A cross-scale infrared pedestrian detection method, system, medium, device and terminal

Abstract

The application belongs to the technical field of pedestrian detection, and discloses a cross-scale infrared pedestrian detection method, system, medium, equipment and terminal, constructs a brightness perception module, obtains a reconstructed background image and a brightness feature image by combining a quadtree-Bessel interpolation operator and a guided filter, designs an EG-Chimp optimization model, iteratively optimizes a constructed objective function, obtains a dynamic feature optimization image, introduces a BiFPN feature pyramid structure under a YOLOv5 algorithm framework, designs a cross-scale feature fusion module to learn weight parameters between different scale features, and introduces a CIOU regression loss function to realize a cross-scale infrared pedestrian detection task. The application cross-fuses sensor technology and target detection algorithms, improves the pedestrian target feature expression capability of a detection network, improves the infrared pedestrian detection performance under multi-scale and partial occlusion conditions, and can realize cross-scale infrared pedestrian detection tasks in various complex environments.

Description

Technical Field

[0001] This invention belongs to the field of pedestrian detection technology, and particularly relates to a cross-scale infrared pedestrian detection method, system, medium, device and terminal. Background Technology

[0002] Currently, pedestrian detection technology aims to use computer vision methods to determine the presence and precise location of pedestrian targets in images or video sequences. Because pedestrians possess both rigid and flexible characteristics, their appearance is easily affected by clothing, size, occlusion, posture, and viewing angle, making pedestrian detection a challenging yet highly valuable research topic in the field of computer vision. Currently, pedestrian detection technology based on visible light / thermal infrared images is widely used in accident warning, intelligent monitoring, vehicle-assisted driving, and human behavior analysis.

[0003] Visible light images, based on the principle of optical reflection, possess rich detail and texture information and conform to the human visual perception system. However, they are highly susceptible to various environmental factors such as changes in lighting conditions, making it difficult to accurately identify pedestrian targets at night or in adverse environments such as heavy rain or fog. Infrared images, due to their unique thermal radiation imaging characteristics, are less affected by changes in lighting and weather conditions, possess strong anti-interference capabilities, and can meet the needs of all-weather detection tasks. Therefore, pedestrian detection technology based on thermal infrared images is of significant research importance and is also a research hotspot.

[0004] Traditional infrared pedestrian detection algorithms mainly use sliding window technology to extract human features and combine them with a classifier to complete the pedestrian detection task. Although such algorithms can achieve accurate target detection in certain specific scenarios, they are difficult to capture the high-level semantic information of infrared pedestrian targets and rely heavily on expert experience in the human feature extraction stage. They also suffer from problems such as poor real-time performance, weak generalization ability, and poor robustness.

[0005] In recent years, with the rapid development of artificial intelligence, computer vision technology has integrated multiple disciplines such as sensors, image processing, and pattern recognition. It utilizes computers and cameras to replace the human brain and eyes in tasks such as target recognition, tracking, and measurement, offering advantages such as high automation, low cost, ease of deployment, and strong real-time monitoring capabilities, thus meeting the current higher requirements for pedestrian detection. Infrared pedestrian detection systems based on computer vision employ deep learning-based algorithms, automatically learning target features using the excellent feature extraction and nonlinear fitting capabilities of convolutional neural networks. This results in excellent detection performance in terms of accuracy, speed, and generalization ability. These algorithms are generally classified into two categories: two-stage detection methods and single-stage detection methods.

[0006] Two-stage detection methods first identify candidate regions, then perform classification and regression on the detection results within these regions. Representative algorithms include Region-Based Convolutional Neural Networks (R-CNN) and Faster Region-Based Convolutional Neural Networks (Faster R-CNN). These algorithms achieve high detection accuracy, but because the backbone feature extraction network extracts only a single layer of feature maps, its ability to represent infrared pedestrian target features is insufficient. Furthermore, determining candidate regions takes time, resulting in generally poor real-time performance and limiting the performance of infrared pedestrian target detection. Single-stage detection methods, based on regression principles, directly use the detection network to predict the probability of pedestrian target category and location coordinates, thus significantly improving the detection speed. Representative algorithms include SSD (Single Shot MultiBox Detection) and the YOLO (You Only Look Once) series. These algorithms can perform real-time target detection tasks well, but due to the thermal sensitivity limitations of infrared sensors and the multi-scale and partially occluded characteristics of infrared pedestrian targets, they are prone to being poorly represented in complex environments, thus affecting detection performance. While deep learning-based infrared pedestrian detection algorithms demonstrate significant advantages over traditional machine learning algorithms in terms of detection accuracy and real-time performance, their detection performance is typically positively correlated with the number of model parameters and the network depth. The larger the number of model parameters and the deeper the network, the greater the computational cost and the longer the runtime, leading to a trade-off between real-time network efficiency and algorithm accuracy.

[0007] Therefore, how to improve the feature representation capability of infrared pedestrian targets, how to improve the detection performance of infrared pedestrian targets under partial occlusion, and how to balance the real-time efficiency of the detection network and the accuracy of the algorithm are the technical problems that urgently need to be solved at this stage.

[0008] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0009] (1) Existing pedestrian detection technologies based on visible light images are easily affected by various environmental factors such as changes in lighting. They are difficult to accurately identify all pedestrian targets at night or in adverse environments such as heavy rain and fog, resulting in insufficient reliability of detection results.

