A hierarchical dependency and position-aware thermal infrared target tracking method

Abstract

This invention belongs to the field of infrared image processing technology, specifically relating to a thermal infrared target tracking method based on hierarchical dependency and position awareness. The method first acquires a sequence of thermal infrared imaging video frames, extracts template and search images, and inputs them into a Transformer encoder to obtain a joint feature sequence. A hierarchical dependency-aware fusion module deeply fuses the outputs of each encoding layer with the joint features to construct a global hierarchical dependency feature. A position-aware gating module enhances position alignment and fine-grained feature representation. Finally, the joint features, dependency-aware features, and gating features are fused, and the target detection result is output through a detection head. This invention specifically addresses the shortcomings of existing Transformer trackers in thermal infrared modalities, such as position information dilution and insufficient hierarchical dependency modeling. It significantly improves tracking robustness and accuracy in complex scenarios such as interference from similar targets, thermal crossover, and partial occlusion, and can be widely adapted to the needs of security monitoring, autonomous driving, military reconnaissance, and other fields.

Description

Technical Field

[0001] This invention belongs to the field of infrared image processing technology, specifically relating to a thermal infrared target tracking method based on hierarchical dependence and position awareness. Background Technology

[0002] Thermal infrared target tracking technology, with its core advantages of not relying on visible light and being able to penetrate adverse weather conditions such as fog, rain, and snow, can accurately identify and lock onto targets at night, in low-visibility, and complex obstructed environments. It has become a core supporting technology in key areas such as nighttime security monitoring, all-weather obstacle avoidance in autonomous driving, low-visibility environment detection in military reconnaissance, and early warning of forest fires. In practical applications, the accuracy and robustness of thermal infrared tracking directly determine the system's emergency response speed, decision-making accuracy, and operational safety. For example, effective tracking of pedestrians or obstacles at night in autonomous driving scenarios can prevent fatal collisions, and continuous locking onto concealed targets in military reconnaissance ensures the effectiveness of mission deployment. Therefore, performance optimization of this technology has significant engineering application value and practical significance.

[0003] Throughout the development of thermal infrared target tracking technology, methods at different stages have all exhibited significant limitations. Early solutions primarily relied on manually designed features (such as histograms of oriented gradients and gray-level co-occurrence matrices) combined with correlation filtering algorithms. Matching models were constructed by manually defining surface information such as target texture and grayscale. However, these methods are limited by the representational capabilities of manually designed features. Faced with the inherent characteristics of thermal infrared images—low contrast, lack of texture detail, and susceptibility to environmental interference—they exhibit poor generalization ability and struggle to handle complex scenarios such as target pose changes and interference from similar targets. With the rise of deep learning technology, trackers based on convolutional neural networks (CNNs) and Siamese network frameworks have gradually become mainstream. These networks automatically learn deep target features, significantly improving tracking performance. However, the local receptive field characteristics of CNNs make it difficult to capture long-range dependencies, and tracking drift still easily occurs in scenarios such as thermal crossover (different targets have similar thermal radiation characteristics) and partial occlusion. In recent years, Transformer-based visual trackers have stood out for their long-range modeling capabilities with their self-attention mechanism. They are divided into two categories: "relational modeling after feature extraction" and "template-search feature early fusion single-stream modeling". They have achieved performance breakthroughs in the visible light (RGB) field. However, these methods are designed for the rich texture and color information of RGB images. When directly transferred to the thermal infrared mode, the adaptability problem becomes prominent.

