Truck black smoke detection method and system based on edge intelligence
Patent Information
- Application Number
- CN202411317343.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2026-10-16
AI Technical Summary
[0005]本发明实施例提供一种基于边缘智能的货车黑烟检测方法及系统，可以解决现有技术中，存在的对于黑烟的检测识别误检率较高问题
[0038]本发明通过引入卷积模块ConvNormLayer、部分卷积Partial Convolution、逐点卷积PointWise Convolution和多路协调注意力MCA，可使得网络模型能够提取更深层次丰富的特征，避免网络模型中的梯度消失和信息丢失的问题，并能精确定位分散的微米级别的颗粒位置信息；同时引入尺度内特征交互模块AIFI、第一卷积模块CBS1、第二卷积模块CBS2和融合模块Fusion，能更为精确地融合不同维度的特征，以精确地分辨出空气中的烟雾信息和其他物质信息，从而精确检测货车的黑烟信息，为道路交通的高级别管理提供可行基础。
Smart Images
Figure CN122928689A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning-based black smoke detection technology, and in particular to a method and system for detecting black smoke from trucks based on edge intelligence. Background Technology
[0002] Over the past decade, my country has made remarkable achievements in the field of environmental protection. However, the country's economic development inevitably has an impact on the environment. In particular, in the field of road transportation, trucks are a key focus and challenge in the field of transportation environmental protection. The illegal emission of black smoke by trucks not only affects urban air quality but also endangers human health.
[0003] In the early days, road traffic management departments inspected illegal trucks manually, but this could not be done in real time around the clock. This led to the development of remote sensing technology, which has the characteristic of wide-area perception and can be applied to the field of black smoke emissions. However, due to its high installation conditions and costs, it is difficult to achieve dense deployment and accurate detection.
[0004] In recent years, the rapid development of machine learning has provided new ideas for black smoke emission detection. However, black smoke is often mixed with air, and in the early and later stages of identification, it is necessary to combine human and machine methods to judge the characteristics of micron-level particles in the black smoke area, which greatly increases the uncertainty of identification and results in a high false detection rate for black smoke. Summary of the Invention
[0005] This invention provides a method and system for detecting black smoke from trucks based on edge intelligence, which can solve the problem of high false detection rate in the existing technology for detecting and identifying black smoke.
[0006] This invention provides a method for detecting black smoke from trucks based on edge intelligence, comprising the following steps:
[0007] Obtain the image dataset of the truck;
[0008] An RT-DETR network model is constructed. In the Backbone of the RT-DETR network model, the ConvNormLayer convolution module, Partial Convolution, PointWise Convolution, and Multi-path Coordinated Attention (MCA) are cascaded into a multi-scale convolution module Stage. The ConvNormLayer convolution module and the multi-scale convolution module Stage are then stacked sequentially in the Backbone to form an improved Backbone.
[0009] An improved Neck section is formed by introducing sequentially cascaded intra-scale feature interaction module AIFI, first convolutional module CBS1, second convolutional module CBS2, and fusion module Fusion into the Neck section of the RT-DETR network model.
[0010] In the Head part of the RT-DETR network model, the decoder and the Head part are merged into an improved Head part;
[0011] The improved BackBone, Neck, and Head parts are concatenated to obtain an improved model of the RT-DETR network model; and the improved model is trained using a truck image dataset to obtain a lightweight black smoke detection model EC-RTDETR.
[0012] The image of the truck to be detected is input into the lightweight black smoke detection model EC-RTDETR. The improved Backbone part extracts features from the truck image to obtain feature maps of different scales containing micron-level particle features. The improved Neck part modifies the number of channels of the feature maps of different scales and performs feature fusion to obtain the final feature map, which is then visualized to form a truck black smoke recognition image. The improved Head part decoder extracts the target from the truck black smoke recognition image to obtain detection results including truck black smoke bounding boxes and confidence scores.
