Obstacle detection method, electronic equipment and storage medium

Abstract

The invention relates to an obstacle detection method, which is applied to rail transit and used for detecting obstacles, and comprises the following steps: establishing a main network based on Officientnet-v2; according to the main network of the Officientnet-v2, extracting an obstacle feature; according to the obstacle features, acquiring a block output result; and obtaining a detection result of the obstacle according to the block output result. According to the obstacle detection method provided by the invention, complex obstacles in rail transit can be detected through the detection method, and the safety of train running is ensured.

Description

Technical Field

[0001] The present application relates to the field of rail transit technology, and in particular to an obstacle detection method, electronic equipment and storage medium. Background Art

[0002] High-speed rail transit has shifted from the large-scale "design and construction" stage to the long-term "operation and maintenance" stage. In addition, high-speed rail has the characteristics of short line construction cycle and short opening and operation time. The deterioration of the infrastructure on the operating line has entered the apparent period. In an extremely complex operating environment, structural damage in some sections, rockfall on the roadbed, and operational garbage can easily pose risks to the normal operation of trains or threaten the safety of personnel. Therefore, efficient and real-time obstacle detection technology has become a necessary support for the safe operation of trains in rail transit.

[0003] The rail transit environment is complex and there are many obstacles. With the rapid movement of the camera on the train, the captured image will be blurred due to the bumps and motion of the train, and the difficulty of identifying obstacles will be greatly increased. Obstacles that are easily encountered in a complex operating environment include: broken tracks, falling rocks, operating garbage and other obstacles. If the obstacles ahead cannot be correctly judged and handled, it is easy to cause rollover, train suspension, casualties and other risks and serious consequences.

[0004] With the rapid development of machine learning target detection methods, more and more researchers have applied target detection algorithms in real scenarios. However, at present, most of the applications of deep learning-based target detection algorithms are limited to structured track applications. In the application of unmanned obstacle detection methods in open-pit mines, traditional methods are mainly based on the following two aspects:

[0005] First, use multi-lidar and combined features to fuse the 64-line radar and 32-line radar point clouds to achieve better negative obstacle detection results. Map the position of the target to the radar coordinates to achieve perception and recognition of the target.

[0006] Second, we need to accurately identify obstacles above the ground, such as fallen rocks, on the running track. Based on the MobileNet backbone feature extraction network, we propose a method for detecting negative obstacles on open-pit mine roads with multi-scale feature fusion, and use computer vision technology to quickly and effectively identify negative obstacles on the running track. However, the above radar-based obstacle detection method has difficulty in realizing real-time data transmission for obstacle recognition, and it is also unable to identify the object category well. The visual-based obstacle detection method cannot complete the depth measurement of target detection, and cannot allow the train to accurately perform obstacle avoidance behavior.

[0007] Therefore, this field is in urgent need of a new solution to solve the above technical problems. Summary of the invention

[0008] The purpose of the present application is to provide an obstacle detection method, electronic device and storage medium, through which complex obstacles in rail transit can be detected to ensure the safety of train travel.

[0009] In a first aspect, an embodiment of the present application provides an obstacle detection method, which is applied to rail transit for detecting obstacles, and the detection method includes:

[0010] Establish the main network based on Efficientnet-v2;

[0011] According to the main network of Efficientnet-v2, obstacle features are extracted;

[0012] According to the obstacle characteristics, obtain the block output result;

[0013] Obtain obstacle detection results based on the block output results.

[0014] In a possible implementation, extracting obstacle features according to the main network of Efficientnet-v2 includes:

[0015] Based on the main network of Efficientnet-v2, a Feffol multi-target detection model is established;

[0016] According to the Feffol multi-target detection model, the obstacle features are extracted.

[0017] In a possible implementation, the obstacle features include at least one or a combination of road collapse, potholes, water accumulation, and falling rocks.

