Methods for detecting obstacles in adverse weather conditions in AGV cargo trolleys used for copper electrode plate transfer.

Abstract

This invention relates to the field of target detection technology and discloses a method for obstacle detection in adverse weather conditions for AGV cargo vehicles used in copper electrode plate transfer. The method includes acquiring real-time images of the AGV vehicle to be detected under adverse weather conditions; inputting the real-time images to be detected into an improved YOLOX model for processing; and outputting obstacle detection results. The improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head. The dynamic self-supervised network includes an environment-aware encoding layer and a dynamic feature reconstructing layer. Based on the improved YOLOX model, this invention solves the key technical problem of feature coupling caused by environmental interference in visual detection systems under adverse weather conditions by introducing a dynamic self-supervised network, and also addresses the core issue of a surge in false alarm rates in the AGV vehicle visual detection system during copper electrode plate transfer under adverse weather conditions.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to an obstacle detection method for AGV cargo trolleys used for copper electrode plate transfer in adverse weather conditions. Background Technology

[0002] In the copper smelting production process, AGV cargo trolleys are widely used for the transfer of copper plates. They mainly run along fixed tracks to achieve efficient material transportation. When they encounter obstacles, they will trigger an alarm device and stop automatically. However, in bad weather, false alarms of AGV cargo trolleys have become the norm.

[0003] The current common solution is to install vision sensors on AGV cargo vehicles to identify obstacles in the track area through real-time image analysis, triggering an alarm and emergency braking when an obstacle is detected. However, traditional vision inspection systems face severe challenges in adverse weather conditions such as rain, snow, and fog. For example, raindrops create a refraction effect on the lens surface, resulting in dense random noise spots in the image, while dense fog significantly reduces image contrast. Rain and fog noise and obstacle features are deeply coupled in the frequency domain, causing a significant decrease in the signal-to-noise ratio of high-frequency components and a significant increase in the aliasing of mid- and low-frequency features. Furthermore, fixed threshold algorithms cannot adapt to dynamic lighting changes, leading to a significant increase in false alarm rates in adverse weather conditions, severely restricting the all-weather reliability of AGV systems.

[0004] Therefore, how to solve the problem of frequent false alarms in the transfer of copper plates by AGV trolleys under adverse weather conditions has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method for detecting obstacles in adverse weather conditions for AGV cargo carts used for copper electrode plate transfer, solving the problem of frequent false alarms during copper electrode plate transfer by AGV carts in adverse weather conditions.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a method for detecting obstacles in adverse weather conditions using an AGV (Automated Guided Vehicle) for transporting copper electrode plates, comprising:

[0007] Acquire real-time images of the AGV (Automated Guided Vehicle) to be inspected under adverse weather conditions;

[0008] The real-time image to be detected is input into the improved YOLOX model for processing, and the obstacle detection result is output; wherein,

[0009] The improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head;

[0010] The dynamic self-supervised network includes an environment-aware coding layer and a dynamic feature recombination layer.

[0011] As one preferred embodiment, the step of inputting the real-time image to be detected into the improved YOLOX model for processing and outputting obstacle detection results includes:

[0012] The backbone network is used to extract features from the real-time image to be detected, thereby obtaining multi-scale features.

[0013] The multi-scale features are fused according to the feature pyramid network to obtain multi-scale fused features;

[0014] The multi-scale fusion features are enhanced using the dynamic self-supervised network to obtain multi-scale enhanced features.

[0015] The multi-scale enhanced features are processed based on the detection head to generate a detection list, which is then processed by NMS to output the obstacle detection results.

[0016] As one preferred embodiment, the step of enhancing the multi-scale fused features based on the dynamic self-supervised network to obtain multi-scale enhanced features includes:

[0017] The multi-scale fusion features are processed by the environmental perception coding layer to extract weather-sensitive features and obtain the multi-scale weight parameters corresponding to the multi-scale weather intensity vector.

[0018] The multi-scale fusion features and the multi-scale weight parameters are input into the dynamic feature recombination layer for processing to obtain multi-scale enhanced features.

[0019] As one preferred embodiment, the environment-aware coding layer includes an adaptive average pooling layer, a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer; wherein,

[0020] The process of processing the multi-scale fused features through the environment-aware coding layer to extract weather-sensitive features and obtain multi-scale weight parameters corresponding to the multi-scale weather intensity vector includes:

[0021] The spatial dimension of each channel in the multi-scale fusion feature is compressed by the adaptive average pooling layer to obtain the multi-scale compressed feature.

[0022] The first convolutional layer is used to reduce the dimensionality of the multi-scale compressed features to obtain multi-scale dimensionality-reduced features.

[0023] The multi-scale dimensionality reduction features are activated based on the first activation layer to extract weather-sensitive features and obtain multi-scale activated features.

[0024] The multi-scale activation features are processed based on the second convolutional layer to obtain multi-scale weather intensity;

[0025] The multi-scale weather intensity is activated by the second activation layer to obtain the multi-scale weather intensity vector and its corresponding multi-scale weight parameters.

[0026] As one preferred embodiment, the dynamic feature recombination layer includes a first average pooling layer and a second average pooling layer; wherein...