[0013] Preferably, the ConvNormLayer, Partial Convolution, PointWise Convolution, and Multipath Coordinated Attention (MCA) modules within the multi-scale convolution module Stage are stacked in a stacking cascade with stacking times of 2, 2, 1, and 1.
[0014] The ConvNormLayer convolution module in the BackBone section replaces the 7x7 convolution kernel with three consecutive 3x3 convolution kernels.
[0015] Preferably, the improved Backbone portion performs feature extraction on the truck image, wherein the Multi-Way Coordinated Attention (MCA) process includes the following steps:
[0016] The Multi-Path Coordinated Attention (MCA) divides the input features into three branches: A, B, and C. Branch A performs one-dimensional average pooling on the feature map in the height direction, branch B performs one-dimensional average pooling on the feature map in the width direction, and branch C is the channel attention branch.
[0017] The characteristic outputs of branches A and B are:
[0018]
[0019] Where: x c represents the input feature of the c-th channel; h represents the height of the feature map; w represents the width of the feature map; This represents the result of one-dimensional average pooling of the input features along the height direction. This represents the result of one-dimensional average pooling of the input features in the width direction; i represents the feature map; W represents the range of feature map i in the height direction; j represents the feature map; H represents the range of feature map j in the width direction.
[0020] Branches A and B concatenate feature maps containing global receptive field and precise location information, and then perform convolution to obtain output feature map e. Feature map e is split to output weights in the height and width directions, which are then used as weights. and The feature map e is then subjected to another convolution operation and sigmoid and split functions to output weights in the height and width directions, which are... and The acquisition process is as follows:
[0021] e = F(C[P(z) w ),z h ])
[0022]
[0023] f = δ(F(e))
[0024]
[0025] Where: C represents the spatial dimension connection operation; δ represents the non-linear activation function; f represents the intermediate feature map; F represents the convolution operation; S represents the split operation; X represents the original input features; P represents the permute operation; w h and w w These represent the product of the weights of the feature map in the height and width directions, respectively; z w This represents the one-dimensional average pooling result of the feature map in the width direction; z h This represents the result of one-dimensional average pooling of the feature map in the height direction; e h Indicates the height of the feature map; e w f represents the width of the feature map; h f represents the height of the intermediate feature map; w This indicates the width of the intermediate feature map.
[0026] Preferably, when the feature map channel number at different scales is modified and feature fusion is performed through the improved Neck section, the intra-scale feature interaction module AIFI includes the following steps:
[0027] The intra-scale feature interaction module AIFI is used to fuse high-level features, and the fusion process is as follows:
[0028] Q = K = V = Flatten (S5)
[0029]
[0030] F5=Reshape(Attention(Q,K,V))
[0031] Where: Q represents Query; K represents Key; V represents Value; Flatten represents vector flattening operation; Softmax represents non-linear activation function; S5 represents feature vector; Attention represents Multi-Head SelfAttention; Reshape represents changing the shape of the input feature map; CCFM represents CNN-based Cross-scale Feature-fusion Module; F5 represents the output feature after passing through the Intra-scale Feature Interaction Module AIFI.
[0032] Preferably, the first convolutional module CBS1 and the second convolutional module CBS2 are used to modify the number of channels in the feature map; the fusion module Fusion is used to fuse features of different dimensions.
[0033] This invention also provides a truck black smoke detection system based on edge intelligence, comprising:
[0034] An edge computing device configured to perform a truck black smoke detection method based on edge intelligence as described above;
[0035] The detection device is configured to identify the truck's license plate and upload the identified license plate to the cloud-based environmental monitoring platform when the edge computing device detects that the confidence score within the black smoke boundary box of the truck exceeds a preset range.
[0036] The environmental monitoring platform is configured to store and manage the uploaded license plates.