[0004] Existing technologies face two major contradictions when applied to thermal infrared target tracking: First, the multi-layer stacking and linear projection process of the Transformer architecture continuously dilutes the initial position encoding information, while thermal infrared images themselves have weak spatial details, further weakening the model's ability to perceive the target's spatial location. At the same time, the self-attention mechanism tends to focus on high-response semantic regions, ignoring fine-grained spatial features such as edges and contours. In scenarios with similar target interference or thermal crossover, it is prone to mismatches due to blurred position information and missing local details. Second, existing trackers lack explicit modeling of the global dependencies between different levels of the feature pyramid (low-level texture, mid-level structure, and high-level semantics). They either over-rely on high-level semantic features, leading to loss of details, or simply stitch together data during multi-scale fusion, ignoring hierarchical correlations. However, thermal infrared images have low signal-to-noise ratios for low-level features and insufficient discriminative power for high-level features. This deficiency makes it impossible for the model to integrate effective information from each level to construct a strong discriminative feature representation, making it prone to tracking failures in scenarios with partial target occlusion or changes in ambient temperature. Therefore, designing a tracking method that can effectively model hierarchical global dependencies and enhance fine-grained position awareness, based on the inherent characteristics of thermal infrared modes, is key to solving the problem of insufficient robustness and accuracy in thermal infrared target tracking under complex scenarios. Summary of the Invention

[0005] To address the problems existing in the background art, one aspect of the present invention provides a thermal infrared target tracking method based on hierarchical dependence and position awareness, comprising:

[0006] S1: Obtain the thermal infrared imaging video frame sequence, and extract the template image and search image from the thermal infrared imaging video frame sequence;

[0007] S2: Input the template image and the search image into the Transformer encoder to obtain the joint feature sequence;

[0008] S3: The output feature sequence of each Transformer encoder layer and the joint feature sequence are deeply fused through the hierarchical dependency-aware fusion module to obtain dependency-aware features;

[0009] S4: Input the joint feature sequence into the position-aware gating module to extract the gating features that have been aligned and enhanced with fine granularity;

[0010] S5: The joint feature sequence, dependent perceptual features and gating features are fused to obtain fused features. The fused features are then input into the detection head to obtain the target detection results in the search image.

[0011] Another aspect of the present invention provides a thermal infrared target tracking system based on hierarchical dependency and position awareness, the system including a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned thermal infrared target tracking method based on hierarchical dependency and position awareness.

[0012] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for tracking thermal infrared targets based on hierarchical dependency and position awareness.

[0013] The present invention has at least the following beneficial effects

[0014] This invention addresses the core problems in thermal infrared target tracking, such as the dilution of position encoding, insufficient fine-grained spatial feature representation, and the lack of global dependency modeling between different levels of the feature pyramid (low-level texture, mid-level structure, and high-level semantics) in Transformer architecture applications. It addresses these issues by designing a hierarchical dependency-aware fusion module to explicitly construct global hierarchical associations across encoding layers, thus compensating for the fragmentation of hierarchical information. A position-aware gating module enhances position alignment and fine-grained feature enhancement, alleviating the problems of ambiguous position information and missing local details. Furthermore, through adaptive fusion of joint feature sequences, dependency-aware features, and gating features, it effectively integrates multi-dimensional information, significantly improving the model's feature discrimination capability in complex scenarios such as similar target interference, thermal crossover, and partial occlusion. Ultimately, this achieves a dual improvement in the robustness and accuracy of thermal infrared target tracking, better adapting to the needs of practical applications such as security monitoring, autonomous driving, and military reconnaissance. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0016] Figure 2 This is a schematic diagram of the model structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 and Figure 2 One aspect of the present invention provides a thermal infrared target tracking method based on hierarchical dependence and position awareness, comprising:

[0019] S1: Obtain the thermal infrared imaging video frame sequence, and extract the template image and search image from the thermal infrared imaging video frame sequence;

[0020] Preferably, the extraction of template image and search image includes: selecting a region containing the target from the first video frame in the thermal infrared imaging video frame sequence to construct a template image, and selecting a search image from the remaining video frames in the thermal infrared imaging video frame sequence other than the first frame.