[0027] The step of inputting the multi-scale fused features and the multi-scale weight parameters into the dynamic feature recombination layer for processing to obtain multi-scale enhanced features includes:

[0028] The low-frequency background in the multi-scale fusion features is extracted by the first average pooling layer, and the high-frequency components are determined based on the multi-scale fusion features and the low-frequency background.

[0029] The low-frequency components in the multi-scale fusion features are extracted using the second average pooling layer, and the high-frequency components and the low-frequency components are weighted and fused according to the multi-scale weight parameters to obtain the multi-scale enhanced features.

[0030] As one preferred embodiment, the detection head includes a regression branch layer, a confidence branch layer, and a classification branch layer; wherein,

[0031] The process of processing the multi-scale enhanced features based on the detection head to generate a detection list includes:

[0032] The bounding box coordinate offset of the multi-scale enhanced features is extracted through the regression branch layer to obtain the target offset;

[0033] The probability of obstacle presence in the multi-scale enhanced features is extracted using the confidence branch layer to obtain the target probability;

[0034] The multi-scale enhancement features are processed according to the classification branch layer to obtain the category probability distribution, which is then combined with the target offset and the target probability to form a detection list.

[0035] As a preferred embodiment, the dynamic self-supervised network further includes a self-supervised optimization layer; the self-supervised optimization layer includes a rotation prediction layer and a contrastive learning layer; the rotation prediction layer includes a rotation prediction head, a two-dimensional convolutional layer, and a first global average pooling layer; the contrastive learning layer includes a contrastive learning head, a gradient inversion layer, a third convolutional layer, and a second global average pooling layer; wherein...

[0036] The method further includes:

[0037] The target features in the multi-scale fusion features are randomly rotated and enhanced by the rotation prediction head to generate target rotation features;

[0038] The target rotation features are processed by the two-dimensional convolutional layer to extract local features, compress channels, and classify them to obtain the target processing features.

[0039] The target processing features are processed using the first global average pooling layer to generate logits, which are then converted into a probability distribution and cross-entropy loss is calculated to obtain the first loss.

[0040] The target features are projected through the contrast learning head to obtain the target projected features;

[0041] The target projection features are inverted by the gradient inversion layer to obtain the target robust features;

[0042] The target robust features are mapped to a low-dimensional space using the third convolutional layer, and the mapping result is processed by the second global average pooling layer to obtain a unit vector for calculating the contrastive loss, thus obtaining the second loss.

[0043] The first loss and the second loss are weighted and fused to obtain the self-supervised loss.

[0044] A second aspect of the present invention provides an obstacle detection system for an AGV cargo trolley used for copper electrode plate transfer in adverse weather conditions, comprising:

[0045] The image acquisition module is used to acquire real-time images of the AGV vehicle under adverse weather conditions.

[0046] The obstacle detection module is used to input the real-time image to be detected into the improved YOLOX model for processing and output the obstacle detection results; wherein,

[0047] The improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head;

[0048] The dynamic self-supervised network includes an environment-aware coding layer and a dynamic feature recombination layer.

[0049] A third aspect of the present invention provides an electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the above-described method for detecting obstacles in adverse weather conditions for an AGV cargo vehicle used for copper electrode plate transfer.

[0050] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the above-described method for detecting obstacles in adverse weather conditions for an AGV cargo trolley used for copper electrode plate transfer.

[0051] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0052] (1) The improved YOLOX model in this invention utilizes a dynamic self-supervised network to more accurately identify obstacles in complex and harsh scenarios, reducing false detections and missed detections; and utilizes a feature pyramid network to further enhance multi-scale feature fusion and improve the ability to detect obstacles.

[0053] (2) The dynamic self-supervised network analyzes weather conditions (such as rain and fog) in real time through the environmental perception coding layer, reducing the impact of image degradation caused by severe weather on detection. The dynamic feature recombination layer enhances obstacle edge and texture information based on the environmental coding results, thereby improving the target recognition in low-quality images. Attached Figure Description

[0054] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of an obstacle detection method for an AGV cargo trolley used for copper electrode plate transfer in adverse weather conditions, provided by a certain embodiment of the present invention;

[0056] Figure 2 This is a flowchart of step S2 provided in a certain embodiment of the present invention;

[0057] Figure 3 This is a flowchart of an obstacle detection method for an AGV cargo trolley used for copper electrode plate transfer in adverse weather conditions, provided by another embodiment of the present invention;

[0058] Figure 4 This is a structural diagram of an AGV cargo trolley obstacle detection system for copper electrode plate transfer in adverse weather conditions, provided in a certain embodiment of the present invention.

