Multi-source obstacle identification method, detection device and detection system

Abstract

The invention belongs to the technical field of agriculture, and particularly relates to a multi-source obstacle identification method, a detection device and a detection system. According to the method, images are collected, an image data set is preprocessed, the preprocessed image data set is divided into a first image data set and a second image data set, the first image data set is introduced into a YOLOv10n network model for training, and the second image data set is introduced for barrier recognition rate verification; if the obstacle recognition rate is lower than the threshold value, the image data set is acquired again for model training and verification; according to the method, the obstacle information can be automatically recognized, the YOLOv10n network model is adopted to improve the obstacle feature recognition capability, and the obstacle recognition accuracy is further improved.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural technology, and in particular relates to a multi-source obstacle recognition method, a detection device and a detection system. Background Art

[0002] In the agricultural production process, due to the multi-faceted terrain, there are often many obstacles on the path of the operating machinery. These obstacles may affect the operation of automated equipment when using automated equipment in modern agriculture, such as drones, unmanned agricultural machinery and other unmanned automated equipment.

[0003] In the existing technology, there are several methods for automatic equipment to identify obstacles: manual inspection, which relies on farm staff to patrol the fields regularly or irregularly to identify and mark obstacles. This method is easily affected by human factors, such as fatigue and distraction, resulting in low recognition accuracy; static image analysis, using pre-shot field images to identify obstacles through image processing techniques such as edge detection and color segmentation. The limitation of this method is that it cannot update obstacle information in real time and has poor adaptability to dynamically changing environments; rule-based machine learning models, but after these models obtain data, due to the certain fluctuations in the actual data collection, the visual module has limited recognition effect on the complex and changeable farm environment, and the recognition degree is not high, which leads to the model trained based on recognition, which is not accurate in acquiring the actual obstacle features. Summary of the invention

[0004] Based on the above problems, the present invention proposes a multi-source obstacle warning recognition method, which acquires images through vision and improves the diversity of data through a hybrid data enhancement algorithm, so that the enhanced data can be more in line with reality. A feature focus model is used to improve the accuracy of feature acquisition, and a recognition model is trained based on YOLOv10n to improve the accuracy of feature acquisition by the model and make model recognition more in line with reality.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides a multi-source obstacle warning identification method, comprising:

[0007] Get image dataset;

[0008] Preprocessing the image data set, and dividing the preprocessed image data set into a first image data set and a second image data set;

[0009] The first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced to verify the obstacle recognition rate; if the obstacle recognition rate is lower than the threshold, the image data set is re-acquired for model training and verification.

[0010] By collecting images and preprocessing the image data set, the preprocessed image data set is divided into a first image data set and a second image data set, and the first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; if the obstacle recognition rate is lower than a threshold, the image data set is re-acquired for model training and verification; so that the method can automatically identify obstacle information, adopt the YOLOv10n network model to improve the ability to recognize obstacle features, and further improve the recognition accuracy of obstacles.

[0011] In some embodiments, the YOLOv10n network model includes:

[0012] Performing feature extraction on the first image data to obtain a first feature set;

[0013] Perform a multi-dimensional transformation on the first feature set, output a second feature set, input the second feature set into the attention model, output a third feature set; and combine the second feature set and the third feature set to output a multi-dimensional fourth feature set;

[0014] Upsampling the second feature set and performing multi-dimensional connection, and outputting the first detection result;

[0015] The first detection result includes first warning information.

[0016] Specifically, the YOLOv10n model includes feature extraction, multi-dimensional transformation, multi-dimensional feature concatenation, and output of multi-dimensional feature results. The introduction of the attention model can make the feature acquisition more focused on the required features of obstacles, thereby making the obstacle recognition more accurate.

[0017] In some implementations, the attention model includes:

[0018] Performing convolution processing on the second feature set;

[0019] The processed second feature set is input into the feature focus model, and processed by the feature focus model to obtain a fifth feature set. After convolution and splicing of the fifth feature set, a third feature set is obtained.

[0020] By setting the attention model, the attention model is used to improve the YOLOv10n network model's attention to information at specific locations.

[0021] In some embodiments, a feature-focused model includes:

[0022] Pool the features in multiple dimensions to obtain pooled features;

[0023] Concatenate the pooled features;

[0024] Through the activation function and the logistic function, the pooled features are weighted to obtain the weighted features;

[0025] Output the processed weighted features.