[0018] In a possible implementation, extracting the obstacle feature according to the Feffol multi-target detection model includes:

[0019] According to the Feffol multi-target detection model, a loss function of obstacles is established;

[0020] The obstacle feature is obtained according to the loss function of the obstacle.

[0021] In a possible implementation manner, the loss function includes a classification loss function, and acquiring the obstacle feature according to the loss function of the obstacle includes:

[0022] Establish the cross entropy function;

[0023] According to the cross entropy function, multiple types of obstacle features are obtained.

[0024] In a possible implementation, the cross entropy function expression is:

[0025]

[0026] Among them, y is the value of whether the obstacle is the foreground, and p is the probability that the obstacle is the foreground.

[0027] In a possible implementation, the loss function includes a border loss function, and obtaining the obstacle feature according to the loss function of the obstacle includes:

[0028] Establish a bounding box loss function;

[0029] Obstacle features of the predicted frame and the real frame are obtained according to the frame loss function.

[0030] In a possible implementation, the border loss function is expressed as:

[0031]

[0032] Among them, b is the center point of the predicted box, bgt is the center point of the real box, ρ is the shortest distance between the two center points, c is the diagonal distance of the minimum closure area that can contain both the predicted box and the real box, α is the weight function, and v is the similarity function that measures the aspect ratio.

[0033] In a second aspect, an embodiment of the present application provides an electronic device, including:

[0034] Memory for storing computer programs;

[0035] A processor is used to implement the above obstacle detection method when executing the computer program.

[0036] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned obstacle detection method are implemented.

[0037] In an embodiment of the present application, a method for detecting obstacles is proposed, which provides a deep detection function through obstacle recognition technology. In the feature extraction stage, the method uses Efficientnet as the backbone network to extract rich feature information. By carefully identifying the different obstacle features extracted by the backbone network, the method can more effectively extract key features, thereby significantly improving the accuracy of the prediction. In addition, the method also provides an accurate and fast obstacle detection technology that can provide reliable and real-time decision-making data support for the obstacle avoidance system. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In addition, in the drawings, the same components are marked with the same reference numerals, and the drawings are not drawn according to the actual scale.

[0039] Figure 1 A flow chart showing the obstacle detection method provided by the present application;

[0040] Figure 2 The following is a flow chart showing the obstacle detection method provided by the embodiment of the present application. Figure 1 ;

[0041] Figure 3 The following is a flow chart showing the obstacle detection method provided by the embodiment of the present application. Figure 2 ;

[0042] Figure 4 The following is a flow chart showing the obstacle detection method provided by the embodiment of the present application. Figure 3 . DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0044] At the current stage of technological development, most object detection algorithms that rely on deep learning technology are mainly applied to scenarios in structured track environments. Especially in unmanned driving systems in open-pit mines, the research and application of obstacle detection methods still face many challenges. Traditional detection methods mainly focus on two core aspects:

[0045] First, the researchers worked on improving the performance of obstacle detection by fusing different types of lidar data. Specifically, they used a multi-lidar system to fuse the point cloud data generated by a 64-line lidar and a 32-line lidar. The purpose of this approach is to achieve better negative obstacle detection results. In this way, the researchers were able to map the target identification and position into the radar coordinate system, thereby achieving effective perception and identification of the target. This fusion strategy has improved the accuracy and reliability of detection to a certain extent.

[0046] Secondly, in order to accurately identify obstacles above the ground, such as fallen rocks on the running track, the researchers proposed a multi-scale feature fusion method for detecting negative obstacles on open-pit mine roads based on the lightweight backbone feature extraction network MobileNet. This method uses computer vision technology to quickly and effectively identify negative obstacles on the running track. In this way, the speed and accuracy of obstacle detection can be significantly improved, thereby providing more reliable environmental perception capabilities for unmanned driving systems.

