Rapid detection algorithm for elevated passenger transport devices in mines

Abstract

This invention relates to a fast detection algorithm for aerial vehicle passenger transport devices in the field of electronic information technology. Based on the YOLOv7 framework, a lightweight network model is obtained by reconstructing the network. A CA attention mechanism is added to FasterNet to obtain FasterNet-CA. An adaptive spatial ASFF structure is adopted to enhance the information interaction between the upper and lower layers of the lightweight network model, resulting in the MOOD algorithm model. The MOOD algorithm model is trained, validated, and tested using a dataset, ultimately yielding a fast detection algorithm for aerial vehicle passenger transport devices in mines. The MOOD algorithm model of this invention is only 47 MB ​​in size, with only 23.3 MB of parameters, achieving an mAP of 96.6% and a detection speed of 121 frames per second (FPS). It meets the accuracy and speed requirements for detecting aerial vehicle passenger transport devices in mines and can be used for reliable detection of behaviors such as not wearing a safety helmet, carrying prohibited items, and passenger status.

Description

Technical Field

[0001] This invention belongs to the field of electronic information technology, specifically relating to a rapid detection algorithm for aerial passenger transport devices in mines. Background Technology

[0002] In recent years, with the rapid development of science and technology and artificial intelligence, and the continuous advancement of "new infrastructure" in mines, the operation of aerial work platforms in mines has gradually evolved towards intelligence, safety, efficiency, and energy conservation. Aerial work platforms in mines, affectionately known as "monkey cars" by workers, are mainly used to assist in transporting workers in underground mines, aiming to shorten the travel time for workers to go up and down the mine and reduce unnecessary physical exertion. Currently, most aerial work platforms in coal mines still rely on traditional sensors and drive devices for operation. However, intelligentization and informatization are the mainstream development trends in modern mine production, and digitizing, intelligentizing, and informatizing them is an inevitable measure to ensure their safety, efficiency, reliability, and green energy conservation.

[0003] As is well known, the coal industry is a high-risk industry, and numerous cameras are typically installed in various tunnels, but most of the image resources have not been effectively utilized. Furthermore, analysis of image data collected during the operation of aerial work platforms in mines revealed that the images exhibit characteristics such as uneven brightness, complex environments, and significant variations in pedestrian scale. Therefore, it is necessary for those skilled in the art to propose a fast, accurate, and highly reliable detection algorithm to reliably detect behaviors such as not wearing safety helmets and carrying prohibited items, as well as the passenger status. This is of great significance for promoting the integration of intelligent technology with the coal industry, fostering high-quality development of the coal industry, and improving worker safety and coal mine production levels.

[0004] With the continuous improvement of deep learning technology, deep learning-based detection algorithms have become the mainstream in existing detection technologies. For example, a method using feature extraction and CBAM attention mechanism based on YOLOv3 can detect pedestrians in coal mines, outperforming classical detection methods. An improved YOLOv4 network based on thermal infrared images is used for personnel detection, employing K-means++ clustering and transfer learning strategies to enhance the model's robustness and generalization ability. An improved mine pedestrian detection algorithm based on YOLOv4-Tiny achieves global and local feature extraction from image information, with an improved mAP of only 91.98%. Furthermore, based on YOLOv5, Vision Transformer and Merge... While Non-Maximum Suppression and Meta-AconC achieve higher detection accuracy, their detection speed decreases. Detectors based on Parallel Feature Transfer Network (PftNet) are used to solve the problem of pedestrian detection in underground mines. PftNet is more suitable for objects of different sizes and aspect ratios, further improving accuracy. With the introduction of Transformer and its application in the field of vision, it has achieved good results in solving large targets, but it has disadvantages such as large number of parameters, slow computing speed, and many limitations in handling small target tasks, making it difficult to apply to the detection of aerial passenger transport devices in mines.