[0059] Figure 5 This is a structural diagram of an electronic device provided in a certain embodiment of the present invention;

[0060] Figure label:

[0061] Among them, 10 is the image acquisition module; 20 is the obstacle detection module; 5000 is the electronic device; 5001 is the processor; 5002 is the bus; 5003 is the memory; and 5004 is the transceiver. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0063] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0064] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the system or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0065] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is merely for describing specific embodiments and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0066] Currently, AGVs used in copper smelting generally employ obstacle detection solutions based on vision sensors, primarily using convolutional neural networks to identify targets in acquired images. Traditional methods rely on a general detection architecture, performing stably under normal lighting conditions, but their performance significantly degrades in rainy or foggy weather. Existing technologies often employ multi-exposure image fusion strategies, enhancing feature representation by superimposing images with different exposure durations. However, high-frequency noise in heavy rainfall environments severely interferes with the fusion effect, leading to a significant decrease in image quality. Existing vision detection solutions have significant limitations in adverse weather conditions, the core issue being the failure to effectively decouple the intrinsic relationship between weather noise and target features. Furthermore, while traditional multi-exposure fusion methods can improve the dynamic range of images, high-frequency noise caused by rain and fog interference exacerbates feature confusion, resulting in blurred obstacle outlines and distorted textures in the fused image. While frequency domain filtering techniques can suppress some noise, the fixed frequency band truncation strategy disrupts spatial continuity, leading to the loss of edge information for small obstacles, and lacks adaptive adjustment capabilities when rain and fog concentrations dynamically change. While generative adversarial network-based methods attempt to improve model generalization through domain transfer, the generated images often exhibit structural distortions, making it difficult to maintain the geometric integrity of key obstacles. At the same time, the complex network structure brings excessive computational load, failing to meet real-time detection requirements and making it difficult for the model to establish robust feature representations, resulting in an exponential increase in false detection rate as the environment deteriorates.

[0067] Based on this, in one embodiment, such as Figure 1 As shown, the first aspect of the present invention provides a method for detecting obstacles in adverse weather conditions for an AGV cargo trolley used for copper electrode plate transfer, comprising:

[0068] S1. Obtain real-time images of the AGV vehicle to be inspected under adverse weather conditions;

[0069] Specifically, a high dynamic range (HDR) camera is deployed at the front end of the AGV used for copper electrode plate transfer, supporting wide dynamic range (WDR) and automatic exposure compensation to adapt to strong light / backlight scenes; and an infrared camera is integrated as an auxiliary to provide thermal imaging information in low light or dense fog conditions, which is fused with visible light images to acquire real-time images of the AGV under adverse weather conditions; then these images are preprocessed and normalized to obtain a standardized tensor [B,C,H,W].

[0070] S2. The real-time image to be detected is input into the improved YOLOX model for processing, and the obstacle detection result is output; wherein, the improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head; the dynamic self-supervised network includes an environment-aware coding layer and a dynamic feature reconstruction layer;

[0071] In one embodiment, step S2 is as follows Figure 2As shown, it includes:

[0072] S21. The backbone network is used to extract features from the real-time image to be detected to obtain multi-scale features. Specifically, the backbone network in this invention adopts CSPDarknet, and then extracts features from the normalized tensor formed by the real-time image to be detected to obtain multi-scale features Dark3, Dark4, and Dark5.

[0073] S22. The multi-scale features are fused according to the feature pyramid network to obtain multi-scale fused features. Specifically, the feature pyramid network in this invention adopts YOLOPAFPN, and then the multi-scale features output by the backbone network are fused to obtain multi-scale fused features: P3 (high resolution, low-level detail features), P4 (medium resolution, mid-level semantic features), P5 (low resolution, high-level abstract features); wherein, these fused features are all four-dimensional tensors [B,C,H,W].

[0074] S23. Enhance the multi-scale fusion features according to the dynamic self-supervised network to obtain multi-scale enhanced features;

[0075] In one embodiment, step S23 includes:

[0076] The multi-scale fusion features are processed by the environmental perception coding layer to extract weather-sensitive features and obtain the multi-scale weight parameters corresponding to the multi-scale weather intensity vector.

[0077] The multi-scale fusion features and the multi-scale weight parameters are input into the dynamic feature recombination layer for processing to obtain multi-scale enhanced features.

[0078] Specifically, this invention utilizes an environmental perception coding layer to enhance weather-sensitive features in multi-scale features, generating a weather intensity vector, and then dynamically adjusting feature weights so that the model pays more attention to key features under severe weather conditions. The dynamic feature reorganization layer uses weight parameters to perform weighted fusion of multi-scale features, enhancing the identifiability of obstacles under severe weather conditions and reducing false detections and false negatives.

[0079] In one embodiment, the environment-aware coding layer includes an adaptive average pooling layer, a first convolutional layer, a first activation layer, a second convolutional layer, and a second activation layer; wherein,

[0080] The process of processing the multi-scale fused features through the environment-aware coding layer to extract weather-sensitive features and obtain multi-scale weight parameters corresponding to the multi-scale weather intensity vector includes:

[0081] The spatial dimension of each channel in the multi-scale fusion feature is compressed by the adaptive average pooling layer to obtain the multi-scale compressed feature.

[0082] The first convolutional layer is used to reduce the dimensionality of the multi-scale compressed features to obtain multi-scale dimensionality-reduced features.

[0083] The multi-scale dimensionality reduction features are activated based on the first activation layer to extract weather-sensitive features and obtain multi-scale activated features.

[0084] The multi-scale activation features are processed based on the second convolutional layer to obtain multi-scale weather intensity;

[0085] The multi-scale weather intensity is activated by the second activation layer to obtain the multi-scale weather intensity vector and its corresponding multi-scale weight parameters.