[0047] However, the above radar-based obstacle detection method also has some limitations in practical applications. Due to the difficulties in real-time transmission and processing of radar data, the real-time identification of obstacles becomes complicated. In addition, the radar system also has certain deficiencies in the recognition of object categories, which limits its accuracy in performing obstacle avoidance behaviors in unmanned driving systems. On the other hand, although the vision-based obstacle detection method performs well in target detection, its ability in depth measurement is limited, which makes it impossible for the train to accurately judge the distance and position of obstacles when performing obstacle avoidance behaviors, thereby affecting the accuracy and safety of obstacle avoidance. Therefore, how to combine the advantages of radar and vision technology to develop a more efficient and accurate obstacle detection method is still an important topic in the current research field.

[0048] The core purpose of this application is to effectively solve at least one aspect of the above technical problems, so a method for detecting obstacles, corresponding electronic equipment and storage medium are proposed. By adopting this detection method, it is possible to accurately detect and identify complex and changeable obstacles in the rail transit system, thereby greatly improving the safety of the train during travel and ensuring the safety of passengers and trains. Specifically, this method can monitor the environment around the track in real time, and quickly and accurately identify various potential obstacles through advanced sensor technology and image processing algorithms. In this way, the train control system can take corresponding measures in time, such as slowing down or stopping, to avoid possible collision accidents. In addition, the method can also improve the accuracy and reliability of its detection through continuous learning and optimization, and further improve the overall safety performance of the rail transit system. Therefore, the detection method, electronic equipment and storage medium proposed in this application can not only effectively deal with the obstacle detection problems encountered in the current rail transit system, but also provide a solid foundation for future technological development.

[0049] Specifically, the first aspect is Figure 1 As shown, an embodiment of the present application provides an obstacle detection method, which is applied to rail transit for detecting obstacles, and the detection method includes:

[0050] S100: Establish the main network based on Efficientnet-v2;

[0051] S200: extracting obstacle features according to the main network of Efficientnet-v2;

[0052] S300: Obtaining a block output result according to the obstacle feature;

[0053] S400: Obtain obstacle detection results according to the block output results.

[0054] In terms of specific implementation, the embodiments of the present application are intended to provide an obstacle detection method for a rail transit system. The main purpose of the method is to ensure the safety of train operation by efficiently and accurately identifying and locating obstacles on the track. The specific steps are as follows:

[0055] First, this method builds a main network model based on the Efficientnet-v2 architecture. Efficientnet-v2 is an advanced deep learning model known for its efficient and accurate image recognition capabilities. By adopting this model as the main network, orbital image data can be effectively processed and analyzed.

[0056] Next, based on this Efficientnet-v2 main network, this method further extracts feature information related to obstacles. This feature information includes but is not limited to the shape, size, color, and position of the obstacle in the image. By extracting these features, potential obstacles on the track can be identified more accurately.

[0057] Subsequently, based on the extracted obstacle features, this method obtains the block output results. Here, "block" refers to one or more convolutional layers or pooling layers in the Efficientnet-v2 network, which play a key role in the feature extraction process. By analyzing the output results of these blocks, the feature representation of obstacles can be further refined, thereby improving the accuracy of detection.

[0058] Finally, based on the output results of these blocks, this method finally obtains the obstacle detection results. The detection results usually include information such as the location coordinates, category, and possible motion state of the obstacle. This information is crucial for the safe operation of the rail transit system and can provide a reliable basis for subsequent decision-making and response measures.

[0059] In summary, the embodiments of the present application provide an efficient and accurate rail transit obstacle detection method by establishing a main network based on Efficientnet-v2, extracting obstacle features, obtaining block output results, and finally obtaining obstacle detection results.

[0060] In one possible implementation, Figure 2 As shown, the S200: extracting obstacle features according to the main network of Efficientnet-v2 includes:

[0061] S210: Establish a Feffol multi-target detection model based on the main network of Efficientnet-v2;

[0062] S220: Extracting the obstacle features according to the Feffol multi-target detection model.