[0005] The inventors discovered through research that the image background of the work platform of the aerial work platform in mines is fixed but complex, mainly divided into periods of no personnel, periods with a small number of people needing to ride, and peak periods for workers going up and down the mine shaft. During peak periods, a large number of workers gather on the work platform, resulting in high density in many images. Although many high-quality object detection algorithms have emerged in the field of vision, the aerial work platform in mines is located in a complex and harsh environment, making it impossible to collect ideal image information. Furthermore, the detection task of the aerial work platform in mines requires high accuracy and speed; directly applying object detection algorithms to its detection is unlikely to achieve satisfactory results. Existing object detection algorithms still have significant limitations for the task of accurately detecting aerial work platforms in mines. Therefore, it is necessary to design a highly accurate and fast algorithm model to detect pedestrian movement on the work platform of the aerial work platform in real time. Summary of the Invention

[0006] Given the complex and harsh environment of underground mines, where images suffer from poor lighting, complex backgrounds, blurred pedestrians, and multi-scale pedestrians, current detection models generally suffer from high computational loads and large parameter counts, resulting in slow detection speeds. The detection performance of current mainstream algorithms is insufficient for detecting aerial passenger transport devices. This invention addresses the detection needs of aerial passenger transport devices in mines by redesigning a lightweight, fast, and highly accurate detection model for reliably detecting behaviors such as not wearing safety helmets and carrying prohibited items, as well as passenger status.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] A rapid detection algorithm for elevated personnel-riding devices in mines includes the following steps:

[0009] S1. Based on the YOLOv7 framework, the network is reconstructed to obtain a lightweight network model: the backbone network in the original YOLOv7 is improved to FasterNet, the cross-connection GSCSP module is used to replace the SPPCSPC module in the YOLOv7 head detection head, and the PConv module is used to improve the ELAN module in the YOLOv7 head detection head. The network is reconstructed to reduce the amount of computation and parameters, resulting in a lightweight network model to improve the processing speed of the model.

[0010] S2. Add the CA attention mechanism to FasterNet to obtain FasterNet-CA, which enhances feature extraction, makes up for the loss of feature extraction information caused by the lightweight network model, and improves detection accuracy.

[0011] S3. Adopting an adaptive spatial ASFF structure enhances the information interaction between the upper and lower layers of the lightweight network model, resulting in the MOOD algorithm model; this improves the accuracy of the predicted bounding boxes and enhances the model's detection performance.

[0012] S4. Collect images in the actual production environment of aerial passenger transport devices in mines, create a dataset, and then use the dataset to train, verify, and test the MOOD algorithm model, ultimately obtaining a fast detection algorithm for aerial passenger transport devices in mines.

[0013] Preferably, in step S1, while ensuring accuracy, this application uses FasterNet as the backbone network of the lightweight network model. FasterNet is a fast and lightweight detection network proposed by Jierun Chen et al. in 2023. The FasterNet is a structure composed of Embedding, FasterNet Block, Merging, and Classifier, and includes four stages. Before stage one, it is processed by an Embedding layer (a Conv 4×4 with a stride of 2), and before other stages, it is processed by a Merging layer (a Conv 2×2 with a stride of 2) for spatial downsampling and channel expansion. Each stage is composed of stacked FasterNet Blocks.

[0014] Preferably, each FasterNet Block consists of one PConv and two subsequent PWConv 1×1 or Conv 1×1, and the three form an inverted residual architecture with more channels in the intermediate layer and a Shortcut connection is placed to reuse input features.

[0015] Compared to YOLOv7's backbone, using FasterNet as the backbone network reduces the number of parameters by 32.5% and the computational cost by 60.1%.

[0016] Preferably, in step S1, to further reduce the number of model parameters, the cross-connection GSCSP module is used instead of the SPPCSPC module in the YOLOv7 Head detection head. The SPPCSPC module consists of ordinary convolutional modules and pooling operations, with a parameter size of 7.6 MB. The GSCSP module is a cross-level partial network designed based on the GSConv module, consisting of ordinary convolutional blocks and GS bottleneck modules, with a parameter size of only 18 MB. Considering that the environment of the elevated passenger vehicle platform can easily interfere with the detection target, leading to false detections and misdetections during network transmission, and that spatial information is gradually transmitted to channels during image backbone transformation, and that each feature map compression and expansion results in some information loss, the GSConv module preserves as much hidden connection as possible between channels, retaining more information. Using the GSConv module to process stitched feature maps is appropriate, reducing redundant and repetitive information, eliminating the need for compression, and improving the performance of the attention module. Therefore, using the GSCSP module can reduce the computational cost and network structure complexity of the model while maintaining sufficient accuracy.