[0021] In this embodiment, the publicly available thermal infrared target tracking dataset LSOTB-TIR is used for model training. The model is built using the PyTorch deep learning framework. The model training consists of 60 epochs, with a batch size of 64, using the AdamW optimizer, and a weight decay of 1×10⁻. 4 The learning rate of the backbone network is initialized to 4 × 10⁻ 4 The input image pair includes a template image and a search image, whose dimensions are preprocessed to 3×128×128 and 3×256×256, respectively. During the testing phase, video can be acquired using an infrared imaging device, and video frames can be extracted using an image processing library (such as OpenCV). These frames are then cropped and normalized according to the aforementioned dimensions to construct the input sequence. By detecting the location and category information of the target in each search image, target tracking can be achieved.

[0022] In this embodiment, high-quality thermal infrared imaging video frame sequences are first obtained through two methods. During the model training and validation phase, publicly available thermal infrared target tracking datasets (such as LSOTB-TIR, NUDT-SIRST, PTB-TIR, etc.) are used. These datasets cover various complex scenarios such as nighttime, fog and snow, low visibility, and partial occlusion, and include different types of tracking objects such as pedestrians, vehicles, and small targets. They also provide accurately annotated target bounding boxes, which can be directly used for subsequent template and search image extraction and model performance validation. During the actual deployment and application phase, video streams of the target scene are acquired in real time through infrared imaging devices (such as infrared thermal imagers, vehicle-mounted infrared cameras, security infrared monitoring equipment, etc.). Continuous thermal infrared video frame sequences are extracted using image processing libraries such as OpenCV, and format normalization, slight denoising, and frame synchronization processing are performed to ensure the continuity and effectiveness of the frame sequences. Image extraction is then performed: From the first frame of the thermal infrared imaging video frame sequence, based on the bounding boxes labeled in the dataset or the target area automatically detected in the actual acquisition scene, an image patch containing the complete target is cropped as a template image (preprocessed to a fixed size, such as 3×128×128) to ensure that the target has a reasonable proportion in the template to facilitate feature modeling; From the other consecutive video frames in the sequence except the first frame, an image patch containing the area where the target may appear is cropped according to the preset search range to serve as a search image (preprocessed to a fixed size, such as 3×256×256) to ensure that the potential motion trajectory of the target is covered, while controlling the image size to balance computational efficiency and tracking accuracy.

[0023] S2: Input the template image and the search image into the Transformer encoder to obtain the joint feature sequence;

[0024] Preferably, the step of inputting the template image and search image into the Transformer encoder for encoding includes: cutting the template image and search image into multiple fixed-size template image blocks and search image blocks respectively; mapping each template image block and search image block to a high-dimensional space through two linear projection matrices respectively; and adding learnable positional encodings to construct template embedding representation sequences and search embedding representation sequences; concatenating the template embedding representation sequences and search embedding representation sequences in the spatial dimension to construct a joint token sequence; and inputting the joint token sequence into the Transformer encoder for encoding to obtain a joint feature sequence.

[0025] In this embodiment, the extracted template image and search image are first preprocessed with feature embedding. They are then cut into fixed-size 16×16 image blocks and flattened. For each template image block and search image block, two independent learnable linear projection matrices are used. and Mapping to a 768-dimensional high-dimensional feature space, and adding learnable positional encodings to the embedding features of the two types of image patches respectively. and This preserves the spatial location information of the image, and then constructs a template embedding representation sequence with dimensions of 64×768. And a search embedding representation sequence of 256×768 Subsequently, the template embedding representation sequence and the search embedding representation sequence are concatenated along the spatial dimension to form a joint token sequence with a dimension of (64+256)×768. Finally, the joint token sequence is input into a backbone network consisting of N=12 stacked standard Transformer encoders. Each encoder layer sequentially performs layer normalization, multi-head self-attention calculation, and feedforward network operations, and residual connections ensure the integrity of feature propagation. After layer-by-layer deep encoding and relation modeling by the 12 encoder layers, the final output is a joint feature sequence that has undergone long-range dependency mining. .