[0037] This invention provides a method and system for detecting black smoke from trucks based on edge intelligence. Compared with existing technologies, its advantages are as follows:
[0038] This invention introduces the ConvNormLayer convolution module, Partial Convolution, PointWise Convolution, and Multi-Path Coordinated Attention (MCA) to enable the network model to extract deeper and richer features, avoiding gradient vanishing and information loss problems in the network model, and accurately locating the positional information of scattered micron-level particles. Simultaneously, the introduction of the Intra-Scale Feature Interaction (AIFI) module, the first convolution module CBS1, the second convolution module CBS2, and the Fusion module enables more precise fusion of features from different dimensions to accurately distinguish smoke information and other substance information in the air, thereby accurately detecting black smoke information from trucks and providing a feasible basis for high-level road traffic management. Attached Figure Description
[0039] Figure 1 A schematic diagram of the overall process of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0040] Figure 2 A schematic diagram of data acquisition for a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0041] Figure 3 A schematic diagram of the overall structure of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0042] Figure 4 A schematic diagram of the model detection steps of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0043] Figure 5 A schematic diagram of the black smoke detection model architecture of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0044] Figure 6 A schematic diagram of the Stage multi-scale convolutional module structure for a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0045] Figure 7 A schematic diagram of the Attention-based Intra-scale Feature Interaction module structure of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention.
[0046] Figure 8 A schematic diagram of the MultiPath CoordinateAttention structure for a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0047] Figure 9 A schematic diagram of the decoder structure of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention;
[0048] Figure 10 This is a schematic diagram of the edge intelligence architecture of a truck black smoke detection method based on edge intelligence provided in an embodiment of the present invention. Detailed Implementation
[0049] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0050] See Figures 1-3 This invention provides a method for detecting black smoke from trucks based on edge intelligence, comprising:
[0051] The data acquisition module within the system provided by this invention uses its camera to monitor roads and collect vehicle image data.
[0052] Construct and train a black smoke detection model for black smoke detection. The black smoke detection model includes:
[0053] The data preprocessing module is used to preprocess the input image.
[0054] The black smoke feature extraction module includes a convolutional module ConvNormLayer and a multi-scale convolutional module Stage, which are stacked in sequence.
[0055] The black smoke feature fusion module includes the Attention-based Intra-scale Feature Interaction module, the convolution module CBS1, the convolution module CBS3, and the fusion module Fusion.
[0056] In the convolutional modules CBS1 and CBS3 and the fusion module Fusion, the convolutional modules CBS1 and CBS3 are used to modify the number of channels in the feature map, and the fusion module Fusion is used to fuse features of different dimensions. EC-RTDETR can scale the detector by modifying the number of RepConv modules and the dimension of the encoder.
[0057] The black smoke detection module includes a decoder and a head. EC-RTDETR selects a fixed number of features from the black smoke feature fusion module as object queries. After passing through the decoder, the head generates the target's bounding box, category, and confidence score. Finally, the image of the vehicle to be detected is input into the EC-RTDETR black smoke detection model for black smoke detection.
[0058] like Figure 3 The diagram shows the overall structure of the system of the present invention, including:
[0059] On the equipment side, it includes edge computing devices and detection devices for data acquisition, image preprocessing, black smoke detection, and license plate recognition.
[0060] Specifically, the black smoke detection model is deployed on edge computing devices to perform image preprocessing and black smoke detection; the detection devices mainly perform license plate recognition based on the detection results and upload them to the cloud-based environmental monitoring platform.
[0061] The cloud includes an environmental monitoring platform for displaying and storing test results.
[0062] The environmental monitoring platform includes the storage of black smoke detection and license plate recognition results, blacklist management, data push for environmental control during heavily polluted weather, and management and maintenance of front-end equipment.
[0063] like Figure 2 As shown, the data acquisition module includes a gantry and a high-definition camera for detection equipment. The high-definition camera is placed at the center of the crossbar of the gantry, with the lens facing the direction of the vehicle's movement and downward at a 45-degree angle to capture vehicle image data and video data. When the EC-RTDETR deployed on the edge device detects black smoke emanating from the truck, it saves a photo of the current time and a video of five seconds before and five seconds after the current time as evidence for law enforcement personnel later.