[0010] (2) Traditional infrared pedestrian detection algorithms are difficult to capture high-level semantic information of infrared pedestrian targets and rely heavily on expert experience in the manual feature extraction stage. They have problems such as poor real-time performance, weak generalization ability and poor robustness, and their detection effect is often difficult to guarantee.

[0011] (3) Existing two-stage detection methods are limited in their ability to express the features of infrared pedestrian targets because the backbone feature extraction network only extracts a single-layer feature map. Furthermore, it takes a certain amount of time to determine the candidate region, resulting in generally poor real-time performance.

[0012] (4) In the existing single-stage detection method, the infrared sensor is limited by thermal sensitivity, the pedestrian target texture features are weak and there is edge blurring, which is easily submerged in the redundant background, thus making it difficult to effectively express the infrared target features and affecting the performance of the detection network.

[0013] (5) In dense multi-target scenarios, infrared pedestrian targets have multi-scale and partial occlusion characteristics, which leads to the deformation or even loss of infrared pedestrian features, posing a certain challenge to infrared pedestrian detection.

[0014] (6) The detection performance of the algorithm is usually positively correlated with the number of detection model parameters and the network depth. However, the larger the number of model parameters and the deeper the network, the greater the computational cost and the longer the running time of the algorithm. Therefore, how to enhance the detection performance of the algorithm without increasing the computational cost is a major challenge for infrared pedestrian detection technology. Summary of the Invention

[0015] To address the problems existing in the prior art, this invention provides a cross-scale infrared pedestrian detection method, system, medium, device, and terminal, and particularly relates to a cross-scale infrared pedestrian target detection method, system, medium, device, and terminal under a dynamic feature optimization mechanism.

[0016] This invention is implemented as follows: a cross-scale infrared pedestrian detection method, the cross-scale infrared pedestrian detection method comprising:

[0017] A brightness perception module is constructed, which combines a quadtree-Bessel interpolation operator and a guided filter to obtain reconstructed background and brightness feature images. An EG-Chimp optimization model is designed, and the constructed objective function is iteratively optimized to obtain a dynamically optimized image. A BiFPN feature pyramid structure is introduced into the YOLOv5 algorithm framework to enhance the feature aggregation capability of the algorithm's neck network. A cross-scale feature fusion module is designed to learn the weight parameters between features of different scales. The loss function is optimized by introducing a CIOU regression loss function to accelerate algorithm convergence, thereby realizing the cross-scale infrared pedestrian detection task.

[0018] Furthermore, the cross-scale infrared pedestrian detection method includes the following steps:

[0019] Step 1: A brightness perception module is constructed by combining a quadtree-Bessel interpolation operator and a guided filter to obtain the reconstructed background image and brightness feature image, laying a good foundation for subsequent image adaptive optimization.

[0020] Step 2: Design the EG-Chimp optimization model to enhance the local contrast and detail expression of the input image while suppressing background information interference. In the YOLOv5 detection algorithm framework, introduce the BiFPN feature pyramid structure and enhance the neck network feature aggregation capability without increasing too much computational cost through scale skip connections.

[0021] Step 3: Construct a cross-scale feature fusion module, and use learnable weight parameters to assign weights to multi-scale features, thereby improving the pedestrian detection effect under multi-scale and partial occlusion conditions, and thus realizing the infrared pedestrian detection task in dense multi-target scenes.

[0022] Furthermore, the construction of the brightness sensing module includes:

[0023] (1) Construct a quadtree-Bessel interpolation operator, and recover a large-scale pixel matrix by adjusting some sampling control points, thereby obtaining the initial background image;

[0024] (2) The local linear fitting characteristics of guided filtering are used to smooth the edge information of the initial background image, thereby avoiding the image block effect caused by uneven sampling of edge control points;

[0025] (3) Use the linear subtraction operator to extract infrared brightness feature images.

[0026] Furthermore, the design of the EG-Chimp optimization model includes:

[0027] (1) Construct an objective function that aims to simultaneously optimize image I by taking into account dynamic features. Oir Gradient change and energy loss:

[0028] F = min(L) SF +λL CON );

[0029] In the formula, L SF L represents the spatial frequency function, which reflects abrupt changes in image space, such as edge variations. A higher L value indicates a clearer image of people or other objects. CON The content loss function can avoid the problem of insufficient preservation of detailed information due to excessive enhancement of pedestrian target brightness features during the optimization process;

[0030]

[0031]

[0032] (2) For the constructed objective function, an ape optimization model is introduced. By simulating two states of ape predation and chaotic food grabbing, the algorithm convergence is accelerated. In this way, while ensuring the spatial frequency of the dynamic feature optimized image, its content loss is suppressed, and finally the dynamic feature optimized image is obtained.

[0033] Furthermore, the BiFPN feature pyramid structure includes:

[0034] (1) Based on the original FPN-PAN pyramid structure of YOLOv5, nodes with only one input edge are deleted to reduce the number of network parameters;

[0035] (2) The neck network transmits strong semantic information from top to bottom and strong localization information from bottom to top, and then performs feature aggregation on information of different scales;

[0036] (3) By adopting scale skip connection, the feature aggregation capability of the neck network is enhanced without increasing the computational cost too much.