[0026] The feature focus model includes multi-dimensional pooling and splicing of the acquired features, processing and activating the data through activation functions and logistic functions, and realizing weighted attention on the multi-dimensional pooled features to enhance the overall function's attention to specific locations.

[0027] In some embodiments, multi-source datasets are acquired, including images and videos.

[0028] Acquire multi-source datasets, mainly vision-related datasets, to extract features from visual images.

[0029] In some implementations, the image data set is preprocessed and divided into a first image data set and a second image data set; the preprocessing includes:

[0030] De-noise, adjust contrast, and optimize size of image datasets;

[0031] Perform data augmentation on the image dataset, including image rotation, and / or image scaling, and / or image cropping, and / or image flipping, and / or image color transformation.

[0032] By performing multi-dimensional transformation on image data, the diversity of data can be improved. At the same time, it can be changed in combination with actual data to make the trained model closer to the actual image.

[0033] In some implementations, the preprocessed image dataset is divided into a first image dataset and a second image dataset:

[0034] The image data set is divided into a first image data set and a second image data set, wherein a proportion of the first image data set in the image data set is greater than a proportion of the second image data set in the image data set;

[0035] The ratio of the first image set to the second image set is 9:1.

[0036] The collected images are divided into training images and test images. The division of the data set should ensure that the types, sizes, shapes and environmental conditions of obstacles in the training set and the test set have similar distributions to avoid deviations in model training and testing.

[0037] In some implementations, the first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; if the obstacle recognition rate is lower than a threshold, the image data set is reacquired for model training and verification; including:

[0038] Perform YOLOv10n network model training using the first image dataset;

[0039] The second image data set is introduced into the YOLOv10n network model to output warning information;

[0040] Determine the accuracy of the warning information. If the accuracy is greater than the threshold, data input and identification can continue;

[0041] If the accuracy is lower than the threshold, the image data is acquired again and the model training is performed again.

[0042] After the model is trained, the test set is input for verification. The accuracy of the model judgment is used to determine whether the model can meet actual usage requirements.

[0043] In a second aspect, the present invention provides an obstacle detection device, comprising:

[0044] An image acquisition module, used to acquire an image data set;

[0045] A data preprocessing module, used to preprocess the image data set, and divide the preprocessed image data set into a first image data set and a second image data set;

[0046] The model training module is used to introduce the first image data set into the YOLOv10n network model for training, and introduce the second image data set for obstacle recognition rate verification; if the obstacle recognition rate is lower than the threshold, the image data set is re-acquired for model training and verification.

[0047] By setting up an image acquisition module, a data preprocessing module, and a model training module, the obstacle detection device can implement the multi-source obstacle warning identification method mentioned in the first aspect.

[0048] In a third aspect, the present invention provides an obstacle detection system, including a processor, a memory, a communication interface, a communication bus and a collection device, wherein the processor and the memory are integrated in a system chip, and the system chip and the communication interface communicate with each other through the communication bus;

[0049] The acquisition device includes a visual module. The system chip is provided with multiple different system ports. The visual module is arranged at the communication interface and connected to the system port through a communication bus. A switching chip is arranged between the visual module and the system port. The visual module is connected to different system ports through the switching chip.

[0050] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute the operation of the multi-source obstacle warning identification method as described in the first aspect.

[0051] The visual system is used to implement the multi-source obstacle warning identification method proposed in the first aspect.

[0052] The multi-source obstacle identification method, detection device and detection system of the present invention have the following beneficial effects:

[0053] By collecting images and preprocessing the image data set, the preprocessed image data set is divided into a first image data set and a second image data set, and the first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; if the obstacle recognition rate is lower than a threshold, the image data set is re-acquired for model training and verification; so that the method can automatically identify obstacle information, and at the same time, a feature focus model is set in the YOLOv10n network model to improve the ability to recognize obstacle features and further improve the recognition accuracy of obstacles. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are only used to illustrate the embodiments and are not to be considered as limiting the present invention. In addition, the same reference symbols are used to represent the same components throughout the accompanying drawings. In the accompanying drawings:

[0055] Figure 1 It is a flow chart of the multi-source obstacle identification method of the present invention;

[0056] Figure 2 It is a flow chart of the multi-source obstacle identification method of the present invention;