[0063] Specifically, by utilizing the main network structure of Efficientnet-v2, the features of obstacles in the scene can be effectively extracted. Specifically, this process includes the following steps:

[0064] First, based on the main network architecture of Efficientnet-v2, a multi-target detection model named Feffol was built. This model inherits the powerful capabilities of Efficientnet-v2 in image processing, especially in feature extraction and information fusion. In this way, the Feffol model can make full use of the pre-trained weights and structural advantages of Efficientnet-v2, thereby showing higher accuracy and efficiency in multi-target detection tasks.

[0065] Next, the constructed Feffol multi-target detection model is used to further extract obstacle features in the scene. These features include information such as the shape, size, position, and relative relationship of the obstacle to other objects. Through the efficient feature extraction capability of the Feffol model, various obstacles in the scene can be accurately identified and located, thus providing reliable data support for subsequent processing and decision-making.

[0066] In summary, by combining the main network of Efficientnet-v2 and the Feffol multi-target detection model, obstacle features can be effectively extracted, providing strong support for the safe navigation and decision-making of intelligent systems in complex environments.

[0067] In a possible implementation, the obstacle features include at least one or a combination of road collapse, potholes, water accumulation, and falling rocks.

[0068] In a specific example, the obstacle features include at least one or a combination of road collapse, potholes, water accumulation, and rockfall. Specifically, these obstacle features may appear alone or in combination. For example, road collapse may refer to a large area of ​​depression on the road surface, which prevents vehicles from passing normally; potholes may refer to smaller depressions or protrusions on the road surface, which affect the driving stability of the vehicle; water accumulation refers to a large area of ​​water accumulation on the road, which may cause the vehicle to slip or the engine to get water; rockfall refers to stones rolling down from the mountain on the road, which may cause damage to the vehicle. The combination of these obstacle features can more comprehensively describe the various obstacles that may be encountered on the road, thereby providing more accurate information for subsequent processing and response.

[0069] In one possible implementation, Figure 3 The step S220: extracting the obstacle features according to the Feffol multi-target detection model includes:

[0070] S221: Establishing a loss function of obstacles according to the Feffol multi-target detection model;

[0071] S222: Acquire the obstacle feature according to the loss function of the obstacle.

[0072] In a specific example, by using the Feffol multi-target detection model, the features of obstacles can be effectively extracted. Specifically, this process includes the following steps:

[0073] First, we need to build a loss function specifically for obstacles based on the Feffol multi-target detection model. This loss function is used to measure the accuracy of the model in detecting obstacles and the error size in predicting the location and category of obstacles. The designed loss function can ensure that the model can better focus on the characteristics of obstacles during training.

[0074] Next, based on the obstacle loss function that has been constructed, the model parameters can be optimized through the back propagation algorithm. In this process, the model will continuously adjust its internal parameters to minimize the value of the loss function. In this way, the model can gradually learn how to extract features related to obstacles from the input image data.

[0075] Finally, through the above steps, the features of obstacles can be successfully obtained. These features include not only the location information of obstacles, but also their attributes such as shape, size, and category. These rich feature information are crucial for subsequent obstacle detection and classification tasks, and can significantly improve the overall performance and accuracy of the system.

[0076] In summary, in this possible implementation, by establishing the loss function of the obstacle and using it to optimize the model parameters, the features of the obstacle are effectively extracted. This process not only improves the accuracy of the model in detecting obstacles, but also provides rich feature information for subsequent tasks.

[0077] In one possible implementation, Figure 4 As shown, the loss function includes a classification loss function, and the step S222: obtaining the obstacle feature according to the loss function of the obstacle includes:

[0078] S222-1: Establish cross entropy function;

[0079] S222-2: Obtain multiple types of obstacle features according to the cross entropy function.

[0080] In a possible implementation, the cross entropy function expression is:

[0081]

[0082] Among them, y is the value of whether the obstacle is the foreground, and p is the probability that the obstacle is the foreground.

[0083] In a specific example, the loss function includes not only a classification loss function, but also other types of loss functions to ensure that the model can learn and predict more accurately. In the process of obtaining the obstacle features, a cross entropy function needs to be established first. The cross entropy function is a commonly used loss function that measures the difference between the probability distribution predicted by the model and the probability distribution of the true label. By optimizing the cross entropy function, the classification accuracy of the model for obstacles can be effectively improved.