[0017] Preferably, in step S1, the ELAN module improves detection performance due to its increased network depth, but this also increases the number of model parameters and computational cost. To further improve the detection performance of aerial work platforms in mines, the ELAN structure is redesigned using the PConv module to obtain the ELAN-PConv module, or ELAN-P module for short, which reduces model complexity while maintaining accuracy. Since feature maps exhibit high similarity and redundancy across different channels, the PConv module is introduced to reduce computational redundancy and memory access. The PConv module only needs to apply the regular Conv module for spatial feature extraction on a portion of the input channels, keeping the remaining channels unchanged. For continuous or regular memory accesses, the first or last consecutive c... p The channels are treated as representative of the entire feature map for computation. Without loss of generality, the input and output feature maps are assumed to have the same number of channels. Therefore, the PConv module replaces the 3×3 ordinary convolutional module in the ELAN module, thus reducing the number of parameters. The computational compression of the PConv module is shown in the following equation:

[0018] (1)

[0019] In equation (1), h and w are the width and height of the feature map, k is the size of the convolution kernel, cp is the number of channels processed by the PConv module, and c is the number of channels processed by the Conv module. The computational complexity of the PConv module can be reduced as shown in the following formula:

[0020] (2)

[0021] The compression of memory access in the PConv module is as follows:

[0022] (3)

[0023] In equation (3), h and w are the width and height of the feature map, k is the size of the convolution kernel, cp is the number of channels used in the PConv module's convolution, and c is the number of channels used in the Conv module's convolution. (The last part is a partial sentence and doesn't need a direct translation.) The compression of memory access in the PConv module is as follows:

[0024] (4)

[0025] Preferably, in step S2, the FasterNet-CA is obtained by adding a CA attention mechanism to the first and third layers of FasterNet, and is inserted after the second Conv module in the FasterNet Block to enhance the ability to extract features and improve detection accuracy. As the network deepens, changes in positional information are often ignored during transmission. However, the CA attention mechanism not only acquires channel information but also considers orientation-related positional information, which helps the model to better locate and identify targets. Furthermore, it forms a pair of orientation-aware and position-sensitive feature maps with almost no additional computation, which can be complementaryly applied to the input feature maps to enhance the representation of the target of interest.

[0026] Preferably, in step S3, since most multi-scale feature extraction uses a one-way information transmission method, it affects the fusion effect between different levels. This method suffers from feature information loss and degradation, which affects the final detection effect and the position of the predicted box. Therefore, an adaptive spatial ASFF structure is used for feature fusion at three levels. The adaptive spatial ASFF structure supports direct interaction between different levels. While retaining one-way information transmission, it fuses two adjacent low-level features into high-level features, thereby avoiding feature information loss and degradation caused by one-way information transmission, as well as the problem of large semantic gaps between different levels. This retains more useful information, resulting in better detection and more accurate localization. Using the adaptive spatial ASFF structure to assign different spatial weights to features at different levels enhances the importance of key levels and mitigates the influence of contradictory information from different targets, as shown in the following formula:

[0027] (5)

[0028] In equation (5), X, Y, and Z are features from Level 1, Level 2, and Level 3, respectively. The weight parameters α, β, and γ are multiplied and added together with the features from different levels to obtain the fused feature adaptive space ASFF structure.

[0029] Preferably, in step S4, the images of workers riding the aerial passenger transport device are processed using the VOC dataset format to create a dataset, which includes a training set, a validation set, and a test set.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] (1) In view of the characteristics of the images of aerial passenger transport devices in mines, this application proposes a dataset of aerial passenger transport devices in mines. Using part of the structure of YOLOv7 Head, the network model based on FasterNet, GSCSP and ELAN-P is redesigned to reduce the computational amount and number of parameters of the network model, so as to improve the processing speed of the model.

[0032] (2) In order to compensate for the loss of feature extraction information caused by lightweight networks, this application adds a CA attention mechanism to FasterNet to enhance feature extraction and improve detection accuracy;

[0033] (3) Due to the harsh environment of the mine, poor lighting, complex background, blurred pedestrians and multi-scale pedestrians, this application adopts an adaptive spatial ASFF structure to strengthen the information interaction between the upper and lower layers, improve the accuracy of the prediction box, and make the model detection effect better.