[0057] Figure 3 It is a schematic diagram of the YOLOv10n network model of the multi-source obstacle identification method of the present invention;

[0058] Figure 4 It is a YOLOv10n network model framework flow chart of the multi-source obstacle identification method of the present invention;

[0059] Figure 5 It is a principle diagram of the attention model of the multi-source obstacle recognition method of the present invention;

[0060] Figure 6 This is a flowchart of the attention model framework of the multi-source obstacle recognition method of the present invention;

[0061] Figure 7 It is a feature-focused model framework flow chart of the multi-source obstacle identification method of the present invention;

[0062] Figure 8 A flow chart for data enhancement of an image data set for a multi-source obstacle recognition method of the present invention;

[0063] Fig. 9 It is a framework diagram of a detection device of a multi-source obstacle identification method of the present invention;

[0064] Fig.10 It is a framework diagram of the detection system of the multi-source obstacle identification method of the present invention.

[0065] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0066] Embodiment 1:

[0067] like Figures 1 to 8 As shown, this embodiment proposes a multi-source obstacle warning identification method, comprising the following steps:

[0068] Step 101: Obtain an image dataset;

[0069] Specifically, in the first step of the present invention, data collection is a basic and critical link. Data collection involves collecting images and video data of various obstacles in the farm environment that may affect crop growth and agricultural machinery operations. These data may include, but are not limited to, telephone poles, people, stones, animals, abandoned farm tools, etc. Data collection can be performed automatically by high-definition cameras installed in the field, or it can be collected regularly by manually carrying equipment. The collected data needs to be representative to cover different obstacle types, sizes, shapes and environmental conditions. The beneficial effect of this step is that through data collection, the model can be more targeted during training, and the detection of obstacles will be closer to the actual scene.

[0070] Step 102: preprocessing the image data set, and dividing the preprocessed image data set into a first image data set and a second image data set;

[0071] Specifically, in this patent application, data preprocessing technology is given a crucial role. It is a key step to improve data quality and ensure the learning efficiency of machine learning models. Its fundamental purpose is to optimize the data such as noise reduction to improve the accuracy of feature capture. At the same time, it also includes changing the data according to the actual situation, such as changing the image with equal exposure rate according to the local exposure rate, so as to obtain the image through relatively few images and match the actual use environment. Further, data processing can also include screening of data, such as screening some invalid feature images and some non-repetitive feature images, to avoid the increase of invalid images causing low data diversity during training. Further, preprocessing can also include resolution and scale transformation of data to adapt to the type of required image data set inside the YOLOv10n network model, to prevent the input data from not matching the data required by the model, resulting in the training model not matching the actual one. And the data is divided into a first image data set and a second image data set to perform data training and data testing respectively, and in order to ensure the consistency of the data, they are all collected in similar scenes to avoid deviations in model training and testing.

[0072] Step 103: introduce the first image data set into the YOLOv10n network model for training, and introduce the second image data set for obstacle recognition rate verification; if the obstacle recognition rate is lower than the threshold, reacquire the image data set for model training and verification.

[0073] Specifically, the first image data set is introduced into the YOLOv10n network model for training, and the first image data set is subjected to feature extraction and feature preservation by the YOLOv10n network model, and common obstacle recognition training is performed according to the first image data set. After the training is completed, the obstacle recognition accuracy is judged by the second image data set to prevent excessive errors in actual use. If the obstacle judgment accuracy is too low, further processing is performed according to the error rate. If the obstacle recognition error rate is too high, the YOLOv10n network model can be reset, and the image data set is re-collected and steps 101 to 103 are re-executed, and verification is performed again until the error rate is lower than the threshold. If the obstacle recognition error rate is slightly higher than the threshold, data training can be further performed on the original YOLOv10n network model, and retested until the error rate threshold is lower than the error rate. If the obstacle recognition error rate is lower than the threshold, the YOLOv10n network model can be uploaded to the interrupt.

[0074] By collecting images and preprocessing the image data set, the preprocessed image data set is divided into a first image data set and a second image data set, and the first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; if the obstacle recognition rate is lower than a threshold, the image data set is re-acquired for model training and verification; so that the method can automatically identify obstacle information, adopt the YOLOv10n network model to improve the ability to recognize obstacle features, and further improve the recognition accuracy of obstacles.