[0037] Furthermore, the construction of the cross-scale feature fusion module includes:

[0038] (1) Scale adjustment is performed on features of different scales. During the upsampling process, the number of channels is adjusted by 1*1 convolution and nearest neighbor interpolation is introduced to improve the resolution. During the downsampling process, 1 / 2 ratio downsampling is completed by 3*3 convolution with stride of 2, and 1 / 4 ratio downsampling is completed by adding a max pooling layer with stride of 2 before the 2-step convolution.

[0039] (2) Use 1*1 convolution to pre-estimate the weight parameters of each input feature and fine-tune them using the soft-max normalization function;

[0040] (3) Weighting the scale-adjusted input features to obtain the fused output features.

[0041] The CIOU loss function improves the regression speed of the predicted bounding box by calculating the overlap area of ​​the two bounding boxes, the distance between their center points, and their aspect ratio.

[0042]

[0043] In the formula, b and b gt Let represent the center points of the predicted bounding box and the target bounding box, respectively. ρ(·) is the Euclidean distance between the center points of the two boxes, and C represents the slant distance of the smallest rectangle covering the predicted bounding box and the target bounding box. The weighting parameter for the aspect ratio of the predicted bounding box is denoted by v, which measures the consistency between the aspect ratio of the predicted bounding box and the actual bounding box.

[0044] Another object of the present invention is to provide a cross-scale infrared pedestrian detection system applying the aforementioned cross-scale infrared pedestrian detection method, the cross-scale infrared pedestrian detection system comprising:

[0045] The brightness feature extraction module is used to construct the brightness perception module. It obtains the reconstructed background image and brightness feature image by combining the quadtree-Bessel interpolation operator and the guided filter.

[0046] The dynamic feature optimization module is used to design the EG-Chimp optimization model and iteratively optimize the constructed objective function to obtain the dynamic feature optimization image;

[0047] A cross-scale feature fusion module is used to introduce the BiFPN feature pyramid structure under the YOLOv5 algorithm framework. The cross-scale feature fusion module is designed to learn the weight parameters between features of different scales.

[0048] The infrared pedestrian detection module is used to optimize the loss function. It introduces the CIOU regression loss function and uses learnable weight parameters to assign weights to multi-scale features to achieve infrared pedestrian detection.

[0049] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the cross-scale infrared pedestrian detection method.

[0050] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the cross-scale infrared pedestrian detection method.

[0051] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned cross-scale infrared pedestrian detection system.

[0052] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:

[0053] First, addressing the technical problems existing in the prior art and the difficulty in solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0054] Pedestrian detection has become a research hotspot in computer vision in recent years. Based on deep learning algorithms, and with image and video compression, storage, transmission, and playback at its core, it holds a dominant position in current target detection technology and has been widely applied in areas such as accident warning, intelligent monitoring, vehicle-assisted driving, and human behavior analysis. However, due to the diverse poses of pedestrians and their multi-scale and partial occlusion characteristics, pedestrian detection technology has become a challenging yet highly valuable research topic in computer vision. Therefore, achieving accurate infrared pedestrian detection in partially occluded and multi-scale, dense, multi-target scenarios is of significant research importance. In traffic safety, it can prevent traffic accidents; in security monitoring, it can prevent suspicious individuals from entering; and in public place management, it can optimize human resource allocation. This invention aims to provide a new approach to pedestrian detection in complex environments.

[0055] This invention provides a cross-scale infrared pedestrian detection network with a dynamic feature optimization mechanism. Addressing the problem that thermal infrared images in complex backgrounds are affected by their own imaging mechanisms, leading to difficulties in effectively representing target features and thus impacting detection accuracy, a dynamic feature optimization mechanism is proposed. This mechanism utilizes a brightness perception module to extract brightness features from the input image and designs an EG-Chimp optimization model to iteratively optimize the constructed objective function, thereby enhancing the local contrast and detail representation of the input image while suppressing background interference. Secondly, to enhance the detection network's performance for multi-scale and partially occluded pedestrian targets, a CSFF-BiFPN structure is proposed. Its scale-skipping connection method effectively enhances the neck network's ability to detect partially occluded and multi-scale pedestrian targets. Furthermore, by constructing a cross-scale feature fusion module and using learnable weight parameters to assign weights to multi-scale features, the problem of missed and false detections of pedestrian targets in multi-scale, dense, multi-target scenes due to insufficient feature representation is effectively solved. Finally, to further accurately locate pedestrian targets, a CIOU loss function is introduced to accelerate network convergence, thereby improving detection performance.

[0056] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0057] This invention proposes a cross-scale infrared pedestrian detection network, named DOFM-CSNet, which integrates sensor technology and target detection algorithms. DOFM-CSNet aims to improve the network's ability to represent pedestrian target features by constructing a dynamic feature optimization algorithm that fuses a brightness perception module and an EG-Chimp optimization model. Furthermore, to further improve infrared pedestrian detection performance under multi-scale and partially occluded conditions, a CSFF-BiFPN feature aggregation structure is designed. This network effectively solves the problem of missed and false detections of infrared pedestrian targets caused by insufficient feature representation capabilities, enabling cross-scale infrared pedestrian detection tasks in various complex environments.

[0058] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0059] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0060] This invention, based on deep learning algorithms and integrating models and data intelligence engines, proposes an infrared pedestrian detection algorithm with a dynamic feature optimization mechanism. It can be extended to various sub-sectors of the smart city market, such as traffic safety, which can prevent traffic accidents; surveillance and security, which can prevent suspicious persons from entering and exiting; and public place management, which can optimize the allocation of human resources. It has significant research significance and commercial value.