[0084] Specifically, according to the cross entropy function, multiple types of obstacle features can be obtained. The expression of the cross entropy function can be expressed as:

[0085]

[0086] Among them, y represents the value of whether the obstacle is the foreground, and the value range is -1 or 1. When the obstacle is the foreground, y takes the value of 1; when the obstacle is the background, y takes the value of -1. p represents the probability of the obstacle being the foreground, and the value range is between 0 and 1. By maximizing the likelihood function, the minimum cross entropy loss can be obtained, so that the model can more accurately predict whether the obstacle is the foreground.

[0087] In practical applications, by continuously iterating and optimizing the cross entropy function, the model's ability to identify obstacle features can be gradually improved. This not only helps to improve the accuracy of classification, but also reduces false positives and false negatives to a certain extent, thereby improving the overall system performance. In this way, obstacle features can be effectively obtained, providing reliable data support for subsequent processing and decision-making.

[0088] In one possible implementation, Figure 4 As shown, the loss function includes a border loss function, and the step S222: obtaining the obstacle feature according to the loss function of the obstacle includes:

[0089] S222-3: Establish border loss function;

[0090] S222-4: Obtain obstacle features of the predicted frame and the real frame according to the frame loss function.

[0091] In another specific example, the loss function includes a border loss function, and the obstacle feature is obtained according to the loss function of the obstacle, and the specific steps are as follows:

[0092] First, we need to establish a bounding box loss function. The purpose of this bounding box loss function is to measure the difference between the predicted box and the true box, so as to help better obtain the characteristics of obstacles. In this way, we can ensure that the predicted box can be as close to the true box as possible, thereby improving the accuracy of the model.

[0093] Next, the predicted box is compared with the true box according to the established bounding box loss function. By calculating the difference between the two, a loss value can be obtained. This loss value reflects the accuracy of the predicted box in locating obstacles. By continuously optimizing this loss value, the parameters of the model can be gradually adjusted so that the predicted box can better match the true box.

[0094] Ultimately, in this way, the characteristics of obstacles can be effectively obtained. These characteristics include information such as the location, shape, size, etc. With this information, we can further perform tasks such as obstacle detection, classification, and tracking, thereby improving the performance and reliability of the entire system.

[0095] In a possible implementation, the border loss function is expressed as:

[0096]

[0097] Among them, b is the center point of the predicted box, bgt is the center point of the real box, ρ is the shortest distance between the two center points, c is the diagonal distance of the minimum closure area that can contain both the predicted box and the real box, α is the weight function, and v is the similarity function that measures the aspect ratio.

[0098] In a possible implementation, the expression of the border loss function can be described in detail as follows:

[0099]

[0100] Among them, b represents the coordinates of the center point of the predicted box, and bgt represents the coordinates of the center point of the real box. d represents the shortest distance between the two center points, which is calculated by the Euclidean distance formula. c represents the diagonal distance of the minimum closure area that can contain both the predicted box and the real box. This distance reflects the relative size of the two boxes in space. α is a weight function used to balance the contribution of different loss terms to ensure that the loss function can effectively guide model optimization during training. v is a similarity function that measures the aspect ratio, which is used to evaluate the similarity in shape between the predicted box and the real box, thereby further optimizing the positioning accuracy of the box. Through the design of this loss function, the performance of the target detection model in the bounding box regression task can be effectively improved.

[0101] As mentioned above, the expression of the similarity function v that measures the aspect ratio is:

[0102]

[0103] Among them, w is the width and h is the height.

[0104] When discussing the similarity function for measuring aspect ratios, we can introduce the above function v, where v represents the similarity function, which is used to measure the similarity between two aspect ratios. The variable w represents the width, and h represents the height. Through this function, we can calculate the similarity between two different aspect ratios for further analysis and comparison.