[0064] like Figure 4 and Figure 5 As shown, the lightweight black smoke detection model EC-RTDETR detects the obtained images. When the confidence score of an object in the image belonging to a black smoke vehicle is greater than or equal to 0.7, the currently detected vehicle is determined to be a black smoke vehicle. EC-RTDETR selects a fixed number of features from the black smoke feature fusion module as object queries. After passing through the decoder, the detection head generates the bounding box, category, and confidence score of the target.
[0065] The specific steps for detecting truck black smoke using the lightweight black smoke detection model EC-RTDETR are as follows:
[0066] Step 1: Create a dataset, including a training set, a validation set, and a test set. The ratio of the data volume of the training set, validation set, and test set is 7:1:2. Then, preprocess the training set and validation set to enhance image diversity and filter out redundant information.
[0067] Data augmentation operations on the training and validation sets include RandomHorizontalFilp, RandomColorJitter, RandomRain, RandomFog, RandomCrop, and Resize.
[0068] Random HorizontalFilp flips images horizontally, which can increase the diversity of images in a dataset.
[0069] RandomColorJitter adds a layer of white noise to the original image before performing uniform quantization; this can increase color richness and contrast, making black smoke features more obvious, thereby improving the detection performance of the model.
[0070] RandomRain and RandomFog are used to simulate weather in complex environments, improving the model's robustness to complex conditions.
[0071] Resizing randomly crops images into different sizes and aspect ratios, then scales the cropped images to a specified size of 640*640. This helps the model adapt to images of different sizes and scales, thereby improving the model's generalization ability and performance. During the inference phase, Resizing scales the input image to the specified size (640*640) and ensures the consistency of the input data, avoiding model validation errors due to images of different sizes. Specifically, each image has a 50% probability of undergoing different data augmentations.
[0072] Step 2: Construct a black smoke detection model.
[0073] The Stage convolution module is a network structure capable of extracting multi-scale features. It incorporates ConvNormLayer, Partial Convolution, PointWise Convolution, and MultiPath Coordinate Attention to optimize the model structure. This allows the network to scale to deeper layers, extracting richer features and avoiding the gradient vanishing and information loss problems common in deep networks. Figure 6 As shown.
[0074] Its workflow is as follows: The image undergoes feature extraction via ConvNormLayer and Stage; the multi-scale convolutional module Stage is stacked (2, 2, 1, 1); S5's high-level semantic features perform intra-scale feature interaction, then features at different scales are extracted from the Backbone for cross-scale feature fusion; finally, the features at different scales are transformed into image features; IoU-aware Query Selection selects 300 image features as the Decoder's initial object query; finally, the Decoder generates the target's bounding box and confidence score using the auxiliary prediction head. Specifically:
[0075] ① Preprocessing module, including:
[0076] Convolution operations are performed using convolutional layers with a kernel size of 2*2; data normalization is performed using batch normalization and ReLU activation functions.
[0077] ② Black smoke feature extraction module.
[0078] The black smoke feature extraction module includes a convolutional module ConvNormLayer and a multi-scale convolutional module Stage; the convolutional module ConvNormLayer and the multi-scale convolutional module Stage are stacked in sequence; specifically:
[0079] The ConvNormLayer convolution module is mainly used to adjust the number of channels in the feature map.
[0080] The ConvNormLayer convolution module replaces the 7x7 convolution kernel with three consecutive 3x3 convolution kernels, reducing the number of model parameters and improving computational efficiency while maintaining the same feature map output size. Data normalization and activation functions are applied after the convolution operation, accelerating model training and enhancing the model's non-linear expressive capabilities. Simultaneously, the stacking of multiple small convolution kernels increases the receptive field, enhancing the model's ability to capture contextual information. The computational cost of a 7x7 convolution is:
[0081] cost = 7 * 7 * C in *C out
[0082] Where: 7*7 represents the size of the convolution kernel; C in C represents the number of channels in the input feature map; out This indicates the number of channels in the output feature map.