[0034] In summary, compared with other mainstream algorithms, the MOOD algorithm model is only 47 MB ​​in size, has only 23.3 MB of parameters, achieves an mAP of 96.6%, and has a detection speed of 121 frames per second (FPS). It can meet the detection accuracy and speed requirements of aerial personnel transport devices in mines and can be used to reliably detect behaviors such as not wearing a safety helmet, carrying prohibited items, and riding status. Attached Figure Description

[0035] Figure 1 This is a network structure diagram of the MOOD algorithm model of the present invention in the embodiment;

[0036] Figure 2 This is a schematic diagram of the FasterNet structure of the present invention in an embodiment;

[0037] Figure 3 This is a schematic diagram of the GS bottleneck module of the present invention in an embodiment;

[0038] Figure 4 This is a schematic diagram of the structure of the GSCSP module of the present invention in an embodiment;

[0039] Figure 5 This is a schematic diagram of the GSConv module of the present invention in an embodiment;

[0040] Figure 6 This is a schematic diagram of the structure of the ELAN-P module of the present invention in an embodiment;

[0041] Figure 7 This is a schematic diagram of the PConv module of the present invention in an embodiment;

[0042] Figure 8This is a schematic diagram of the FasterNet Block structure of the present invention in an embodiment;

[0043] Figure 9 This is a schematic diagram of the FasterNet-CA structure of the present invention in an embodiment;

[0044] Figure 10 This is a schematic diagram of the adaptive spatial ASFF structure of the present invention in an embodiment;

[0045] Figure 11 This is a diagram illustrating the data collection information of the dataset presented in this embodiment.

[0046] Figure 12 This is a performance comparison chart of the CA attention mechanism of the present invention in each layer of FasterNet in the embodiments;

[0047] Figure 13 This is a comparison of the parameter quantities of the CA attention mechanism of the present invention in each layer of FasterNet in the embodiment;

[0048] Figure 14 The image shows the prediction box result and a partial magnified view of the present invention in the embodiment.

[0049] Figure 15 The images show the detection results and magnified views of each model for hat, state1, and state2 in the examples.

[0050] Figure 16 The examples show the detection results and magnified views of each model for hat and state1.

[0051] Figure 17 The images show the detection results and magnified views of the head for each model in the embodiments.

[0052] Figure 18 The images show the detection results and magnified views of illegal carrying for each model in the examples. Detailed Implementation

[0053] The technical solutions 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 protection scope of the present invention.

[0054] Example

[0055] like Figure 1 As shown, the fast detection algorithm for aerial passenger transport devices in mines proposed in this embodiment includes the following steps:

[0056] S1. Based on the YOLOv7 framework, the network is reconstructed to obtain a lightweight network model: the backbone network in the original YOLOv7 is improved to FasterNet, the cross-connection GSCSP module is used to replace the SPPCSPC module in the YOLOv7 head detection head, and the PConv module is used to improve the ELAN module in the YOLOv7 head detection head. The network is reconstructed to reduce the amount of computation and parameters, resulting in a lightweight network model to improve the processing speed of the model.

[0057] S2. Add the CA attention mechanism to FasterNet to obtain FasterNet-CA, which enhances feature extraction, compensates for the loss of feature extraction information caused by the lightweight network model, and improves detection accuracy.

[0058] S3. Adopting an adaptive spatial ASFF structure enhances the information interaction between the upper and lower layers of the lightweight network model, resulting in the MOOD algorithm model; this improves the accuracy of the predicted bounding boxes and enhances the model's detection performance.

[0059] S4. Collect images in the actual production environment of aerial passenger transport devices in mines, create a dataset, and then use the dataset to train, verify, and test the MOOD algorithm model, ultimately obtaining a fast detection algorithm for aerial passenger transport devices in mines.

[0060] In step S1, while ensuring accuracy, this application uses FasterNet as the backbone network of the lightweight network model. FasterNet (its structure is as follows) Figure 2The FasterNet (shown) is a fast and lightweight detection network proposed by Jierun Chen et al. in 2023. The FasterNet structure consists of Embedding, FasterNet Blocks, Merging, and a Classifier, and includes four stages. Before stage one, the network is processed by an Embedding layer (a 4×4 Conv layer with a stride of 2). Before the other stages, a Merging layer (a 2×2 Conv layer with a stride of 2) is used for spatial downsampling and channel expansion. Each stage is composed of stacked FasterNet Blocks. Each FasterNet Block consists of one PConv layer and two subsequent PWConv layers (1×1 or Conv1×1), forming an inverse residual architecture. The intermediate layers have more channels and a Shortcut connection is placed to reuse input features. Compared to the YOLOv7 backbone, using FasterNet as the backbone reduces the number of parameters by 32.5% and the computational cost by 60.1%.