[0026] S3: The output feature sequence of each Transformer encoder layer and the joint feature sequence are deeply fused through the hierarchical dependency-aware fusion module to obtain dependency-aware features;

[0027] Preferably, step S3 includes:

[0028] S31: The first... in the Transformer encoder The output feature sequence of each Transformer encoding layer is represented as follows: ,Will Max pooling along the spatial dimension yields hierarchical features. ;

[0029] S32: Concatenate all hierarchical features to obtain a hierarchical sequence. ,in, This indicates the number of Transformer coding layers in the Transformer encoder;

[0030] S33: Hierarchical sequence Input dependency establishment layer extracts global hierarchical dependency features The dependency-building layer consists of three cascaded dilated convolutional layers, each of which includes a DWC convolution, a PReLU activation function, and a Dropout layer cascaded in sequence.

[0031] S34: Incorporate global hierarchical dependency features Concatenating the concatenated features with the joint feature sequence yields the concatenated features. ; splicing features Input a Transformer encoding layer and perform feature fusion to obtain features ;

[0032] S35: From the characteristics Extract the features of the corresponding parts of the joint feature sequence as dependency-aware features. .

[0033] In this embodiment, the input is first defined as the output feature sequence of all N Transformer coding layers in the Transformer encoder. The output feature sequence of each Transformer encoding layer is represented as follows: With joint feature sequences First, process the output feature sequence of each coding layer. Perform max pooling along the spatial dimension (token sequence dimension) to extract the most discriminative channel features from each layer, thus obtaining the corresponding hierarchical features. Then, all N hierarchical features are concatenated in the order of the coding layers to construct a complete hierarchical sequence. Next, the hierarchical sequence H is input into a dependency building layer consisting of three cascaded dilated convolutions. Each dilated convolution is cascaded with a depthwise separable convolution (DWC), a PReLU activation function, and a Dropout layer. The kernel size of the three dilated convolutions is set to 3, the stride to 1, and the dilation rates to 1, 2, and 3, respectively, with padding values ​​of 2, 4, and 6 to ensure that the output length remains constant. This structure explicitly mines global hierarchical dependencies across all coding layers and outputs global hierarchical dependency features. Finally, the global hierarchical dependency feature D and the joint feature sequence are combined. By concatenating along the feature dimensions, we obtain the concatenated features. Then splicing features An additional Transformer encoding layer is input, and a deep semantic fusion of dependency features and spatial features is achieved through a self-attention mechanism. The final output is a feature that combines hierarchical dependency information and spatial feature representation. Extracting partial features corresponding to the joint feature sequence from this feature as dependency-aware features. .

[0034] S4: Input the joint feature sequence into the position-aware gating module to extract the gating features that have been aligned and enhanced with fine granularity;

[0035] Preferably, step S4 includes:

[0036] S41: Extract features from the corresponding part of the template image from the joint feature sequence. Features corresponding to the search image ;

[0037] S42: Features and characteristics Features are obtained by mapping through two linear layers respectively. and characteristics ;

[0038] S43: Features Features are obtained sequentially through a DWC convolution, a GELU activation function, and a nearest neighbor upsampling layer. ;

[0039] S44: Features and characteristics Features are obtained by performing Hadamard product operations. , will feature Features are obtained by processing through a linear layer and a Dropout layer in sequence. ; Features and characteristics Gating features are obtained by splicing. .

[0040] In this embodiment, we first start from the joint feature sequence Features of the corresponding part of the template image are extracted by sequence dimension. Features corresponding to the search image This ensures that the two types of features retain the core features of the template target and the scene features of the search region, respectively; then the features are... and characteristics Two independent learnable linear layers are input separately, and then uniformly mapped to the same intermediate feature dimension through a linear transformation to obtain the features. and characteristics This lays the foundation for subsequent feature interactions; subsequently, the features will be... The data is sequentially fed into a depthwise separable convolutional (DWC) layer, a GELU activation function, and a nearest neighbor upsampling layer. The DWC convolution reduces computation while preserving the spatial correlation of features, the GELU activation function performs non-linear feature transformation, and the nearest neighbor upsampling adapts the feature size to match the feature size. Consistency is achieved, ultimately resulting in enhanced template features. Then the features With features Performing the Hadamard product operation strengthens the fine-grained correlation between the two types of features at corresponding positions through element-wise multiplication, achieving position alignment and feature enhancement, and obtaining the feature... ;Then the features The features are obtained by sequentially passing them through a linear layer for dimensionality adjustment and a Dropout layer (to suppress overfitting). Finally, the features and characteristics The features are concatenated along the feature dimension, and the original core features are combined with the fine-grained features enhanced by position. The final output is a gated feature that has been aligned by position and enhanced by fine-grained position.