[0061] (2) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0062] This invention addresses three major technical challenges in current pedestrian detection technology: weak feature representation capabilities for infrared pedestrian targets, difficulty in balancing detection accuracy and model lightweighting, and poor detection performance under multi-scale and partially occluded pedestrian conditions. It proposes a cross-scale infrared pedestrian detection algorithm with a dynamic feature optimization mechanism. By combining a brightness perception module and a dynamic feature optimization model, it enhances the local contrast and detail representation capabilities of the input image while suppressing background interference, thereby improving the algorithm's feature representation capabilities for infrared pedestrian targets. Furthermore, it introduces a BiFPN feature pyramid structure based on the YOLOv5 detection network, employing a scale-skipping connection approach to enhance the neck network feature aggregation capability without significantly increasing computational costs, thus improving the algorithm's detection accuracy. Finally, a cross-scale feature fusion module is constructed to learn the weight parameters between features at different scales, enabling accurate detection of infrared pedestrians at multiple scales and under partially occluded conditions in complex environments. In summary, this invention solves the pressing technical problems currently facing the field. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of the cross-scale infrared pedestrian detection method provided in an embodiment of the present invention;

[0065] Figure 2 This is a schematic diagram of the cross-scale infrared pedestrian detection method provided in an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram comparing the dynamic features before and after optimization provided in an embodiment of the present invention;

[0067] Figure 4 Figure (a) is a schematic diagram comparing the feature pyramids before and after the improvement provided in the embodiments of the present invention, and Figure (b) is a schematic diagram of the FPN+PAN feature pyramid and Figure (c) is a schematic diagram of the CSFF-BiFPN feature pyramid.

[0068] Figure 5 Figure (a) is a schematic diagram of a cross-scale feature fusion module provided in an embodiment of the present invention, and Figure (b) is a schematic diagram of a two-scale feature fusion module.

[0069] Figure 6 This is a schematic diagram of the loss comparison curves before and after the improvement provided in the embodiments of the present invention;

[0070] Figure 7 This is a schematic diagram comparing the average accuracy values ​​before and after the improvement provided in this embodiment of the invention;

[0071] Figure 8 Figure 1 is a schematic diagram comparing the detection results of different algorithms provided in the embodiments of the present invention; wherein, Figure (a) is a schematic diagram of the detection results of each algorithm in a scene with partial pedestrian occlusion, Figure (b) is a schematic diagram of the detection results of each algorithm in a multi-scale dense multi-target scene, and Figure (c) is a schematic diagram of the detection results of each algorithm in a scene with overexposure. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0073] To address the problems existing in the prior art, the present invention provides a cross-scale infrared pedestrian detection method, system, medium, device and terminal. The present invention will be described in detail below with reference to the accompanying drawings.

[0074] I. Explanatory and Illustrative Embodiments. To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory and illustrative description of the embodiments described in the claims.

[0075] like Figure 1 As shown, the cross-scale infrared pedestrian detection method provided in this embodiment of the invention includes the following steps:

[0076] S101, construct a brightness perception module, which obtains the reconstructed background image and brightness feature image by combining a quadtree-Bessel interpolation operator and a guided filter;

[0077] S102, an EG-Chimp optimization model is designed to iteratively optimize the constructed objective function, thereby enhancing the local contrast and detail expression of the input image while suppressing background information interference, and obtaining a dynamically optimized image, laying the foundation for subsequent infrared pedestrian detection tasks;

[0078] S103 introduces the BiFPN feature pyramid structure within the YOLOv5 algorithm framework to enhance the feature aggregation capability of the neck network.

[0079] S104, Design a cross-scale feature fusion module to learn the weight parameters between features of different scales, and improve the pedestrian detection performance in multi-scale and partially occluded situations;

[0080] S105 optimizes the loss function by introducing the CIOU (Complete intersection over union) regression loss function to improve the regression speed and accuracy of the predicted boxes during training, thus achieving cross-scale infrared pedestrian detection tasks.

[0081] As a preferred embodiment, such as Figure 2 As shown, the cross-scale infrared pedestrian detection method provided in this embodiment of the invention specifically includes:

[0082] 1. Dynamic Feature Optimization Mechanism

[0083] Due to the thermal radiation imaging mechanism of infrared sensors, infrared images often have lower resolution and weaker detail resolution compared to visible light images, posing a challenge to the accurate detection of pedestrian targets. To address this, this invention proposes a DFOM (Dynamic Feature Optimization Mechanism) to enhance the contrast and detail representation of pedestrian targets in input images. This mechanism consists of two parts: a brightness perception module and an EG-Chimp optimization model.

[0084] 1.1 Brightness Sensing Module

[0085] To address the challenge of accurately representing infrared pedestrian target features due to low resolution and contrast in thermal infrared images, an LPM (Luminance Perception Module) was designed. First, a quadtree-Bessel interpolation operator was constructed to recover a large-scale pixel matrix by adjusting some sampling control points, thus obtaining the initial background image. Then, to avoid image block artifacts caused by uneven sampling of edge control points, the local linear fitting characteristics of guided filtering were utilized to smooth image edge information, ultimately yielding the reconstructed background image and brightness feature image.

[0086] 1.1.1 Quadtree-Bezier Interpolation

[0087] To accurately reconstruct the infrared background image, a quadtree decomposition method is introduced, which estimates the background contour information by sampling as many control points as possible outside the infrared brightness region. First, the input image is adjusted to a size of 512×512, and then decomposition is performed using Equation (1).