[0086] Specifically, in the environment-aware coding layer, P3, P4, and P5 feature maps are received from the feature pyramid network output, but the operations on these features are independent. Taking P3 as an example, the processing of multi-scale fusion features by the environment-aware coding layer is explained: This invention compresses the spatial dimension (H×W) of each channel in the P3 feature map into a single value using AdaPtiveAvgPool2d(1) (adaptive average pooling layer), obtaining the P3 compressed feature [B,C,1,1]; the number of channels in the P3 compressed feature is reduced from C to the median value using Conv1x1 (first convolutional layer), obtaining the P3 dimensionality-reduced feature; the P3 dimensionality-reduced feature is activated by the ReLU function (first activation layer), mainly to reduce the number of parameters and retain key weather-sensitive features. To enhance nonlinear expressive power, P3 activation features are obtained. Conv1×1 (the second convolutional layer) is used to process the P3 activation features, outputting a 3D weather intensity, i.e., P3 weather intensity. Finally, after activating the P3 weather intensity using the Sigmoid function (the second activation layer), the P3 weather intensity vector [B,3] is output, along with its corresponding P3 weight parameters (β-rain, β-fog, β-snow) ∈ [0,1]. Here, β is the weight parameter, and [0,1] represents the degree of severe weather; rain, fog, and snow represent rain, fog, and snow, respectively, indicating severe weather types. Similarly, the same processing procedure applies to P4 and P5. Combining the P3, P4, and P5 weight parameters yields multi-scale weight parameters. Other multi-scale features can also be obtained by combining the features corresponding to each fused feature, which will not be elaborated upon here. Furthermore, the output parameters β-rain, β-fog, and β-snow of the environmental perception coding layer are the core control unit for the model to achieve "dynamic weather countermeasures," quantifying the intensity of weather interference and adjusting the feature recombination strategy.

[0087] This invention significantly reduces computation by compressing the spatial dimension of multi-scale features, while retaining important information from the channel dimension to avoid information loss. A first convolutional layer maps high-dimensional features to a low-dimensional space, reducing the complexity of subsequent calculations and improving inference speed. Nonlinear activation extracts key weather-related features, enhancing the model's ability to perceive severe weather. A second convolutional layer and a second activation layer generate a weather intensity vector, quantifying the sensitivity of features at different scales to weather conditions and providing a basis for dynamic weight allocation. Weight parameters are dynamically adjusted according to weather conditions, allowing the model to focus more on key features (such as obstacle edges) and reduce noise interference under severe weather conditions. Containing only two convolutional and two activation layers, it has a small number of parameters, making it suitable for deployment on embedded devices in AGV vehicles.

[0088] In one embodiment, the dynamic feature recombination layer includes a first average pooling layer and a second average pooling layer; wherein,

[0089] The step of inputting the multi-scale fused features and the multi-scale weight parameters into the dynamic feature recombination layer for processing to obtain multi-scale enhanced features includes:

[0090] The low-frequency background in the multi-scale fusion features is extracted by the first average pooling layer, and the high-frequency components are determined based on the multi-scale fusion features and the low-frequency background.

[0091] The low-frequency components in the multi-scale fusion features are extracted using the second average pooling layer, and the high-frequency components and the low-frequency components are weighted and fused according to the multi-scale weight parameters to obtain the multi-scale enhanced features.

[0092] Specifically, this invention still uses P3 as an example to illustrate the processing of multi-scale fusion features and multi-scale weight parameters by the dynamic feature recombination layer: taking the P3 fusion feature and its corresponding β parameter as input, the dynamic feature recombination layer obtains the low-frequency background AvgPool3x3(P3) through AvgPool3x3 (the first average pooling layer), and then subtracts the low-frequency background from the original P3 fusion feature to obtain the high-frequency component P3-high=P3-AvgPool3×3(P3), thereby separating high-frequency noise such as rain and fog, providing a basis for subsequent suppression; subsequently, the dynamic feature recombination layer... The feature reassembly layer extracts blurred details from the P3 fusion feature using AvgPool5×5 (the second average pooling layer), extracts the contour in the fog, and preserves the main object contour, resulting in the low-frequency component P3-low = AvgPool5×5(P3), which enhances the overall shape information of the blurred target in the fog. Finally, the high-frequency and low-frequency components are fused using the β parameter corresponding to the P3 fusion feature, weighted by β-rain and β-fog, to obtain the P3 enhanced feature: P3-enh = β-rain * P3-high + β-fog * P3-low. Furthermore, a high β-rain weight weakens high-frequency noise, while a high β-fog weight enhances low-frequency contours. The fusion weights of high-frequency and low-frequency components can be adaptively adjusted to achieve a balance between noise suppression and target enhancement. The same operation applies to P4 and P5.

[0093] This invention, by separating low-frequency background and high-frequency components, can capture different information in images or data more meticulously. Processing and fusing these components separately results in multi-scale enhanced features that retain overall structural information while highlighting detailed information, thus enhancing the expressive power of the features. By using multi-scale weight parameters to weight and fuse high-frequency and low-frequency components, the proportions of these components in the final features can be dynamically adjusted according to different task requirements and data characteristics. This allows the model to better adapt to different scenarios and data distributions, improving its generalization ability. Furthermore, based on an improved YOLOX model, this invention addresses the key technical challenge of feature coupling caused by environmental interference in visual detection systems under adverse weather conditions by introducing a dynamic self-supervised network. This network extracts weather intensity parameters in real time through an environment-aware coding layer and dynamically adjusts the feature recombination weights, achieving an adaptive balance between high-frequency noise suppression and low-frequency contour enhancement, ensuring stable detection under different weather conditions.