[0105] The present invention adds the aspect ratio of the target frame into the calculation, and adds the aspect ratio design of the target frame, so that it can be better fed back to the prediction frame, so that the prediction frame is closer to the real detection frame, and the efficiency and accuracy of the detection are improved. Therefore, the present invention adds the operator:

[0106]

[0107] Used to calculate the aspect ratio of the target box and feed it back to the prediction box

[0108] The present invention takes the aspect ratio of the target frame into consideration during the calculation process. By designing the aspect ratio of the target frame in detail, this information can be fed back into the prediction frame more accurately. In this way, the prediction frame can be closer to the actual shape and size of the real detection frame, thereby significantly improving the efficiency and accuracy of the detection. To achieve this goal, the present invention introduces a special operator:

[0109] The main function of this operator is to calculate the aspect ratio of the target box and effectively pass this ratio information to the predicted box. In this way, the predicted box can better simulate the shape of the real detection box and ensure the accuracy and reliability of the detection results.

[0110] In summary, the present invention provides a deep detection function for obstacle identification. In the feature extraction stage, Efficientnet is used as the backbone network to extract features, and the feature layers of different sizes extracted by the backbone network are fused to achieve better feature extraction effects. Specifically, the present invention uses Efficientnet as the basic backbone network, which is efficient and accurate and can effectively extract key feature information from the image. On this basis, the present invention further fuses the extracted feature layers of different scales. By fusing features at different levels, the model can capture both the detailed information and the global information of the object, thereby improving the comprehensiveness and accuracy of feature extraction.

[0111] In addition, in order to further improve the performance of target detection, the present invention adopts FocalLoss in the classification loss function. Focal Loss is a loss function designed for the problem of class imbalance, which can make the model pay more attention to samples that are difficult to classify, thereby improving the accuracy of classification. In the border regression loss function, the present invention adopts CIoULoss, which is an improved border loss function, which not only considers the position and size of the border, but also the aspect ratio of the border, so that the model can more accurately predict the aspect ratio of the target and improve the accuracy of border regression.

[0112] In order to further improve the accuracy of prediction, the present invention also designs new operators and adds corresponding feedback parameters. These newly designed operators can better handle feature fusion and information transmission, and the added feedback parameters can provide more adjustment space, so that the model can be better optimized and adjusted during the training process, thereby further improving the accuracy of prediction.

[0113] Through the above series of technical improvements and optimizations, the present invention successfully implements an accurate and fast obstacle detection algorithm that can provide decision data for the obstacle avoidance system. The algorithm can not only accurately identify and locate obstacles, but also provide detailed information about obstacles, such as aspect ratio, etc., thereby providing more abundant decision support for the obstacle avoidance system, enabling the obstacle avoidance system to make decisions and respond more intelligently and efficiently.

[0114] In a second aspect, an embodiment of the present application provides an electronic device, including:

[0115] Memory for storing computer programs;

[0116] A processor is used to implement the above obstacle detection method when executing the computer program.

[0117] Specifically, the device has the following components and functions. Specifically, the electronic device includes a memory, the main function of which is to store computer programs. These computer programs can be software codes in various forms, including but not limited to applications, operating systems, middleware, etc. The memory can be any form of non-volatile storage device, such as flash memory, hard disk drive or solid state drive, etc., to ensure that the stored program remains unchanged after the device is powered off.

[0118] In addition, the electronic device is also equipped with a processor, whose main responsibility is to implement and run the above obstacle detection method when executing a computer program stored in the memory. The processor can be a central processing unit (CPU), a microcontroller or other form of digital signal processor (DSP), depending on the complexity of the device and the application requirements. In the process of executing the program, the processor will perform a series of calculations and logical operations according to the program instructions, thereby realizing the function of detecting obstacles in the surrounding environment.