[0088]

[0089] In the formula, QTD(·) represents the quadtree decomposition operator; N is the number of decompositions; I N This represents the decomposition of an image block; MIN is the minimum gray value allowed for decomposition; threshold represents the decomposition threshold, which is the maximum gray value within the image block. and minimum grayscale value Decomposition is performed when the difference is greater than the threshold; otherwise, the decomposition ends.

[0090] The quadtree decomposition algorithm can segment the input image into multiple image blocks. After uniformly sampling 16 control points from each image block, the infrared background image is reconstructed using the Bessel interpolation operator. The mathematical description is as follows:

[0091] I b (x, y) = XMPM T Y T (2)

[0092] In the formula, (x, y) represents the interpolation sampling point; P represents a 4×4 matrix composed of 16 sampling control points; X and Y represent the variable interpolation coefficients associated with the interpolation sampling point; and M represents the constant coefficient interpolation matrix. The mathematical descriptions of each parameter are shown in equations (3) to (4).

[0093]

[0094]

[0095] 1.1.2 Guided Filtering

[0096] To address the block artifact issue in the reconstructed background image caused by inconsistent selection of control points in the edge regions of different image blocks during sampling, this invention employs guided filtering to smooth the initial background image, mathematically described as follows:

[0097]

[0098] In the formula, I Bir Indicates an infrared background image; w k Represents an image patch centered at pixel k; a k and b k There are two linear parameters, and their specific values ​​are shown in equations (6) and (7), respectively.

[0099]

[0100]

[0101] In the formula, μ k and These represent the guiding image I. ir In w k The pixel mean and variance in the data; represent The pixel mean; |w| is the value of w. k The number of pixels contained within an image patch; ε represents the regularization parameter to avoid a k The value is too large.

[0102] Finally, the infrared brightness feature image I can be extracted using the linear subtraction operator. Lir The mathematical description is as follows:

[0103] I Lir =I ir -I Bir (8)

[0104] 1.2 EG-Chimp Optimization Model

[0105] To mitigate the adverse effects of complex backgrounds on accurate pedestrian target detection, this invention proposes the EG-Chimp optimization model, which generates dynamically optimized images (I) with locally high contrast through iterative optimization. Oir The mathematical description is as follows:

[0106] I Oir =α×I Lir +I Bir (9)

[0107] In the formula, α is the optimization parameter. The optimization design steps are as follows.

[0108] 1.2.1 Construction of the objective function

[0109] The objective function constructed in this invention aims to simultaneously consider the gradient changes and energy loss of the dynamically optimized image, using spatial frequency L. SF and content loss L CON express.

[0110] (1) Spatial frequency function

[0111] Spatial frequency (SF) reflects spatial abrupt changes in an image, such as edge variations. It is used to measure the gradient distribution of an image. A higher SF value indicates that the subject, such as a person, is clearer in the image. The mathematical description is as follows:

[0112]

[0113] (2) Content loss function

[0114] To avoid the problem of insufficient preservation of detailed information due to excessive enhancement of pedestrian target brightness features during the optimization process, this invention utilizes the l2 function to construct a content loss function, mathematically described as follows:

[0115]

[0116] In summary, to enhance the brightness features of pedestrians in the original infrared image while also preserving detailed information, the mathematical description of the constructed objective function is as follows:

[0117] F = min(L) SF +λL CON (12)

[0118] In the formula, λ is the regularization parameter that maintains the balance of the two terms. In this invention, λ = 1 / 500.

[0119] 1.2.2 Adaptive Optimization of Objective Function

[0120] Metaheuristic optimization algorithms, due to their simplicity, flexibility, and efficiency, have become a major technique for solving global optimization problems. However, with the increase in problem complexity, these algorithms are prone to slow convergence and getting trapped in local optima. Therefore, for the objective function constructed in this invention, i.e., formula (12), an ape optimization algorithm is introduced. By simulating two states—predation and chaotic food grabbing—the algorithm converges faster, thereby ensuring the optimization of image spatial frequency through dynamic features while suppressing content loss. It classifies apes into aggressive X-type... Attacker Drive-away type X Chaser Interceptor X Bamier and the pursuit type X Driver Four categories. During the hunt, apes can change their positions according to the location of their prey, as described mathematically below:

[0121]

[0122] In the formula, t represents the current iteration number, and X... prey Let X be the position vector of the prey. chimp Let a be the position vector of the ape, and a, c, and m be the coefficient vectors. The mathematical description is as follows:

[0123]

[0124] In the formula, f decreases nonlinearly from 2.5 to 0 during the iteration process, r1 and r2 are random vectors in the range [0, 1], and m is a chaotic vector obtained based on Gaussian chaotic mapping, representing the social phenomenon of apes fighting for food in the hunting process.

[0125] Once the prey is surrounded, the attacker is assisted by herders, interceptors, and chasers to complete the final hunting task. To improve the algorithm's convergence speed, the model simulates a chaotic food-grabbing process between humans and apes in the later stages of iteration, mathematically described as follows:

[0126]

[0127] In the formula, χ is a random parameter in [0, 1].

[0128] During the hunt, apes update their position based on the locations of attackers, herders, interceptors, and pursuers, and then attack their prey. Simultaneously, they search for prey globally by dispersing their activity. Ultimately, through multiple iterations, a dynamically optimized image can be obtained, such as... Figure 3 As shown.