[0094] S24. Based on the detection head, the multi-scale enhanced features are processed to generate a detection list, which is then processed by NMS to output the obstacle detection results.

[0095] In one embodiment, the detection head includes a regression branch layer, a confidence branch layer, and a classification branch layer; wherein,

[0096] The process of processing the multi-scale enhanced features based on the detection head to generate a detection list includes:

[0097] The bounding box coordinate offset of the multi-scale enhanced features is extracted through the regression branch layer to obtain the target offset;

[0098] The probability of obstacle presence in the multi-scale enhanced features is extracted using the confidence branch layer to obtain the target probability;

[0099] The multi-scale enhancement features are processed according to the classification branch layer to obtain the category probability distribution, which is then combined with the target offset and the target probability to form a detection list.

[0100] Specifically, this invention inputs multi-scale augmented features P3-enh, P4-enh, and P5-enh into the detection head for independent prediction. The bounding box coordinate offsets of the multi-scale augmented features are extracted using a Conv3x3 (regression branch layer), and the target offset is output. The obstacle presence probability in the multi-scale augmented features is extracted using a Conv3×3 (confidence branch layer), and the target probability is output. The multi-scale augmented features are processed using a Conv3x3 (classification branch layer), and the class probability distribution is output. The outputs of each layer in the detection head are integrated. For each preset anchor box, its corresponding target offset, target probability, and class probability distribution are combined into a detection result entry. All detection result entries corresponding to the preset anchor boxes are summarized to form a detection list [P3-Pred, P4-Pred, P5-Pred], which is then output. The detection list includes the coordinates, confidence scores, and class probabilities of all detection boxes. Finally, the detection list is filtered for redundant detection boxes using non-maximum suppression (NMS), and the final detection box list is output. The final detection box list shows the coordinates, category, and confidence level of the target obstacle, i.e., the obstacle detection result.

[0101] This invention extracts bounding box coordinate offsets through a regression branch layer to accurately determine the target's position in the image; extracts the probability of obstacle presence through a confidence branch layer to provide a confidence score for each detected target, reducing false detections; and obtains the category probability distribution through a classification branch layer, enabling accurate target classification and thus improving the accuracy of obstacle detection.

[0102] In one embodiment, the dynamic self-supervised network further includes a self-supervised optimization layer; the self-supervised optimization layer includes a rotation prediction layer and a contrastive learning layer; the rotation prediction layer includes a rotation prediction head, a two-dimensional convolutional layer, and a first global average pooling layer; the contrastive learning layer includes a contrastive learning head, a gradient inversion layer, a third convolutional layer, and a second global average pooling layer; wherein,

[0103] The method further includes:

[0104] The target features in the multi-scale fusion features are randomly rotated and enhanced by the rotation prediction head to generate target rotation features;

[0105] The target rotation features are processed by the two-dimensional convolutional layer to extract local features, compress channels, and classify them to obtain the target processing features.

[0106] The target processing features are processed using the first global average pooling layer to generate logits, which are then converted into a probability distribution and cross-entropy loss is calculated to obtain the first loss.

[0107] The target features are projected through the contrast learning head to obtain the target projected features;

[0108] The target projection features are inverted by the gradient inversion layer to obtain the target robust features;

[0109] The target robust features are mapped to a low-dimensional space using the third convolutional layer, and the mapping result is processed by the second global average pooling layer to obtain a unit vector for calculating the contrastive loss, thus obtaining the second loss.

[0110] The first loss and the second loss are weighted and fused to obtain the self-supervised loss.

[0111] Specifically, the dynamic self-supervised network in this invention also includes a self-supervised optimization layer, which is used to quantify the self-supervised loss of the improved YOLOX model and then optimize the model parameters. The self-supervised optimization layer uses the P5 fusion feature in the multi-scale fusion feature as the target feature, and performs random rotation enhancement on the target feature by rotating the prediction head. The feature map is randomly rotated by any angle k∈{0°,90°,180°,270°} to generate the rotated feature, namely the target rotation feature X-rot. Then, it is processed by two layers of nn.Conv2d (two-dimensional convolutional layers). The spatial local features of the target rotation feature are extracted by the 3×3 convolution in the two-dimensional convolutional layer, and the target rotation feature is channel compressed and classified by the 1×1 convolution in the two-dimensional convolutional layer to obtain the target processed feature. The target processed feature is processed by the first global average pooling GAP to generate logits (the original output value of the last layer of the model). The logits are converted into a probability distribution by the Softmax function, and the cross-entropy loss is calculated according to the actual rotation angle and the predicted probability distribution to obtain the first loss loss-rot.