[0083] The computational cost of a 3x3 convolution is:
[0084] cost = 3 * (3 * 3 * C) in *C out)
[0085] Compared to 7x7 convolution, 3x3 convolution reduces the number of parameters, decreasing the computational cost to about half, and thus reducing the training time of the model.
[0086] The multi-scale convolution module Stage is mainly used for feature extraction.
[0087] The multi-scale convolution module Stage consists of ConvNormLayer, Partial Convolution, PointWise Convolution, and MultiPath CoordinateAttention. Stage uses a combination of Partial Convolution and PointWise Convolution to reduce memory access and improve the model's inference speed, which is beneficial for deployment on edge devices. MultiPath CoordinateAttention combines feature information from both channel and spatial dimensions to improve the model's computational and multi-scale feature fusion capabilities.
[0088] MultiPath CoordinateAttention is:
[0089] Attention mechanisms enable models to focus on key information in the input data, thereby improving model performance and efficiency. MultiPath CoordinateAttention is an attention mechanism improved upon CoordinateAttention, used to enhance the model's attention to important features, such as... Figure 8 As shown; MultiPath CoordinateAttention embeds positional information into channel attention to enhance feature aggregation and selects appropriate channels for dimensionality reduction, thus reducing model complexity; MultiPath CoordinateAttention divides the input features into three branches (A, B, C); branch A represents one-dimensional average pooling of the feature map in the height direction, and branch B represents one-dimensional average pooling of the feature map in the width direction; the feature outputs of branches A and B are:
[0090]
[0091] Where: x c represents the input feature of the c-th channel; h represents the height of the feature map; w represents the width of the feature map; This represents the result of one-dimensional average pooling of the input features along the height direction. This represents the result of one-dimensional average pooling of the input features in the width direction; i represents the feature map; W represents the range of feature map i in the height direction; j represents the feature map; H represents the range of feature map j in the width direction.
[0092] Branches A and B concatenate feature maps containing global receptive field and precise location information for information exchange, and after convolution, obtain the output feature map e; feature map e outputs the weights in the height and width directions through a split operation. and Then, the feature map e undergoes a convolution operation and the sigmoid and split functions to output weights in the height and width directions. and ); and and The multiplication operation was applied to capture long-distance dependencies in the horizontal dimension while preserving precise positional relationships in the vertical dimension; the calculation process is as follows:
[0093] e = F(C[P(z) w ),z h ])
[0094]
[0095] f = δ(F(e))
[0096]
[0097] Where: C represents the spatial dimension connection operation; δ represents the non-linear activation function; f represents the intermediate feature map; F represents the convolution operation; S represents the split operation; X represents the original input features; P represents the permute operation; w h and w w These represent the product of the weights of the feature map in the height and width directions, respectively; z w This represents the one-dimensional average pooling result of the feature map in the width direction; z h This represents the result of one-dimensional average pooling of the feature map in the height direction; e h Indicates the height of the feature map; e w f represents the width of the feature map; h f represents the height of the intermediate feature map; w This indicates the width of the intermediate feature map.
[0098] After global average pooling, a convolution operation is performed to obtain the output feature map z; feature map z and f are multiplied to obtain the output feature map r; this can effectively capture the relationship between channels; finally, the original features X and W are...h W w After multiplying r, the final output Y is obtained; the calculation process is as follows:
[0099] z = F(GAP(X))
[0100] r = f * z
[0101] Y = X * w h *w w *r
[0102] Where: GAP represents global average pooling; r represents the product of the intermediate feature map f and the output feature map z of the C branch.
[0103] ③ Black smoke feature fusion module.