[0061] In step S1, to further reduce the number of model parameters, the cross-connection GSCSP module is used instead of the SPPCSPC module in the YOLOv7 head detection head. The SPPCSPC module consists of ordinary convolutional modules and pooling operations, and its parameter size reaches 7.6 MB. The GSCSP module (its structure is as follows) Figure 4 As shown in the figure, it is in the GSConv module (its structure is as follows) Figure 5 Based on the above, a cross-level partial network is designed, consisting of ordinary convolutional blocks and GS bottleneck modules (the structure of which is shown in the figure). Figure 3 The system consists of components (as shown) and has only 18 MB of parameters. The environment of the elevated passenger platform can easily interfere with the detection target, leading to false positives and incorrect detections during network transmission. Furthermore, during image transformation on the backbone, spatial information is gradually transmitted to the channels, and each compression and expansion of the feature map results in information loss. The GSConv module, however, preserves as much hidden connection as possible between each channel, retaining more information. Using the GSConv module to process stitched feature maps is ideal, reducing redundant and repetitive information, eliminating the need for compression, and improving the performance of the attention module. Therefore, using the GSCSP module can reduce the computational cost and network complexity of the model while maintaining sufficient accuracy.

[0062] In addition, in step S1, the ELAN module improves detection performance due to its increased network depth, but this also increases the number of model parameters and computational load. To further improve the detection performance of aerial work platforms in mines, the ELAN structure is redesigned using the PConv module, resulting in the ELAN-PConv module, or ELAN-P module for short. Its structure is as follows... Figure 6 As shown. This is used to reduce model complexity while maintaining accuracy. Because feature maps have high similarity and redundancy across different channels, the PConv module is introduced to reduce computational redundancy and memory access. The PConv module only applies the regular Conv module for spatial feature extraction on a portion of the input channels, keeping the remaining channels unchanged. For continuous or regular memory access, the first or last consecutive c... p The channels are treated as representative of the entire feature map for computation. Without loss of generality, the input and output feature maps are assumed to have the same number of channels. Therefore, the PConv module is used instead of the 3×3 ordinary convolutional module in the ELAN module, thus reducing the number of parameters.

[0063] Figure 7 The diagram below shows the structure of the PConv module. The computational complexity of the PConv module is compressed as shown in the following formula:

[0064] (1)

[0065] In equation (1), h and w are the width and height of the feature map, k is the size of the convolution kernel, cp is the number of channels processed by the PConv module, and c is the number of channels processed by the Conv module. The computational complexity of the PConv module can be reduced as shown in the following formula:

[0066] (2)

[0067] The compression of memory access in the PConv module is as follows:

[0068] (3)

[0069] In equation (3), h and w are the width and height of the feature map, k is the size of the convolution kernel, cp is the number of channels used in the PConv module's convolution, and c is the number of channels used in the Conv module's convolution. (The last part is a partial sentence and doesn't need a direct translation.) The compression of memory access in the PConv module is as follows:

[0070] (4)

[0071] Specifically, in step S2, the FasterNet-CA is obtained by adding a CA attention mechanism to the first and third layers of FasterNet, and is inserted after the second Conv module in the FasterNet Block to enhance the ability to extract features and improve detection accuracy. As the network deepens, changes in positional information are often ignored during transmission. However, the CA attention mechanism not only acquires channel information but also considers orientation-related positional information, which helps the model to better locate and identify targets. Furthermore, it forms a pair of orientation-aware and position-sensitive feature maps with almost no additional computation, which can be complementaryly applied to the input feature maps to enhance the representation of the target of interest. Figure 8 This is a schematic diagram of the FasterNet Block structure. Figure 9 This is a schematic diagram of the FasterNet-CA structure.

[0072] In this embodiment, in step S3, since most multi-scale feature extraction uses a one-way information transmission method, it affects the fusion effect between different levels. This method suffers from feature information loss and degradation, which affects the final detection effect and the position of the predicted box. Therefore, an adaptive spatial ASFF structure is used for feature fusion at three levels. The adaptive spatial ASFF structure supports direct interaction between different levels. While retaining one-way information transmission, it fuses two adjacent low-level features into high-level features, thereby avoiding the feature information loss and degradation caused by one-way information transmission and the problem of large semantic gaps between different levels. This retains more useful information, resulting in better detection and more accurate localization. The adaptive spatial ASFF structure assigns different spatial weights to features at different levels, enhancing the importance of key levels and mitigating the influence of contradictory information from different targets, as shown in the following formula:

[0073] (5)

[0074] In equation (5), X, Y, and Z are features from Level 1, Level 2, and Level 3, respectively. The weight parameters α, β, and γ are multiplied and summed with the features from different levels to obtain the fused feature adaptive space ASFF structure, as shown below. Figure 10 As shown.