[0129] 2. Neck network based on CSFF-BiFPN

[0130] 2.1 CSFF-BiFPN Pyramid Structure

[0131] To improve feature aggregation across different scales, the original Yolov5 network utilizes an FPN+PAN structure to transmit strong semantic information from bottom to top while simultaneously transmitting strong localization information from top to bottom. However, this structure directly adds and aggregates feature maps from different scales after resizing, failing to fully utilize the cross-scale information at the input, thus affecting the final detection accuracy. To address this issue, this invention constructs a CSFF-BiFPN feature pyramid structure, improving the network's feature aggregation capability through scale-skipping connections. The steps are as follows:

[0132] (1) Delete nodes with only one input edge to reduce the number of network parameters.

[0133] (2) Drawing on the FPN-PAN structure, strong semantic and strong positioning information are transmitted from two directions respectively.

[0134] (3) To further utilize the cross-scale information of the input features, the network feature aggregation capability is improved by using scale skip connections.

[0135] (4) The cross-scale feature fusion modules CSFF_2 and CSFF_3 assign weights to feature information from different inputs.

[0136] The feature pyramid structure comparison diagram provided in the embodiments of the present invention is as follows: Figure 4 As shown.

[0137] 2.2 Cross-scale feature fusion module

[0138] The FPN-PAN structure directly adds input features from different scales after scale adjustment, which can easily lead to insufficient feature fusion. To address this, this invention designs cross-scale feature fusion modules CSFF_2 and CSFF_3, which adaptively learn the weight parameters of input features at different scales. The model structure is shown below. Figure 5 .

[0139] Let I denote the input features of level i. i Where i∈[1,2,3]. First, scale adjustment is performed on features of different scales. During upsampling, a 1*1 convolution is used to adjust the number of channels and nearest neighbor interpolation is introduced to improve resolution. During downsampling, a 3*3 convolution with a stride of 2 is used to complete downsampling at a ratio of 1 / 2, and a max pooling layer with a stride of 2 is added before the two convolution steps to complete downsampling at a ratio of 1 / 4. Next, the weight parameters of each input feature are pre-estimated using a 1*1 convolution and fine-tuned using a soft-max normalization function. Finally, the scale-adjusted input features are weighted to obtain the fused output features, which are mathematically described as follows:

[0140]

[0141]

[0142] In the formula, This represents the feature information after level i is adjusted to level l. γ, ψ, δ, and η represent the learnable weight parameters obtained after normalization using the soft-max function through the fusion of features across various scales. Their mathematical description is as follows:

[0143]

[0144]

[0145] In the formula, λ γ , λ ψ , λ δ and λ η This represents the estimated weight parameters obtained using a 1*1 convolution.

[0146] The cross-scale infrared pedestrian detection system provided in this embodiment of the invention includes:

[0147] The brightness feature extraction module is used to construct the brightness perception module. It obtains the reconstructed background image and brightness feature image by combining the quadtree-Bessel interpolation operator and the guided filter.

[0148] The dynamic feature optimization module is used to design the EG-Chimp optimization model and iteratively optimize the constructed objective function to obtain the dynamic feature optimization image;

[0149] A cross-scale feature fusion module is used to introduce the BiFPN feature pyramid structure under the YOLOv5 algorithm framework. The cross-scale feature fusion module is designed to learn the weight parameters between features of different scales.

[0150] The infrared pedestrian detection module is used to optimize the loss function. It introduces the CIOU regression loss function and uses learnable weight parameters to assign weights to multi-scale features to achieve infrared pedestrian detection.

[0151] II. Application Examples. To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides application examples of the technical solution of the claims on specific products or related technologies.

[0152] This invention, based on deep learning algorithms and focusing on image and video compression, storage, transmission, and playback, achieves accurate infrared pedestrian target detection in scenarios with partial pedestrian occlusion and in multi-scale, dense, multi-target scenarios. It has been widely applied in tasks such as accident early warning, intelligent monitoring, vehicle-assisted driving, and human behavior analysis, and has significant research value. Specific application examples are analyzed below:

[0153] 1. In terms of traffic safety, integration with technologies such as the Internet of Things and big data can enable connection with urban traffic command and dispatch application platforms, thereby reducing the occurrence of traffic accidents;

[0154] 2. In terms of surveillance and security, it can enhance the coverage of information technology while ensuring urban emergency command, monitoring and dispatch, and security protection, thereby achieving key tasks such as security deployment and personnel tracking;

[0155] 3. In terms of public place management, improve and optimize the allocation of human resources to enhance the management efficiency of personnel safety and epidemic prevention and control.

[0156] III. Evidence of the Relevant Effects of the Embodiments. The embodiments of the present invention have achieved some positive effects during research and development or use, and indeed possess significant advantages compared to existing technologies. The following description, in conjunction with data, charts, and other materials from the experimental process, illustrates these advantages.

[0157] The model training and result analysis provided in this embodiment of the invention are as follows:

[0158] 1. Network model training

[0159] The hardware and software platform configuration in this invention experiment is as follows: Operating system: Microsoft Windows 10; GPU: NVIDIA GeForce GTX 1660 Ti; CPU: Intel Core i5-10400F@2.90GHz×6 CPUs; CUDA: 11.1; Deep learning framework: PyTorch.