[0104] The black smoke feature fusion module includes an attention-based intra-scale feature interaction module, a convolutional module CBS1, a convolutional module CBS3, and a fusion module Fusion; specifically:
[0105] The Attention-based Intra-scale Feature Interaction module is primarily used to fuse high-level features, such as... Figure 7 As shown. Compared to the shallow features extracted by the model, high-level features have higher-level semantic features, which are beneficial for solving complex classification and detection tasks; therefore, the high-level features are subjected to the Attention-based Intra-scale Feature Interaction module to reduce computational cost and improve computational speed; the AIFI calculation process is as follows:
[0106] Q = K = V = Flatten (S5)
[0107]
[0108] F5=Reshape(Attention(Q,K,V))
[0109] Where: Q represents Query; K represents Key; V represents Value; Flatten represents vector flattening operation; Softmax represents non-linear activation function; S5 represents feature vector; Attention represents Multi-Head SelfAttention; Reshape represents changing the shape of the input feature map; CCFM represents CNN-based Cross-scale Feature-fusion Module; F5 represents the output feature after passing through Attention-based Intra-scale Feature Interaction Module; d k For the input dimensions of Query and Key; QK T Calculate the similarity score between Q and K.
[0110] The convolutional modules CBS1 and CBS3, along with the fusion module Fusion, are primarily used to modify the number of channels in the feature map and fuse features from different dimensions. CBS1 and CBS3 modify the number of channels in the feature map, while Fusion fuses features from different dimensions. EC-RTDETR can scale the detector by modifying the number of RepConv modules and the encoder dimension.
[0111] ④ Black smoke detection module.
[0112] The black smoke detection module includes a decoder and a detection head; EC-RTDETR selects a fixed number of features from the black smoke feature fusion module as object queries; after passing through the decoder, the detection head generates the target's bounding box, category, and confidence score, such as... Figure 9 As shown.
[0113] Step 3: Construction of the environmental monitoring platform.
[0114] The environmental monitoring platform is a crucial tool for environmental protection departments to detect and enforce regulations against illegal black smoke emissions from trucks. It effectively improves management and enforcement efficiency. It comprises four main functions: 1. Black smoke detection and license plate capture: Utilizing the method proposed in this invention, it accurately and effectively identifies trucks illegally emitting black smoke and captures their license plates, providing this information to law enforcement agencies as evidence. 2. Blacklist management: It registers trucks illegally emitting black smoke into a blacklist, allowing law enforcement personnel to focus their enforcement efforts on specific areas and blacklisted vehicles. 3. Heavy pollution weather environmental control data push: During periods of heavy pollution, it pushes monitoring data to traffic control departments for effective truck control and transportation channel management. 4. Front-end equipment management and maintenance: It monitors the status of front-end detection equipment in real time and issues alarms for any anomalies or malfunctions, enabling maintenance personnel to promptly repair and handle them. Figure 10 As shown.
[0115] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A method for detecting black smoke from trucks based on edge intelligence, characterized in that, Includes the following steps: Obtain the image dataset of the truck; An RT-DETR network model is constructed. In the Backbone of the RT-DETR network model, the ConvNormLayer convolution module, Partial Convolution, PointWise Convolution, and Multi-path Coordinated Attention (MCA) are cascaded into a multi-scale convolution module Stage. The ConvNormLayer convolution module and the multi-scale convolution module Stage are then stacked sequentially in the Backbone to form an improved Backbone. An improved Neck section is formed by introducing sequentially cascaded intra-scale feature interaction module AIFI, first convolutional module CBS1, second convolutional module CBS2, and fusion module Fusion into the Neck section of the RT-DETR network model. In the Head part of the RT-DETR network model, the decoder and the Head part are merged into an improved Head part; The improved BackBone, Neck, and Head parts are concatenated to obtain an improved model of the RT-DETR network model; and the improved model is trained using a truck image dataset to obtain a lightweight black smoke detection model EC-RTDETR. The image of the truck to be detected is input into the lightweight black smoke detection model EC-RTDETR. The improved Backbone part extracts features from the truck image to obtain feature maps of different scales containing micron-level particle features. The improved Neck part modifies the number of channels of the feature maps of different scales and performs feature fusion to obtain the final feature map, which is then visualized to form a truck black smoke recognition image. The improved Head part decoder extracts the target from the truck black smoke recognition image to obtain detection results including truck black smoke bounding boxes and confidence scores.