[0075] More specifically, in step S4, the images of workers riding the aerial passenger transport device are processed using the VOC dataset format to create a dataset, which includes a training set, a validation set, and a test set.

[0076] To verify the above technical solution, the following experiment was conducted in this embodiment:

[0077] 1. Experimental environment:

[0078] The operating system used is Windows 10, the CPU is an Intel(R) Core(TM) i7-12700F CPU @2.90GHz, the RAM is 32GB, and the GPU is an NVIDIA GeForce GTX 3080 with 12GB of VRAM. The YOLOv7 model uses the PyTorch deep learning framework and Python as the programming language. The parameters are set as follows: learning rate 0.01, momentum 0.937, weight decay 0.0005, and the network is trained for 100 epochs.

[0079] 2. Dataset:

[0080] To meet the testing requirements of aerial work platforms in mines, the data used in the experiment came from the actual work platform of such platforms. Based on image analysis of workers riding the platforms, it was found that large-scale use of the platforms is only during workers' commuting hours, with the platforms remaining unoccupied for the vast majority of the time. When workers need to go down into the mine, they hold the aerial work platform and attach it to the operating wire rope system. The process of a worker riding the platform is divided into two states: state 1 and state 2. State 1 involves holding the aerial work platform and waiting to be attached; state 2 involves the process of attaching the platform until the worker is seated. To ensure worker safety, the system checks whether workers are wearing safety helmets and whether they are carrying any foreign objects during the ride. When an unoccupied worker is detected, the auxiliary wire rope system stops operating. The dataset was created using the VOC dataset format, resulting in an XML file. Figure 11 As shown, the mining passenger carrying device dataset includes more than 6,500 images of various scenes in 5 types (denoted as hat, head, state1, state2 and illegal carrying, respectively), and is divided into training set, validation set and test set according to a randomly generated ratio of 8:1:1.

[0081] 3. Experimental Results and Analysis:

[0082] 3.1 Lightweight Network Experiment: Due to the complex and harsh environment of mines, problems such as poor lighting, complex backgrounds, blurred pedestrians, and multi-scale pedestrians exist. Current mainstream algorithms have large parameter counts and high computational costs, resulting in slow detection speeds and poor real-time detection performance, making them unsuitable for detecting aerial passenger transport devices in mines. To select a lightweight backbone network suitable for aerial passenger transport device detection, experiments compared MobileNetV3, GhostNetV1, GhostNetV2, ShuffleNetV2, PP-LCNet, and FasterNet. As shown in Table 1, FasterNet is slightly inferior to other models in terms of parameters, FLOPs, and model size, but the differences are not significant. FasterNet has unique advantages in mAP and FPS, reaching 0.941 and 98.04 respectively, both superior to other lightweight models. Therefore, this embodiment redesigns a lightweight backbone network based on FasterNet.

[0083] Table 1 Comparison of Lightweight Network Parameters

[0084]

[0085] 3.2 Ablation Experiment: To compensate for the loss of feature extraction information caused by the lightweight network, a FasterNet Block-CA suitable for aerial passenger vehicle detection was reconstructed based on FasterNet. The FasterNet Block-CA was added to different positions on the network backbone and compared. The results are as follows: Figure 12 As shown, the performance of FasterNet Block-CA is compared across different layers of the backbone network, with a comparison of the precision, recall, and mean accuracy (mAP) of each model.

[0086] like Figure 13 As shown, under similar parameter conditions, experimental data demonstrates that FasterNetBlock-CA2 achieved a recall of 0.928, while FasterNet Block-CA1-3 achieved precision and mean precision of 0.928 and 0.945, respectively. Therefore, FasterNet Block-CA1-3 was chosen, meaning CA attention was added to the first and third layers of FasterNet. Adding the CA attention mechanism to the first and third layers allows the network to better focus on a wide range of location information, which is more in line with the requirements of aerial passenger vehicle detection tasks.

