the difference in detector pixel response; image contrast is enhanced by dynamic range adjustment, and random rotation and cropping are added to expand the dataset.

[0068] Object detection model training: Vehicle locations in the dataset were annotated to construct labeled data covering rainy, foggy, and urban scenes. Training was performed using a modified YOLOv5 model (Backbone's C3 module with the temperature-aware CBAM). The optimized IoU loss dynamically adjusted the gradient parameters. Training results showed that the model achieved a detection accuracy (mean Average Precision) of 92.5% in rainy and foggy scenes, an 8% improvement over the baseline model.

[0069] Ranging model construction and validation: 100 sets of infrared images of vehicles in rainy and foggy environments, along with their actual distance data (calibrated by lidar), were collected to form the ranging training sample. The target detection box information (center coordinates, width, and height) was input into a Gaussian radial basis function neural network. The weights were optimized using mean squared error (MSE), and a regularization term (λ = 0.01) was added to prevent overfitting. After error compensation, the model achieved an average ranging error of 0.3m (standard deviation ±0.05m), significantly improving upon the traditional monocular ranging method (average error of 1.2m).

[0070] Effect verification: In rainy and foggy weather, the system successfully achieved real-time vehicle ranging, and the output distance data was highly consistent with the lidar calibration value, verifying the reliability and practicality of the invention in harsh environments.

[0071] The beneficial effects of the present invention are:

[0072] High-precision target detection and enhanced adaptability: By improving the YOLOv5 model and introducing the temperature-aware CBAM module, the feature expression of high-temperature targets in infrared images is effectively enhanced, significantly improving the target detection accuracy and robustness in complex scenarios (such as rain and fog, nighttime, and high temperature interference).

[0073] Dynamically optimize training efficiency: Adopting an improved IoU Loss loss function, by dynamically adjusting the gradient penalty intensity, it accelerates the convergence speed of low IoU samples (difficult samples) while suppressing the overfitting risk of high IoU samples, significantly improving model training efficiency and generalization ability.

[0074] Accurate ranging and error compensation: A physical ranging model is constructed by combining a Gaussian radial basis function neural network and a system error compensation mechanism is introduced to effectively solve the nonlinear error problem in the geometric ranging model, achieving high stability and high-precision distance estimation under various ambient temperatures.

[0075] Multi-scenario adaptability: By establishing a dataset covering typical scenarios such as urban, outdoor, and rainy and foggy scenes, and combining it with preprocessing techniques (non-uniformity compensation and dynamic range adjustment), the algorithm's adaptability to complex lighting, differences in reflective characteristics, and occlusion environments is significantly improved, meeting the diverse needs of practical applications.

[0076] Collaborative optimization of hardware and algorithms: Through the deep integration of high-precision infrared thermal imagers and AI models, an end-to-end automated ranging process is achieved, reducing the cost of manual intervention. It is suitable for real-time ranging needs in multiple fields such as industry, military, and medical care.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An AI automatic ranging method based on an infrared image model, characterized by: include: S1. Deploy high-precision infrared thermal imagers to collect infrared images of multiple scenes and distances; S2, preprocessing the original infrared image to obtain a preprocessed image; S3. Create a data set containing a variety of typical scenes based on the preprocessed images and annotate the images; S4. Use the improved yolov5 model to train the data set to obtain accurate target detection box information; S5. Sampling image samples of different ambient temperatures and their corresponding actual physical distance data to form distance measurement training samples; S6. Input the precise target detection frame information and ranging training samples into the designed Gaussian radial basis function neural network model for training to obtain a physical ranging model. The physical ranging model is used to compensate for the system error and obtain the distance between the target and the infrared thermal imager.

2. The AI ​​automatic ranging method based on an infrared image model according to claim 1, characterized in that: The preprocessing includes: non-uniformity compensation and dynamic range adjustment.

3. The AI ​​automatic ranging method based on infrared image model according to claim 1, characterized in that: The various typical scenarios in step S3 include: urban, outdoor, rainy and foggy, nighttime, and high temperature interference.

4. The AI ​​automatic ranging method based on infrared image model according to claim 1, characterized in that: The improved yolov5 model described in step S4 specifically refers to: introducing temperature-aware CBAM after the C3 module of the yolov5 network backbone to enhance high-level semantic features.

5. The AI ​​automatic ranging method based on infrared image model according to claim 1, characterized in that: During training in step S4, the optimized IoU Loss loss function is used, as follows: Among them, IoU represents the intersection-over-union ratio, which measures the difference between the predicted box B and the real box B. g t The degree of overlap measures the overlap between the predicted box and the real box. The larger the value, the better the match. C represents the context area, which is the entire image or the local area where the target may appear in the image. α is a custom power parameter used to adjust the sensitivity of the IoU term. Among them, when α = 1, it degenerates to the standard IoU Loss; When α>1, a stronger gradient penalty is imposed on low IoU samples to accelerate convergence; When 0<α<1, reduce the gradient of high IoU samples to avoid overfitting.

6. The AI ​​automatic ranging method based on infrared image model according to claim 1, characterized in that: The designed Gaussian radial basis function neural network model specifically refers to: inputting precise target detection box information into the radial basis function neural network, training the radial basis function neural network through supervised learning and selecting the center method, thereby constructing a ranging model.

7. The AI ​​automatic ranging method based on infrared image model according to claim 6, characterized in that: The radial basis function uses a Gaussian kernel function, as shown in the following formula: Among them, x represents the input, r is the center point of the Gaussian kernel function, σ is the variance of the Gaussian kernel function, L is the norm of xr, and ||xr|| represents the distance from x to r.

8. A storage medium, characterized in that: The storage medium stores instructions and data for implementing an AI automatic ranging method based on an infrared image model as described in any one of claims 1 to 7.

9. An AI automatic ranging method and device based on an infrared image model, characterized by: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement an AI automatic ranging method based on an infrared image model as described in any one of claims 1 to 7.

Images