[0075] In some embodiments, the YOLOv10n network model includes:

[0076] Step 201: extracting features from first image data to obtain a first feature set;

[0077] Specifically, through multiple convolutional layers, the Conv convolutional layer is a basic building block in CNN, which is used to extract data features of input data, and through the C2F operation layer, the first feature set is divided into multiple layers. The beneficial effect of this step is that the first feature set can be extracted from the first image data set, and the first feature set is subjected to feature conversion so that the features in the first feature set remain consistent, such as consistent resolution.

[0078] Step 202: Perform a multi-dimensional transformation on the first feature set, output a second feature set, input the second feature set into the attention model, output a third feature set; and combine the second feature set and the third feature set to output a multi-dimensional fourth feature set;

[0079] Specifically, the spatial dimension of the first feature set is reduced by the spatial downsampling layer SCDown, and the first feature set is further divided into multiple spatial dimension layers by the C2F operation layer and the spatial downsampling layer SCDown, so that the first feature set is subjected to multi-dimensional transformation. Further, SPPF spatial pyramid pooling is also provided to capture the multi-scale spatial information of the image, and the second feature set is output by SPPF spatial pyramid pooling; preferably, the attention model is input through the second feature set, and the third feature set is output through the attention model to improve the selection accuracy of the third feature set for a specific position. Optionally, the attention model may not be provided, although the computing power can be reduced, but the overall accuracy is lower than the YOLOv10n network model provided with the attention model. And finally the second feature set and the third feature set are combined into the fourth feature set for data fusion. The beneficial effect of this step is that multi-scale transformation is performed, the representation ability of the feature is improved, the adaptability and flexibility of the model are improved, and the performance is better optimized. At the same time, the attention model is provided, so that the positioning is more accurate when locating the feature position.

[0080] Step 203: upsampling the second feature set and performing multi-dimensional connection, and outputting the first detection result;

[0081] Specifically, the data dimension is increased through the concat connection module, the C2F operation layer, and the upsampling Upsampling module. The concat connection module is used in the neural network to splice multiple tensors or feature maps along a certain dimension. The upsampling Upsampling module is used to increase the spatial dimension of the feature map, and is usually used in a generation model or a segmentation model. After using the upsampling Upsampling module to increase the dimension, the Conv convolution layer, the concat connection module, the C2F operation layer, and the spatial downsampling SCDown are further combined to reduce the dimension transformation, and the first detection result is output. The first detection result is output by v10Detect to achieve target detection. The beneficial effect of this step is that through the transformation of the dimension, the noise of the image can be suppressed and the balance of feature extraction can be achieved. At the same time, in the deep learning model, features can be processed at different scales.

[0082] The first detection result includes first warning information.

[0083] The beneficial effect of this method is that the YOLOv10n network model is a target detection algorithm model. In this embodiment, YOLOv10n is used, which is an improved model of the YOLO model. The YOLOv10n network model includes feature extraction, multi-dimensional transformation, multi-dimensional feature splicing, and outputs multi-dimensional feature results. The introduction of the attention model can make the feature acquisition more focused on the required features of the obstacle, thereby making the obstacle identification more accurate.

[0084] In some embodiments, the attention model includes:

[0085] Among them, the attention model can be expressed as PSCoordAtt;

[0086] Step 301: performing convolution processing on the second feature set;

[0087] Specifically, convolution processing is performed on the second feature set to further extract features from the second feature set.

[0088] Step 302: Input the processed second feature set into the feature focus model, and process it with the feature focus model to obtain a fifth feature set. After convolution and splicing of the fifth feature set, a third feature set is obtained.

[0089] Specifically, the processed second feature set is input into the feature focus model, and the model's focus on features is further improved through the feature focus model, and the fifth feature set is obtained. After convolution and splicing of the fifth feature set, it is output as the third feature set for judgment by the YOLOv10n network model.

[0090] The beneficial effect of this method is that by setting the attention model, the attention model is used to improve the YOLOv10n network model's attention to information at a specific location.

[0091] In some embodiments, the feature-focused model may be represented as CoordAtt:, which includes:

[0092] Step 401: Pooling the features in multiple dimensions to obtain pooled features;

[0093] Specifically, after initializing the feature focus model and the attention model, the input second feature set features are multi-dimensionally pooled, including horizontal pooling and vertical pooling. The beneficial effect of this method is that it can improve the analysis efficiency of the second feature set, improve the utilization and value of data, and at the same time improve the diversity of data.