[0087] Figure 12 , Figure 13In the diagram, FasterNet Block-CA1, FasterNet Block-CA2, FasterNet Block-CA3, FasterNet Block-CA4, FasterNet Block-CA1-2, FasterNet Block-CA1-3, FasterNet Block-CA1-4, and FasterNet Block-CA1~4 respectively represent the application of FasterNet Block-CA to the first, second, third, and fourth layers of the backbone network, the first and second layers, the first and third layers, the first and fourth layers, and each layer.

[0088] To further illustrate the effectiveness of each module in this embodiment, an ablation experiment was conducted to analyze the impact of different modules on the model detection performance. As shown in Table 2, the model with the backbone feature extraction network replaced by FasterNet-CA (FC) reduced the number of parameters by 12.1 MB and FLOPs by 60% compared to YOLOv7, but the mAP only decreased by 0.005. The model with the SPPCSPC module in the head detection head replaced by the GSCSP module reduced the number of parameters and computation, but the detection speed FPS decreased. The ELAN-P module (an improvement of the ELAN module using the PConv module) outperformed YOLOv7 in terms of number of parameters, computation, average accuracy, and detection speed. The model with the adaptive spatial ASFF structure increased the complexity of the model and decreased the detection speed, but the mAP improved by 0.021. FC+ASFF, ELAN-P+ASFF, GSCSP+ELAN-P, and FC+GSCSP+ELAN-P were experiments with different combinations of methods, while the MOOD (FC+GSCSP+ELAN-P+ASFF) algorithm model outperformed the other comparative models in both detection accuracy and detection speed. Experiments demonstrate that using lightweight FasterNet-CA, GSCSP, and PConv can significantly reduce the number of parameters and computational cost while improving detection speed, resulting in poor mAP performance. While the ASFF structure increases the number of parameters and computational cost significantly, it also enhances information exchange between non-adjacent layers, preventing information loss and degradation during transmission, retaining more effective information, and making the detection boxes more accurate. This leads to better detection results and fully meets the detection requirements of aerial passenger transport devices.

[0089] Table 2: Comparison of Ablation Experiment Results

[0090]

[0091] 3.3 Comparative Experiment: To address the issue of inaccurate prediction boxes, an adaptive spatial ASFF structure was adopted for connection. By allowing information to interact between hierarchical levels, information loss and degradation during transmission were avoided. Figure 14 The image shows the predicted bounding boxes and a magnified view of a portion of them. The red rectangles represent the predicted bounding boxes without the adaptive spatial ASFF structure, while the yellow rectangles represent the predicted bounding boxes with the adaptive spatial ASFF structure. As can be seen from the image, the predicted bounding boxes with the adaptive spatial ASFF structure are closer to the detected object, resulting in better target localization.

[0092] To objectively and fully demonstrate the advantages of the design model, a comparative experiment was conducted between the MOOD algorithm model and Faster R-CNN, SSD, YOLOX, YOLOv5, YOLOv7, and DETR models. All experimental models used the same dataset of aerial work platform for mining for training and inference validation. Table 3 shows the comparison results of parameters, computational cost (FLOPs), detection speed (FPS), average detection accuracy (mAP), and model weights for each model.

[0093] As shown in Table 3, under the same dataset, although the YOLOX and YOLOv5 models have fewer parameters and lower computational cost than the MOOD algorithm model, and the YOLOv5 model's weight file is also smaller, the MOOD algorithm model outperforms all other comparison models in terms of detection speed and average detection accuracy. In contrast, the YOLOv7 model achieves a lightweight network model, faster detection speed, and higher detection accuracy.

[0094] Table 3: Statistical Table of Performance Comparison Tests for Different Models

[0095]

[0096] like Figure 15 As shown: Faster R-CNN failed to detect some occluded hats when detecting hats; YOLOv5 failed to detect state1. The detection results of the YOLOX and MOOD algorithms are closer to the actual targets, and their detection location information is more accurate compared to other models. The confidence scores of the MOOD algorithm model for detecting hats, state1, and state2 are 0.90, 0.81, and 0.93, respectively. Compared to other models, the MOOD algorithm model has better detection performance.

[0097] from Figure 16As shown, due to the small size of the target (hat), the complexity of the scene, and the large scale of pedestrians, all models exhibited varying degrees of missed detections. However, for the detected targets, the MOOD algorithm model achieved confidence scores of 0.92 and 0.97 for hat and state1, respectively. This not only demonstrates more accurate location detection but also superior confidence scores compared to other comparative models.

[0098] from Figure 17 As shown, all models can detect the head well. The MOOD algorithm model has a confidence score of 0.83 for detecting the head. Compared with other models, the MOOD algorithm model has more accurate detection boxes.