[0112] The original P5 fused features are projected into a low-dimensional space using a contrastive learning head, which typically consists of one or more fully connected layers. A gradient inversion layer is used to invert the gradient of the target projected features, i.e., the gradient is inverted and multiplied by a coefficient of 0.3 during backpropagation. This forces the backbone network to suppress weather-related features, generating weather-independent robust features, which are then directly passed to the input features during forward propagation. A 1×1 convolution (the third convolutional layer) maps the target robust features from high-dimensional to low-dimensional space. The mapping result is then processed by a second global average pooling layer (GAP) and normalization to generate unit vectors. The contrastive loss is calculated using the InfoNCE loss function, resulting in the second loss, loss-con. Finally, the self-supervised loss, loss-ssl = loss-rot + 0.5 * loss-con, is calculated using a weighted fusion formula, and the self-supervised loss is output.

[0113] This invention employs a self-supervised optimization layer to constrain the dynamic self-supervised network to learn more discriminative features, reducing overfitting to weather noise. Self-supervised contrastive learning and rotation prediction enhance the model's generalization ability to unseen weather conditions. Furthermore, a gradient inversion (GRL) mechanism is used to suppress the activation intensity of noise-related feature channels, forcing the network to learn robust feature representations independent of weather, fundamentally decoupling the mixed features of noise and the target. Finally, weighted fusion and backpropagation of multiple losses are used for optimization, thereby improving target detection accuracy and adaptability to harsh environments. This allows the model to adaptively adjust feature extraction strategies based on real-time rain and fog intensity, thus solving the core problem of a surge in false alarm rates in AGV visual detection systems during copper electrode plate transport in severe weather.

[0114] After the detection head outputs the detection list, the detection loss L-det is calculated based on the original prediction results and the true labels (including bounding box coordinates, category labels, and target presence labels). The detection loss typically includes bounding box regression loss (e.g., SmoothL1 loss), confidence loss (e.g., binary cross-entropy loss), and classification loss (e.g., cross-entropy loss). These are weighted and fused with the self-supervised loss to obtain the total loss L-total = L-det + 0.3 * loss - ssl. Then, backpropagation is performed using the total loss to optimize the network parameters of the environment-aware coding layer, dynamic feature reorganization layer, and dynamic self-supervised network, blocking the gradient of the backbone network. This invention, through multi-band feature reorganization and self-supervised learning mechanisms, enables the improved YOLOX model to automatically enhance target contours and suppress weather noise in adverse environments such as rain and fog, significantly improving detection accuracy in complex weather scenes while maintaining efficient inference speed. A flowchart of another method for obstacle detection in adverse weather conditions using an AGV cargo vehicle for copper electrode plate transport is shown below. Figure 3As shown, through the synergistic innovation of dynamic self-supervised networks and environmental perception mechanisms, the technical limitations of traditional visual detection schemes are overcome. Based on a feature decoupling strategy using gradient inversion layers and contrastive learning, the frequency-domain spatial coupling problem between rain / fog noise and target semantic features is fundamentally solved. The environmental perception coding layer dynamically adjusts the fusion weights of high-frequency suppression and low-frequency enhancement in real time, overcoming the failure defect of fixed filtering algorithms when rain / fog concentration changes abruptly. The dynamic self-supervised network adaptively adjusts the feature extraction strategy according to the real-time rain / fog intensity, improving network robustness and reducing dependence on large-scale labeled data, enabling the system to perform well in both small and large target detection. The overall design balances accuracy and real-time performance, meeting the real-time obstacle detection requirements of AGV vehicles transporting copper electrode plates in adverse weather conditions.

[0115] This application proposes a method for obstacle detection in adverse weather conditions for AGV (Automated Guided Vehicle) cargo vehicles transporting copper plates, addressing the frequent false alarms during such transport. The method acquires real-time images of the AGV under adverse weather conditions, processes these images using an improved YOLOX model, and outputs obstacle detection results. The improved YOLOX model comprises a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head. The dynamic self-supervised network includes an environment-aware coding layer and a dynamic feature reconstruction layer. The improved YOLOX model utilizes the dynamic self-supervised network to more accurately identify obstacles in complex and adverse scenarios, reducing false positives and false negatives. The feature pyramid network further enhances multi-scale feature fusion, improving obstacle detection capabilities. The dynamic self-supervised network analyzes weather conditions (such as rain and fog) in real-time through the environment-aware coding layer, reducing the impact of image degradation caused by adverse weather on detection. The dynamic feature reconstruction layer strengthens obstacle edge and texture information based on the environment coding results, improving target recognizability in low-quality images.

[0116] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0117] In another embodiment, such as Figure 4 As shown, a second aspect of the present invention provides an obstacle detection system for an AGV cargo trolley used for copper electrode plate transfer in adverse weather conditions, comprising:

[0118] Image acquisition module 10 is used to acquire real-time images of the AGV vehicle under adverse weather conditions.

[0119] The obstacle detection module 20 is used to input the real-time image to be detected into the improved YOLOX model for processing and output the obstacle detection result; wherein,

[0120] The improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head;

[0121] The dynamic self-supervised network includes an environment-aware coding layer and a dynamic feature recombination layer.

[0122] It should be noted that the various modules in the aforementioned AGV cargo cart obstacle detection system for copper electrode plate transfer in adverse weather conditions can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module. For specific limitations regarding the AGV cargo cart obstacle detection system for copper electrode plate transfer in adverse weather conditions, please refer to the limitations of the AGV cargo cart obstacle detection method for copper electrode plate transfer in adverse weather conditions described above; both have the same function and role, and will not be repeated here.