[0094] Step 402: concatenate the pooled features;

[0095] Specifically, the pooled features are spliced, and the second feature set splicing can be performed using the concat connection module as described above.

[0096] Step 403: weighting the pooled features through activation functions and logistic functions to obtain weighted features;

[0097] Specifically, through the activation function, that is, the h_swish function, which itself is a type of Swish activation function. This function can enhance the nonlinear expression ability of the model, so that the model can capture the distribution of the more complex second feature set, so as to more accurately capture the effective feature bits and improve the ability to capture and judge the feature bits. Furthermore, the logistic function, that is, h_sigmoid, is also used: this function is an optimized version of the sigmoid function. Through this function, in the binary classification problem, any numerical value can be mapped to the interval of (0,1), and the weight of the feature points can be more effectively increased, so that the model can focus more on the feature points. Furthermore, the horizontal pooling and vertical pooling can be weighted separately to make the weight distribution more reasonable, and thus make the obtained feature points closer to the required obstacle feature points, so that the judgment is more accurate.

[0098] Among them, the h_swish function can be specifically expressed as:

[0099]

[0100] The logistic function, also known as h_sigmoid, can be specifically expressed as:

[0101]

[0102] Step 404: Output the processed weighted features.

[0103] The beneficial effects of this method are: the feature focus model includes multi-dimensional pooling and splicing of the acquired features, processing and activating the data through activation functions and logistic functions, and realizing weighted attention on the multi-dimensional pooled features to enhance the overall function's attention to specific locations.

[0104] In some embodiments, multi-source datasets are acquired, including images and videos.

[0105] Specifically, a multi-source dataset is obtained, including visual images and visual videos.

[0106] Acquire multi-source datasets, mainly vision-related datasets, to extract features from visual images.

[0107] In some embodiments, the image data set is preprocessed and divided into a first image data set and a second image data set; the preprocessing includes:

[0108] De-noise, adjust contrast, and optimize size of image datasets;

[0109] Perform data augmentation on the image dataset, including image rotation, and / or image scaling, and / or image cropping, and / or image flipping, and / or image color transformation.

[0110] Specifically, the image data set is subjected to denoising, contrast adjustment, and size optimization; for example, the image data set is subjected to denoising, contrast adjustment, and size optimization by an image processor. This operation is a conventional image optimization operation after image acquisition, and its purpose is to enable the image to be used normally. This operation is a conventional image processing operation, and therefore no further examples will be given for discussion.

[0111] Data enhancement is performed on the image dataset, including image rotation, and / or image scaling, and / or image cropping, and / or image flipping, and / or image color transformation. Specifically, it may include image rotation (Image Rotation), scaling (Scaling), cropping (Cropping), flipping (Flipping), and color transformation (Color Transformation). For example, the imgaug library can be used to implement a flexible and scalable data enhancement method to improve the robustness and accuracy of the model in a changing environment. The model's adaptability to lighting changes is enhanced by simulating different lighting conditions through brightness adjustment (iaa.Multiply((0.8,1.2))). The code first reads the image and its corresponding label file, and then parses and converts the coordinate information in the label file into a BoundingBox object. Then, the BoundingBoxesOnImage object is used to associate these bounding boxes with the image, and a defined enhancement sequence is applied. The enhanced image and bounding box information are saved, where the bounding box coordinates are converted to a normalized format to meet the needs of the YOLOv10n target detection model. The whole process is automated, which effectively improves the generalization ability and performance of the model in practical applications. Combined with the application of the YOLOv10n network model, the present invention not only improves the accuracy of obstacle recognition, but also greatly improves the efficiency and robustness of the entire system, bringing substantial breakthroughs to the field of modern agricultural technology.

[0112] These technologies can not only enrich the data set, but also simulate the diversity of real farm environments, so as to train a more robust model. Image rotation technology is used to rotate the image at different angles to simulate the perspective of obstacles in different directions, thereby enhancing the model's adaptability to directional changes. Scaling technology simulates the observation effect of obstacles at different distances by changing the size of the image, thereby improving the model's ability to recognize scale changes. Cropping technology simulates the recognition problem of local obstacles by selecting different parts of the image, thereby improving the model's ability to learn local features. Flipping technology increases the model's ability to recognize symmetry changes by flipping the image horizontally or vertically. Color transformation technology adjusts the color attributes of the image to simulate different lighting conditions, thereby improving the model's robustness to color changes. Through these advanced data enhancement technologies, a richer and more diverse data set can be constructed. The beneficial effect of this method is that it can increase the versatility, diversity, and robustness of the data, making the collected data more suitable for actual use, and making the feature points collected by the trained model closer to actual application.