[0160] To verify the advantages of the proposed algorithm, 1000 random sample images were selected from the KAIST dataset for experiments, containing 5384 pedestrian targets. 700 images were used for training, and the remaining 300 images were used as the test dataset. During network model training, the input image resolution was uniformly adjusted to 640×640, and asynchronous stochastic gradient descent with a momentum term of 0.937 was used for training. Each training batch contained 32 images, fed into the training network in 29 iterations. The learning rate was set to 0.01 for the first 200 iterations, and then reduced to 0.001 in the last 100 iterations as the number of iterations increased. To prevent overfitting, the weight decay regularization term was set to 5E-3, and the CIOU loss function was used to improve the regression speed of the predicted bounding boxes by calculating the overlap area, center distance, and aspect ratio of the two bounding boxes. The mathematical description is as follows:

[0161]

[0162] In the formula, b and b gt Let represent the center points of the predicted bounding box and the target bounding box, respectively. ρ(·) is the Euclidean distance between the center points of the two boxes, and C represents the slant distance of the smallest rectangle covering the predicted bounding box and the target bounding box. The weighting parameter for the aspect ratio of the predicted bounding box is denoted by v, which measures the consistency between the aspect ratio of the predicted bounding box and the actual bounding box.

[0163] 2. Image contrast analysis

[0164] To verify the effectiveness of the dynamic feature optimization mechanism proposed in this invention, entropy, Brenner gradient function, Discrete Cosine Transform (DCT), and Vollaths function were selected as objective evaluation indicators. Entropy characterizes the richness of information contained in an image; Brenner estimates its sharpness by calculating the second-order gradient; DCT highlights the image's frequency distribution information; and Vollaths assesses the degree of defocusing of an image based on the similarity between two points in space. Larger values ​​for Entropy, Brenner, DCT, and Vollaths indicate higher image contrast and stronger feature representation ability after dynamic feature optimization. 1000 randomly selected samples from the KAIST dataset were used for testing and verification, and the quantitative analysis results are shown in Table 1.

[0165] Table 1. Average Evaluation Metrics of 1000 Images

[0166]

[0167] As shown in Table 1, the dynamic feature optimization algorithm proposed in this invention is the best among the above four evaluation indicators, thus proving that the original image not only contains richer information after dynamic feature optimization, but also has the strongest contrast and the best feature expression ability.

[0168] 3. Quantitative experimental analysis

[0169] To objectively evaluate the detection performance of the algorithm of this invention, the DFOM-CSNet detection network was compared with the original YOLOv5 detection framework. The loss comparison curves are shown below. Figure 6 As shown in the figure, the average accuracy comparison curve is as follows: Figure 7 As shown.

[0170] from Figure 6 It can be seen that the YOLOv5 detection network reaches equilibrium after approximately 75 rounds and eventually stabilizes at 0.018. In contrast, the DFOM-CSNet detection algorithm proposed in this invention achieves a loss value of 0.02 after only 20 training rounds, and eventually stabilizes at around 0.009, demonstrating better training performance. Figure 7 As can be seen, the final average accuracy of the YOLOv5 detection network is approximately 0.895, while the detection accuracy of the algorithm of this invention can reach 0.913.

[0171] To quantitatively evaluate the advantages of the detection network of this invention, ten classic algorithms—Faster-RCNN, SSD, Sparse R-CNN, NAS-FCOS, TOOD, VarifocalNet, Auto-assign, YOLOv3, YOLOv4, and YOLOv5—were selected for comparative experiments on the KAIST test set. All object detection algorithms used the same data samples and parameter settings. The detection results are shown in Table 2. The best results are indicated in bold.

[0172] Table 2 Comparison results of different detection algorithms

[0173]

[0174] As shown in Table 2, the algorithm of this invention is significantly superior to the other 10 comparative algorithms in terms of detection accuracy. Although the algorithm of this invention is slightly slower than the YOLOv5 detection network in terms of detection speed, it still has a significant advantage over the other comparative algorithms.

[0175] 4. Qualitative Experimental Analysis

[0176] To further objectively evaluate the advantages of the algorithm of this invention, comparative tests were conducted on scenes with partial pedestrian occlusion, multi-scale dense multi-target scenes, and overexposed scenes. The comparison results are as follows: Figure 8As shown. For ease of observation and subsequent analysis, the missed detection and false detection areas of each algorithm have been marked with dashed boxes.

[0177] The first set of experiments shows that, except for the DFOM-CSNet detection algorithm, the other comparative algorithms all suffer from varying degrees of missed detections when handling pedestrian occlusion scenarios. The second set of experiments shows that in complex environments, Faster-RCNN, SSD, NAS-FCOS, TOOD, and YOLOv3 algorithms have significant missed detection problems. Furthermore, the SparseRCNN, VarifocalNet, Auto-assign, YOLOv4, and YOLOv5 detection frameworks lack the ability to aggregate feature information across different scales, making it difficult to accurately detect pedestrian targets at multiple scales. The DFOM-CSNet detection algorithm proposed in this invention enhances the network's feature aggregation ability by designing a CSFF-BiFPN feature pyramid structure and constructing a cross-scale fusion module to weight features at different scales, thereby effectively improving the detection performance of pedestrian targets in multi-scale, dense, and multi-target scenarios. The third set of experiments shows that in blurred nighttime scenes, except for Sparse RCNN, due to excessive daytime illumination, infrared images cannot accurately represent the detailed information of pedestrian targets, leading to numerous missed detections in various algorithms. The DFOM-CSNet detection algorithm proposed in this invention can effectively enhance the local contrast of infrared images and improve the network feature aggregation capability by designing a dynamic feature optimization mechanism, thereby improving the detection accuracy of the algorithm.