2. The method for detecting black smoke from trucks based on edge intelligence according to claim 1, characterized in that, The ConvNormLayer, Partial Convolution, PointWise Convolution, and Multipath Coordinated Attention (MCA) modules within the Stage multi-scale convolution module are stacked in concatenation with stacking times of 2, 2, 1, and 1. The ConvNormLayer convolution module in the BackBone section replaces the 7x7 convolution kernel with three consecutive 3x3 convolution kernels.
3. The method for detecting black smoke from trucks based on edge intelligence according to claim 1, characterized in that, The improved Backbone portion extracts features from the truck image, and when the Multi-Way Coordinated Attention (MCA) is executed, it includes the following steps: The Multi-Path Coordinated Attention (MCA) divides the input features into three branches: A, B, and C. Branch A performs one-dimensional average pooling on the feature map in the height direction, branch B performs one-dimensional average pooling on the feature map in the width direction, and branch C is the channel attention branch. The feature outputs of branches A and B are: Where: x c represents the input feature of the c-th channel; h represents the height of the feature map; w represents the width of the feature map; This represents the result of one-dimensional average pooling of the input features along the height direction. This represents the result of one-dimensional average pooling of the input features in the width direction; i represents the feature map; W represents the range of feature map i in the height direction; j represents the feature map; H represents the range of feature map j in the width direction. Branches A and B concatenate feature maps containing global receptive field and precise location information, and then perform convolution to obtain output feature map e. Feature map e is split to output weights in the height and width directions, which are then used as weights. and The feature map e is then subjected to another convolution operation and sigmoid and split functions to output weights in the height and width directions, which are... and The acquisition process is as follows: e=F(C[P(z w ),z h ]) f = δ(F(e)) Where: C represents the spatial dimension connection operation; δ represents the non-linear activation function; f represents the intermediate feature map; F represents the convolution operation; S represents the split operation; X represents the original input features; P represents the permute operation; w h and w w These represent the product of the weights of the feature map in the height and width directions, respectively; z w This represents the one-dimensional average pooling result of the feature map in the width direction; z h This represents the result of one-dimensional average pooling of the feature map in the height direction; e h Indicates the height of the feature map; e w f represents the width of the feature map; h f represents the height of the intermediate feature map; w This indicates the width of the intermediate feature map.
4. The method for detecting black smoke from trucks based on edge intelligence according to claim 1, characterized in that, When the feature map channel counts at different scales are modified and feature fusion is performed through the improved Neck section, the intra-scale feature interaction module AIFI includes the following steps: The intra-scale feature interaction module AIFI is used to fuse high-level features, and the fusion process is as follows: Q = K = V = Flatten (S5) F5=Reshape(Attention(Q,K,V)) Where: Q represents Query; K represents Key; V represents Value; Flatten represents vector flattening operation; Softmax represents non-linear activation function; S5 represents feature vector; Attention represents Multi-Head SelfAttention; Reshape represents changing the shape of the input feature map; CCFM represents CNN-based Cross-scale Feature-fusion Module; F5 represents the output feature after passing through the Intra-scale Feature Interaction Module AIFI.
5. The method for detecting black smoke from trucks based on edge intelligence according to claim 1, characterized in that, The first convolutional module CBS1 and the second convolutional module CBS2 are used to modify the number of channels in the feature map; the fusion module Fusion is used to fuse features of different dimensions.
6. A truck black smoke detection system based on edge intelligence, characterized in that, include: An edge computing device configured to perform a truck black smoke detection method based on edge intelligence as described in any one of claims 1 to 5; The detection device is configured to identify the truck's license plate and upload the identified license plate to the cloud-based environmental monitoring platform when the edge computing device detects that the confidence score within the black smoke boundary box of the truck exceeds a preset range. The environmental monitoring platform is configured to store and manage the uploaded license plates.