[0113] In some embodiments, the preprocessed image dataset is divided into a first image dataset and a second image dataset:

[0114] The image data set is divided into a first image data set and a second image data set, wherein a proportion of the first image data set in the image data set is greater than a proportion of the second image data set in the image data set;

[0115] The ratio of the first image set to the second image set is 9:1.

[0116] Specifically, after data collection, data partitioning is usually required. Data set partitioning is to divide the preprocessed and enhanced data set into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the performance of the model. Usually, the data set is divided into 90% training set and 10% test set, but this ratio can be adjusted according to actual conditions. The partitioning of the data set should ensure that the types, sizes, shapes, and environmental conditions of obstacles in the training set and the test set have similar distributions to avoid deviations in model training and testing.

[0117] The beneficial effect of this method is that the collected images are divided into training images and test images. The division of the data set should ensure that the types, sizes, shapes and environmental conditions of obstacles in the training set and the test set have similar distributions to avoid deviations in model training and testing.

[0118] In some embodiments, the first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; if the obstacle recognition rate is lower than a threshold, the image data set is re-acquired for model training and verification; including:

[0119] Perform YOLOv10n network model training using the first image dataset;

[0120] The second image data set is introduced into the YOLOv10n network model to output warning information;

[0121] Determine the accuracy of the warning information. If the accuracy is greater than the threshold, data input and identification can continue;

[0122] If the accuracy is lower than the threshold, the image data is acquired again and the model training is performed again.

[0123] Specifically, in the model training stage, the improved YOLOv10n network model is first trained using the training set. The improved YOLOv10n network is a deep learning model specifically designed for real-time object detection. The model training includes the optimization of the loss function, where the category classification loss function is used to ensure that the model can correctly identify the category of obstacles, and the bounding box regression loss function is used to accurately predict the location of obstacles. In addition, the attention mechanism is introduced to improve the model's ability to recognize obstacle features by focusing on key areas in the image.

[0124] After the model training is completed, the test set is used to verify the model. The obstacle recognition results of the test set will be used to evaluate the performance indicators such as the accuracy, recall rate and F1 score of the model. If the model performs poorly on the test set, it may be necessary to return to the data preprocessing or model training stage for adjustment.

[0125] After the model is trained, the test set is input for verification. The accuracy of the model judgment is used to determine whether the model can meet actual usage requirements.

[0126] Embodiment 2:

[0127] like Fig. 9 As shown, this embodiment is based on Embodiment 1 and proposes an obstacle detection device 500, including:

[0128] The image acquisition module 501 is used to acquire an image data set; it is used to execute step 101, and collect an image data set through the image acquisition module to perform YOLOv10n network model training and verification.

[0129] The data preprocessing module 502 is used to preprocess the image data set and divide the preprocessed image data set into a first image data set and a second image data set; it is used to execute step 102, and preprocess and divide the image data set of the image acquisition module acquisition end through the data preprocessing module to perform YOLOv10n network model training and verification.

[0130] The model training module 503 is used to introduce the first image data set into the YOLOv10n network model for training, and introduce the second image data set for obstacle recognition rate verification; if the obstacle recognition rate is lower than the threshold, the image data set is re-acquired for model training and verification. It is used to execute step 103, after the image is acquired by the image acquisition module and the data is preprocessed by the data preprocessing module, the YOLOv10n network model is trained and verified by the model training module, so that the YOLOv10n network model can recognize obstacles.

[0131] By setting up an image acquisition module, a data preprocessing module, and a model training module, the obstacle detection device can implement the multi-source obstacle warning identification method mentioned in the first aspect.

[0132] Embodiment 3:

[0133] like Fig.10 As shown, an obstacle detection system 600 includes a processor 601, a memory 602, a communication interface 603, a communication bus 605 and a collection device. The processor 601 and the memory 602 are integrated in a system chip, and the system chip and the communication interface communicate with each other through the communication bus 605;

[0134] The memory 602 is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of the multi-source obstacle warning identification method of the first aspect.