[0123] A third aspect of the present invention provides an electronic device comprising:

[0124] Processor, memory, and bus;

[0125] The bus is used to connect the processor and the memory;

[0126] The memory is used to store operation instructions;

[0127] The processor is configured to execute instructions by calling the operation instructions, causing the processor to perform operations corresponding to the adverse weather obstacle detection method for an AGV cargo vehicle used for copper electrode plate transfer as shown in the first aspect of this application.

[0128] In one alternative embodiment, an electronic device is provided, such as Figure 5 As shown, Figure 5 The illustrated electronic device 5000 includes a processor 5001 and a memory 5003. The processor 5001 and the memory 5003 are connected, for example, via a bus 5002. Optionally, the electronic device 5000 may also include a transceiver 5004. It should be noted that in practical applications, the transceiver 5004 is not limited to one type, and the structure of this electronic device 5000 does not constitute a limitation on the embodiments of this application.

[0129] Processor 5001 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 5001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0130] Bus 5002 may include a path for transmitting information between the aforementioned components. Bus 5002 may be a PCI bus or an EISA bus, etc. Bus 5002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0131] The memory 5003 may be a ROM or other type of static storage device capable of storing static information and instructions, RAM or other type of dynamic storage device capable of storing information and instructions, or it may be an EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0132] The memory 5003 is used to store application code that executes the scheme of this application, and its execution is controlled by the processor 5001. The processor 5001 is used to execute the application code stored in the memory 5003 to implement the content shown in any of the foregoing method embodiments.

[0133] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers.

[0134] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for detecting obstacles in adverse weather conditions for an AGV cargo vehicle used for copper electrode plate transfer, as shown in the first aspect of this application.

[0135] Another embodiment of this application provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the aforementioned method embodiments.

[0136] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0137] In summary, this invention relates to the field of target detection technology and discloses a method for obstacle detection in adverse weather conditions for AGV cargo vehicles used for copper electrode plate transfer. The method includes acquiring real-time images of the AGV vehicle to be detected under adverse weather conditions; inputting the real-time images to be detected into an improved YOLOX model for processing; and outputting obstacle detection results. The improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network, and a detection head. The dynamic self-supervised network includes an environment-aware encoding layer and a dynamic feature reconstructing layer. Based on the improved YOLOX model, this invention solves the key technical problem of feature coupling caused by environmental interference in visual detection systems under adverse weather conditions by introducing a dynamic self-supervised network, and also addresses the core problem of a surge in false alarm rates in the AGV vehicle visual detection system during copper electrode plate transfer under adverse weather conditions.

[0138] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0139] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are 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 improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for detecting obstacles in adverse weather conditions for an AGV carrying vehicle for transporting copper pole plates, characterized in that, The method comprises the steps of: acquiring a real-time image to be detected of an AGV in harsh weather; inputting the real-time image to be detected into an improved YOLOX model for processing, and outputting an obstacle detection result; wherein the improved YOLOX model comprises a backbone network, a feature pyramid network, a dynamic self-supervised network and a detection head; the dynamic self-supervised network comprises an environment perception encoding layer and a dynamic feature reorganization layer; the step of inputting the real-time image to be detected into the improved YOLOX model for processing, and outputting an obstacle detection result comprises: extracting features of the real-time image to be detected through the backbone network to obtain multi-scale features; fusing the multi-scale features according to the feature pyramid network to obtain multi-scale fusion features; enhancing features of the multi-scale fusion features according to the dynamic self-supervised network to obtain multi-scale enhanced features; processing the multi-scale enhanced features based on the detection head to generate a detection list and perform NMS processing, and outputting an obstacle detection result; wherein NMS is non-maximum suppression; the dynamic self-supervised network further comprises a self-supervised optimization layer; the self-supervised optimization layer comprises a rotation prediction layer and a contrast learning layer; the rotation prediction layer comprises a rotation prediction head, a two-dimensional convolution layer and a first global average pooling layer; the contrast learning layer comprises a contrast learning head, a gradient inversion layer, a third convolution layer and a second global average pooling layer; wherein the method further comprises: randomly rotating and enhancing target features in the multi-scale fusion features through the rotation prediction head to generate target rotation features; extracting local features, compressing channels and classifying the target rotation features through the two-dimensional convolution layer to obtain target processing features; processing the target processing features through the first global average pooling layer to generate logits to convert into a probability distribution and perform cross-entropy loss calculation to obtain a first loss; projecting the target features through the contrast learning head to obtain target projection features; inverting gradients of the target projection features through the gradient inversion layer to obtain target robust features; mapping the target robust features to a low-dimensional space through the third convolution layer, and processing the mapping results through the second global average pooling layer to obtain unit vectors to calculate a contrast loss and obtain a second loss; performing weighted fusion processing on the first loss and the second loss to obtain a self-supervised loss.

2. The AGV vehicle obstacle detection method for harsh weather of the copper pole plate transfer according to claim 1, characterized in that, the step of enhancing features of the multi-scale fusion features according to the dynamic self-supervised network to obtain multi-scale enhanced features comprises: processing the multi-scale fusion features through the environment perception encoding layer to extract weather-sensitive features to obtain multi-scale weight parameters corresponding to multi-scale weather intensity vectors; inputting the multi-scale fusion features and the multi-scale weight parameters into the dynamic feature reorganization layer for processing to obtain multi-scale enhanced features.