[0041] S5: The joint feature sequence, dependent perceptual features and gating features are fused to obtain fused features. The fused features are then input into the detection head to obtain the target detection results in the search image.

[0042] Preferably, the fusion of joint feature sequences, dependency-aware features, and gating features to obtain fused features includes:

[0043]

[0044] in, and Represents the learnable weight parameters; Indicates fusion characteristics; It represents the Hadamardi (or Hadama) stack; Indicates dependency-aware features; This indicates a gating feature.

[0045] In this embodiment, the adaptive fusion formula is first used. Achieving deep integration of three types of features, among which, and These are learnable weight parameters with the same dimension as the features, which can dynamically adjust the contribution ratio of different features during model training. The Hadamard product, representing element-wise multiplication, strengthens local correlations between features and highlights effective information. This fusion process fully leverages the complementary advantages of the three types of features, preserving the basic long-range dependency information of the joint features while incorporating hierarchical global dependencies, precise location awareness, and fine-grained detail features. This effectively compensates for the expression deficiencies of single features in complex scenarios, ultimately generating a fused feature with strong discriminative power, location accuracy, and hierarchical correlation. ;Then the fusion features will be The input is a center-based detection head network, which performs classification and regression tasks in parallel. The classification branch outputs the probability of target presence to distinguish between background and target, while the regression branch outputs the bounding box coordinates of the target in the search image (such as center coordinates, width and height). Through the collaborative decision-making of the two, the target position in the search image is accurately located, completing single-frame target detection. By combining the detection results of continuous video frames, the target motion trajectory can be constructed, achieving stable and accurate tracking of thermal infrared targets. In this embodiment, the detection network can adopt the CenterNet architecture, FCOS (FullyConvolutional One-Stage Object Detection) architecture, and YOLOv8 decoupled detection head architecture, etc. Before inputting the fused features into the detection head network, a reshape operation is required to reshape the fused features into a two-dimensional spatial feature map.

[0046] Another aspect of the present invention provides a thermal infrared target tracking system based on hierarchical dependency and position awareness, the system including a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned thermal infrared target tracking method based on hierarchical dependency and position awareness.

[0047] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for tracking thermal infrared targets based on hierarchical dependency and position awareness.

[0048] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0049] In summary, this invention addresses the core issues of diluted position encoding, insufficient fine-grained spatial feature representation, and lack of global dependency modeling between different levels of the feature pyramid (low-level texture, mid-level structure, and high-level semantics) in thermal infrared target tracking using the Transformer architecture. It addresses these issues by designing a hierarchical dependency-aware fusion module to explicitly construct global hierarchical associations across encoding layers, thus mitigating the fragmentation of hierarchical information. Furthermore, a position-aware gating module enhances position alignment and fine-grained feature enhancement, alleviating the problems of ambiguous position information and missing local details. Finally, through adaptive fusion of joint feature sequences, dependency-aware features, and gating features, it effectively integrates multi-dimensional information, significantly improving the model's feature discrimination capability in complex scenarios such as similar target interference, thermal crossover, and partial occlusion. Ultimately, this invention achieves a dual improvement in the robustness and accuracy of thermal infrared target tracking, better adapting to the needs of practical applications such as security monitoring, autonomous driving, and military reconnaissance.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A thermal infrared target tracking method based on hierarchical dependency and position awareness, characterized in that, include: S1: Obtain the thermal infrared imaging video frame sequence, and extract the template image and search image from the thermal infrared imaging video frame sequence; S2: Input the template image and the search image into the Transformer encoder to obtain the joint feature sequence; S3: The output feature sequence of each Transformer encoder layer and the joint feature sequence are deeply fused through the hierarchical dependency-aware fusion module to obtain dependency-aware features; S4: Input the joint feature sequence into the position-aware gating module to extract the gating features that have been aligned and enhanced with fine granularity; S5: The joint feature sequence, dependent perceptual features and gating features are fused to obtain fused features. The fused features are then input into the detection head to obtain the target detection results in the search image.