[0178] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0179] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-scale infrared pedestrian detection method, characterized in that, The cross-scale infrared pedestrian detection method includes: A brightness perception module is constructed, which combines a quadtree-Bessel interpolation operator and a guided filter to obtain reconstructed background and brightness feature images. An EG-Chimp optimization model is designed, and the constructed objective function is iteratively optimized to obtain a dynamically optimized image. A BiFPN feature pyramid structure is introduced into the YOLOv5 algorithm framework to enhance the feature aggregation capability of the algorithm's neck network. A cross-scale feature fusion module is designed to learn the weight parameters between features of different scales. The loss function is optimized by introducing a CIOU regression loss function to accelerate algorithm convergence, thereby achieving the cross-scale infrared pedestrian detection task. The design of the EG-Chimp optimization model includes: (1) Construct an objective function while taking into account dynamic features to optimize the image. Gradient change and energy loss: ； In the formula, The spatial frequency function reflects changes in spatial abruptness of an image, such as edge variations. The higher the value, the clearer the person or target in the image. Represents the content loss function; ； ； (2) For the constructed objective function, an ape optimization model is introduced. By simulating two states of ape predation and chaotic food grabbing, the dynamic feature optimization image is finally obtained.

2. The cross-scale infrared pedestrian detection method as described in claim 1, characterized in that, The cross-scale infrared pedestrian detection method includes the following steps: Step 1: Construct a brightness perception module by combining a quadtree-Bessel interpolation operator and a guided filter, and extract brightness features from the input image; Step 2: Design the EG-Chimp optimization model, build the YOLOv5 detection algorithm framework and introduce the BiFPN feature pyramid structure, while performing scale jump connections. Step 3: Construct a cross-scale feature fusion module and use learnable weight parameters to assign weights to multi-scale features to achieve infrared pedestrian detection in dense multi-target scenes.

3. The cross-scale infrared pedestrian detection method as described in claim 1, characterized in that, The construction of the brightness sensing module includes: (1) Construct a quadtree-Bessel interpolation operator, and recover a large-scale pixel matrix by adjusting some sampling control points, thereby obtaining the initial background image; (2) The edge information of the initial background image is smoothed by utilizing the local linear fitting characteristics of guided filtering; (3) Use the linear subtraction operator to extract infrared brightness feature images.

4. The cross-scale infrared pedestrian detection method as described in claim 1, characterized in that, The BiFPN feature pyramid structure includes: (1) Based on the original FPN-PAN pyramid structure of YOLOv5, delete nodes with only one input edge; (2) The neck network transmits strong semantic information from top to bottom and strong localization information from bottom to top, and then performs feature aggregation on information of different scales; (3) The ability of neck network feature aggregation is enhanced by using scale skip connections.

5. The cross-scale infrared pedestrian detection method as described in claim 1, characterized in that, The construction of the cross-scale feature fusion module includes: (1) Scale adjustment is performed on features of different scales. During the upsampling process, the following is utilized: Convolution adjusts the number of channels and introduces nearest-neighbor interpolation; during downsampling, a stride of 2 is used. Convolution complete The downsampling is performed proportionally, and this is accomplished by adding a max-pooling layer with a stride of 2 before the two convolutional steps. Proportional downsampling; (2) Utilization The convolutional parameters are used to pre-estimate the weight parameters of each input feature and then fine-tuned using a soft-max normalization function. (3) Weighting the scale-adjusted input features to obtain the fused output features; The CIOU loss function improves the regression speed of predicted bounding boxes by calculating the overlap area between two bounding boxes, the distance between their center points, and their aspect ratio. ； In the formula, and These represent the center points of the predicted bounding box and the target bounding box, respectively. The Euclidean distance between the center points of the two frames. This represents the slant distance of the smallest rectangle between the predicted bounding box and the target bounding box. To limit the weight parameters of the aspect ratio of the predicted bounding box, This parameter measures the consistency between the aspect ratio of the predicted bounding box and the actual bounding box.

6. A cross-scale infrared pedestrian detection system applying the cross-scale infrared pedestrian detection method as described in any one of claims 1 to 5, characterized in that, The multi-scale infrared pedestrian detection system includes: The brightness feature extraction module is used to construct the brightness perception module. It obtains the reconstructed background image and brightness feature image by combining the quadtree-Bessel interpolation operator and the guided filter. The dynamic feature optimization module is used to design the EG-Chimp optimization model and iteratively optimize the constructed objective function to obtain the dynamic feature optimization image; A cross-scale feature fusion module is used to introduce the BiFPN feature pyramid structure under the YOLOv5 algorithm framework. The cross-scale feature fusion module is designed to learn the weight parameters between features of different scales. The infrared pedestrian detection module is used to optimize the loss function. It introduces the CIOU regression loss function and uses learnable weight parameters to assign weights to multi-scale features to achieve infrared pedestrian detection.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the cross-scale infrared pedestrian detection method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the cross-scale infrared pedestrian detection method as described in any one of claims 1 to 5.

9. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the cross-scale infrared pedestrian detection system as described in claim 6.

Images