3. The AGV vehicle obstacle detection method for harsh weather of the copper pole plate transfer according to claim 2, characterized in that, the environment perception encoding layer comprises an adaptive average pooling layer, a first convolution layer, a first activation layer, a second convolution layer and a second activation layer; wherein The multi-scale fusion feature is processed by the environment perception encoding layer to extract weather-sensitive features, so as to obtain multi-scale weather intensity vectors corresponding to multi-scale weight parameters, including: The spatial dimensions of each channel in the multi-scale fusion feature are compressed by the adaptive average pooling layer to obtain multi-scale compressed features; The multi-scale compressed features are processed by the first convolution layer for dimension reduction to obtain multi-scale reduced features; The multi-scale reduced features are activated by the first activation layer to extract weather-sensitive features, so as to obtain multi-scale activated features; The multi-scale activated features are processed by the second convolution layer to obtain multi-scale weather intensity; The multi-scale weather intensity is activated by the second activation layer to obtain multi-scale weather intensity vectors and corresponding multi-scale weight parameters.

4. The AGV for copper pole plate transfer according to claim 2, wherein, The dynamic feature reorganization layer includes a first average pooling layer and a second average pooling layer; wherein, The multi-scale fusion feature and the multi-scale weight parameter are input into the dynamic feature reorganization layer for processing to obtain multi-scale enhanced features, including: The low-frequency background in the multi-scale fusion feature is extracted by the first average pooling layer, and the high-frequency component is determined according to the multi-scale fusion feature and the low-frequency background; The low-frequency component in the multi-scale fusion feature is extracted by the second average pooling layer, and the high-frequency component and the low-frequency component are weighted and fused according to the multi-scale weight parameter to obtain multi-scale enhanced features.

5. The AGV for copper pole plate transfer according to claim 1, wherein, The detection head includes a regression branch layer, a confidence branch layer and a classification branch layer; wherein, The multi-scale enhanced features are processed based on the detection head to generate a detection list, including: The boundary box coordinate offset of the multi-scale enhanced features is extracted by the regression branch layer to obtain a target offset; The obstacle existence probability in the multi-scale enhanced features is extracted by the confidence branch layer to obtain a target probability; The multi-scale enhanced features are processed by the classification branch layer to obtain a class probability distribution, which is combined with the target offset and the target probability to form a detection list.

6. A bad weather obstacle detection system for AGV load carrier of copper pole plate transfer, characterized in that, including: An image acquisition module is configured to acquire a real-time image to be detected of an AGV in severe weather; An obstacle detection module is configured to input the real-time image to be detected into an improved YOLOX model for processing, and output an obstacle detection result; wherein, The improved YOLOX model includes a backbone network, a feature pyramid network, a dynamic self-supervised network and a detection head; The dynamic self-supervised network includes an environment perception encoding layer and a dynamic feature reorganization layer; The real-time image to be detected is input into the improved YOLOX model for processing to output an obstacle detection result, including: The multi-scale features are extracted from the real-time image to be detected by the backbone network; The multi-scale fusion features are obtained by fusing the multi-scale features according to the feature pyramid network; The multi-scale enhanced features are obtained by feature enhancement of the multi-scale fusion features according to the dynamic self-supervised network; Based on the detection head, the multi-scale enhanced features are processed to generate a detection list and processed by NMS to output an obstacle detection result; wherein, NMS is non-maximum suppression; The dynamic self-supervised network further comprises a self-supervised optimization layer; the self-supervised optimization layer comprises a rotation prediction layer and a contrast learning layer; the rotation prediction layer comprises a rotation prediction head, a two-dimensional convolution layer and a first global average pooling layer; the contrast learning layer comprises a contrast learning head, a gradient inversion layer, a third convolution layer and a second global average pooling layer; wherein, The system further comprises: The target feature in the multi-scale fusion feature is randomly rotated and enhanced by the rotation prediction head to generate a target rotation feature; The target rotation feature is processed by the two-dimensional convolution layer to extract local features, compress channels and classify to obtain a target processing feature; The target processing feature is processed by the first global average pooling layer to generate logits to convert to a probability distribution and calculate cross-entropy loss, thereby obtaining a first loss; The target feature is projected by the contrast learning head to obtain a target projection feature; The target projection feature is inverted by the gradient inversion layer to obtain a target robust feature; The target robust feature is mapped to a low-dimensional space by the third convolution layer, and the mapping result is processed by the second global average pooling layer to obtain a unit vector to calculate a contrast loss, thereby obtaining a second loss; The first loss and the second loss are weighted and fused to obtain a self-supervised loss.

7. An electronic device, comprising: The computer readable storage medium comprises a stored computer program, wherein the device in which the computer readable storage medium is located executes the computer program to implement the AGV cargo carrier obstacle detection method for copper pole plate transfer in harsh weather as claimed in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored computer program, wherein the device in which the computer readable storage medium is located executes the computer program to implement the AGV cargo carrier obstacle detection method for copper pole plate transfer in harsh weather as claimed in any one of claims 1 to 5.

Images