[0135] Specifically, the obstacle detection system 600 may include: a processor 601 (processor), a communication interface 603 (Communications Interface), a memory 602 (memory), and a communication bus 605

[0136] The memory 602 is used to store the program 604. The memory may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0137] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system or other device. In addition, the embodiments of the present invention are not directed to any particular programming language.

[0138] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. Similarly, in order to simplify the present invention and help understand one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, the various features of the embodiments of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. Wherein, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, wherein each claim itself is a separate embodiment of the present invention.

[0139] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and further may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive.

[0140] It should be noted that the above embodiments illustrate the present invention rather than limit it, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising a number of different elements and by means of a suitably programmed computer. In a unit claim enumerating a number of devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be understood as limitations on the order of execution.

Claims

1. A multi-source obstacle warning identification method, characterized in that: include: Get image dataset; Preprocessing the image data set, and dividing the preprocessed image data set into a first image data set and a second image data set; The first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; If the obstacle recognition rate is lower than the threshold, the image data set is reacquired for model training and verification.

2. The multi-source obstacle warning identification method according to claim 1, characterized in that: The YOLOv10n network model includes: Performing feature extraction on the first image data to obtain the first feature set; Performing a multi-dimensional transformation on the first feature set, outputting a second feature set, and inputting the second feature set into the attention model, outputting a third feature set; and combining the second feature set and the third feature set to output the multi-dimensional fourth feature set; Upsampling the second feature set and performing multi-dimensional connection, and outputting a first detection result; The first detection result includes first warning information.

3. The multi-source obstacle warning identification method according to claim 2, characterized in that: The attention model includes: Performing convolution processing on the second feature set; The processed second feature set is input into a feature focus model, and processed by the feature focus model to obtain a fifth feature set, and the fifth feature set is convolved and spliced ​​to obtain the third feature set.

4. The multi-source obstacle warning identification method according to claim 3, characterized in that: The feature-focused model includes: Pool the features in multiple dimensions to obtain pooled features; Splicing the pooled features; The pooling features are weighted by an activation function and a logistic function to obtain weighted features; The processed weighted features are output.

5. The multi-source obstacle warning identification method according to claim 1, characterized in that: The obtaining of the image data set comprises: Obtain multi-source datasets, including images and videos.

6. The multi-source obstacle warning identification method according to claim 1, characterized in that: The image data set is preprocessed and divided into a first image data set and a second image data set; the preprocessing includes: De-noising, contrast adjustment, and size optimization are performed on the image dataset; The image dataset is subjected to data enhancement, including image rotation, and / or image scaling, and / or image cropping, and / or image flipping, and / or image color transformation.

7. The multi-source obstacle warning identification method according to claim 6, characterized in that: The step of dividing the preprocessed image data set into a first image data set and a second image data set comprises: dividing the image data set into a first image data set and a second image data set, wherein the proportion of the first image data set in the image data set is greater than the proportion of the second image data set in the image data set; The ratio of the first image set to the second image set is 9:

1.

8. The multi-source obstacle warning identification method according to claim 1, characterized in that: The first image data set is introduced into the YOLOv10n network model for training, and the second image data set is introduced for obstacle recognition rate verification; If the obstacle recognition rate is lower than the threshold, the image data set is re-acquired for model training and verification; including: Perform YOLOv10n network model training using the first image dataset; Introducing the second image data set into the YOLOv10n network model, and outputting warning information; Determine the accuracy of the warning information. If the accuracy is greater than a threshold, continue to input data and identify; If the accuracy is lower than the threshold, the image data is reacquired and the model training is performed again.

9. An obstacle detection device, characterized in that: include: An image acquisition module, used to acquire an image data set; A data preprocessing module, used to preprocess the image data set, and divide the preprocessed image data set into a first image data set and a second image data set; A model training module, used for introducing the first image data set into the YOLOv10n network model for training, and introducing the second image data set for obstacle recognition rate verification; If the obstacle recognition rate is lower than the threshold, the image data set is reacquired for model training and verification.

10. An obstacle detection system, characterized in that: It includes a processor, a memory, a communication interface, a communication bus and a collection device, wherein the processor and the memory are integrated in a system chip, and the system chip and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation of the multi-source obstacle warning identification method as described in any one of claims 1-8.

Images