2. The thermal infrared target tracking method based on hierarchical dependency and position awareness according to claim 1, characterized in that, The extraction of template images and search images includes: selecting a region containing the target from the first video frame in the thermal infrared imaging video frame sequence to construct a template image, and selecting search images from the remaining video frames in the thermal infrared imaging video frame sequence excluding the first frame.

3. The thermal infrared target tracking method based on hierarchical dependency and position awareness according to claim 1, characterized in that, The step of inputting the template image and search image into the Transformer encoder for encoding includes: cutting the template image and search image into multiple fixed-size template image blocks and search image blocks respectively; mapping each template image block and search image block to a high-dimensional space through two linear projection matrices respectively; and adding learnable positional encodings to construct template embedding representation sequences and search embedding representation sequences; concatenating the template embedding representation sequences and search embedding representation sequences in the spatial dimension to construct a joint token sequence; and inputting the joint token sequence into the Transformer encoder for encoding to obtain a joint feature sequence.

4. The thermal infrared target tracking method based on hierarchical dependence and position awareness according to claim 1, characterized in that, Step S3 includes: S31: The first... in the Transformer encoder The output feature sequence of each Transformer encoding layer is represented as follows: ,Will Max pooling along the spatial dimension yields hierarchical features. ; S32: Concatenate all hierarchical features to obtain a hierarchical sequence. ,in, This indicates the number of Transformer coding layers in the Transformer encoder; S33: Hierarchical sequence Input dependency establishment layer extracts global hierarchical dependency features The dependency-building layer consists of three cascaded dilated convolutional layers, each of which includes a DWC convolution, a PReLU activation function, and a Dropout layer cascaded in sequence. S34: Incorporate global hierarchical dependency features Concatenating the concatenated features with the joint feature sequence yields the concatenated features. ; splicing features Input a Transformer encoding layer and perform feature fusion to obtain features ; S35: From the characteristics Extract the features of the corresponding parts of the joint feature sequence as dependency-aware features. .

5. The thermal infrared target tracking method based on hierarchical dependency and position awareness according to claim 1, characterized in that, Step S4 includes: S41: Extract features from the corresponding part of the template image from the joint feature sequence. Features corresponding to the search image ; S42: Features and characteristics Features are obtained by mapping through two linear layers respectively. and characteristics ; S43: Features Features are obtained sequentially through a DWC convolution, a GELU activation function, and a nearest neighbor upsampling layer. ; S44: Features and characteristics Features are obtained by performing Hadamard product operations. , will feature Features are obtained by processing through a linear layer and a Dropout layer in sequence. ; Features and characteristics Gating features are obtained by splicing. .

6. The thermal infrared target tracking method based on hierarchical dependency and position awareness according to claim 1, characterized in that, The process of fusing joint feature sequences, dependency-aware features, and gated features to obtain fused features includes: in, and Represents the learnable weight parameters; Indicates fusion characteristics; It represents the Hadamardi (or Hadama) stack; Indicates dependency-aware features; Indicates gating characteristics; This represents a joint feature sequence.

7. A thermal infrared target tracking system based on hierarchical dependence and position awareness, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute a thermal infrared target tracking method based on hierarchical dependency and position awareness as described in any one of claims 1 to 6.

8. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements a thermal infrared target tracking method based on hierarchical dependency and position awareness as described in any one of claims 1 to 6.

Images