[0099] from Figure 18 As shown, all models correctly detected illegal carrying, but the detection box of the YOLOv7 model did not completely cover the target information, while the MOOD algorithm model not only detected the target more accurately, but also had a confidence level of 0.92 in detecting illegal carrying.

[0100] In summary, the technical solution in this embodiment has the following characteristics:

[0101] (1) In view of the characteristics of the images of aerial passenger transport devices in mines, a dataset of aerial passenger transport devices in mines is proposed. Using part of the structure of YOLOv7 Head, the network model based on FasterNet, GSCSP and ELAN-P is redesigned to reduce the computational amount and number of parameters of the network model in order to improve the processing speed of the model.

[0102] (2) In order to compensate for the loss of feature extraction information caused by the lightweight network, a CA attention mechanism is added to the FasterNet module to enhance feature extraction and improve detection accuracy;

[0103] (3) Due to the harsh environment of the mine, poor lighting, complex background, blurred pedestrians and multi-scale pedestrians, this embodiment adopts an adaptive spatial ASFF structure to strengthen the information interaction between the upper and lower layers, improve the accuracy of the prediction box, and make the model detection effect better.

[0104] Compared with other mainstream algorithms, the MOOD algorithm model in this embodiment is only 47 MB ​​in size, has only 23.3 MB of parameters, achieves an mAP of 96.6%, and has a detection speed of 121 frames per second (FPS). It can meet the detection accuracy and speed requirements of aerial passenger transport devices in mines and can be used to reliably detect behaviors such as not wearing a safety helmet, carrying prohibited items, and passenger status.

[0105] The embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A rapid detection algorithm for aerial passenger transport devices in mines, characterized in that, Includes the following steps: S1. Based on the YOLOv7 framework, the network is reconstructed to obtain a lightweight network model: the backbone network in the original YOLOv7 is improved to FasterNet, the cross-connection GSCSP module is used to replace the SPPCSPC module in the YOLOv7 head detection head, and the PConv module is used to improve the ELAN module in the YOLOv7 head detection head. The network is reconstructed to reduce the amount of computation and parameters, resulting in a lightweight network model. The FasterNet is a structure composed of Embedding, FasterNet Block, Merging, and Classifier, and includes four stages. The first stage is processed by the Embedding layer, and the other stages are processed by the Merging layer for spatial downsampling and channel expansion. Each stage is composed of stacked FasterNet Blocks. Each FasterNet Block consists of one PConv and two subsequent PWConv 1×1 or Conv 1×1, and the three form an inverted residual architecture with a Shortcut connection to reuse input features. The ELAN structure is redesigned using PConv modules to obtain ELAN-PConv modules, or ELAN-P modules for short. ELAN-P modules are obtained by replacing the 3×3 ordinary convolutional modules in the ELAN modules with PConv modules. S2. Add the CA attention mechanism to FasterNet to obtain FasterNet-CA, so as to enhance feature extraction and make up for the problem of loss of feature extraction information caused by the lightweight network model. The FasterNet-CA is obtained by adding the CA attention mechanism to the first and third layers of FasterNet, and the CA attention mechanism is inserted after the second Conv module in the FasterNet Block; S3. Adopting an adaptive spatial ASFF structure to enhance the information interaction between the upper and lower layers of the lightweight network model, the MOOD algorithm model is obtained. S4. Collect images in the actual production environment of aerial passenger transport devices in mines, create a dataset, and then use the dataset to train, verify, and test the MOOD algorithm model, ultimately obtaining a fast detection algorithm for aerial passenger transport devices in mines.

2. The rapid detection algorithm for aerial passenger transport devices in mines according to claim 1, characterized in that: In step S1, the GSCSP module is a cross-level partial network designed based on the GSConv module. It consists of ordinary convolutional blocks and GSbottleneck modules, with only 18 MB of parameters.

3. The rapid detection algorithm for aerial passenger transport devices in mines according to claim 1, characterized in that: In step S3, the adaptive spatial ASFF structure supports direct interaction at different levels and assigns different spatial weights to features at different levels.

4. The rapid detection algorithm for aerial passenger transport devices in mines according to claim 1, characterized in that: In step S4, the images of workers riding the aerial passenger transport device are processed using the VOC dataset format to create a dataset, which includes a training set, a validation set, and a test set.

Images