[0119] The obstacle detection method can include a variety of technologies, such as image recognition, laser scanning, ultrasonic detection, etc., depending on the design and application scenario of the electronic device. For example, if the electronic device is an autonomous vehicle, the obstacle detection method may use radar and cameras to identify and locate other vehicles, pedestrians or stationary objects on the road. In this way, the processor can process sensor data in real time, analyze potential obstacles, and take corresponding obstacle avoidance measures to ensure the safe driving of the vehicle.

[0120] In summary, the electronic device provided in the embodiments of the present application can effectively implement the obstacle detection method through the collaborative work of its memory and processor, thereby providing safety protection and enhancing user experience in various application scenarios.

[0121] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned obstacle detection method are implemented.

[0122] Specifically, this computer-readable storage medium has a storage function and can save and record data information thereon. Specifically, a section of computer program code is pre-stored on the storage medium. This section of program code contains instructions and data for implementing specific functions. When the processor accesses and executes this section of computer program, it can complete the task of detecting and identifying obstacles in the surrounding environment according to the steps and logic defined in the program. In this way, the processor can effectively process and analyze data from sensors or other input devices, thereby accurately identifying key information such as the location, shape and type of obstacles. Such an implementation not only improves the automation level of the system, but also enhances its adaptability and reliability in complex environments.

[0123] In an embodiment of the present application, an innovative obstacle detection method is proposed, which uses advanced obstacle recognition technology to achieve a deep detection function. In the feature extraction stage, the method uses Efficientnet as the backbone network, so that rich and detailed feature information can be extracted. By meticulously identifying and analyzing the various obstacle features extracted by the backbone network, the method can more effectively extract key features, thereby significantly improving the accuracy of the prediction. In addition, the method also provides an accurate and fast obstacle detection technology, which can provide reliable and real-time decision-making data support for the obstacle avoidance system, ensuring the efficient operation and safety of the system.

[0124] It should be noted that the phrases "one embodiment", "an embodiment", "an exemplary embodiment", "some embodiments", etc. mentioned in the specification indicate that the described embodiments may include certain features, structures or characteristics, but not every embodiment may include the certain features, structures or characteristics. In addition, such phrases do not necessarily refer to the same embodiment. In addition, when describing certain features, structures or characteristics in conjunction with an embodiment, it is within the knowledge of those skilled in the art to implement such features, structures or characteristics in conjunction with other embodiments, whether explicitly or not explicitly described.

[0125] It should be easily understood that “on,” “above,” and “over” in the present disclosure should be interpreted in the broadest manner, so that “on” not only means “directly on something,” but also includes the meaning of “on something” with intervening features or layers therebetween, and “above” or “over” not only includes the meaning of “above” or “over,” but also may include the meaning of “above” or “over something” with no intervening features or layers therebetween (i.e., directly on something).

[0126] In addition, spatially relative terms, such as "below," "below," "beneath," "above," "above," etc., may be used herein for ease of description to describe the relationship of one element or feature relative to other elements or features as shown in the figures. Spatially relative terms are intended to encompass different orientations of the device in use or operation in addition to the orientation shown in the figures. The device may have other orientations (rotated 90 degrees or at other orientations), and the spatially relative descriptors used herein may likewise be interpreted accordingly.

[0127] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting obstacles, applied to rail transportation for detecting obstacles, characterized in that: The detection method comprises: Establish the main network based on Efficientnet-v2; According to the main network of Efficientnet-v2, obstacle features are extracted; According to the obstacle characteristics, obtain the block output result; Obtain obstacle detection results based on the block output results.

2. The obstacle detection method according to claim 1, characterized in that: According to the main network of Efficientnet-v2, obstacle features are extracted, including: Based on the main network of Efficientnet-v2, a Feffol multi-target detection model is established; According to the Feffol multi-target detection model, the obstacle features are extracted.

3. The obstacle detection method according to claim 2, characterized in that: The obstacle characteristics include at least one or a combination of road collapse, potholes, water accumulation, and falling rocks.

4. The obstacle detection method according to claim 2, characterized in that: The step of extracting the obstacle features according to the Feffol multi-target detection model includes: According to the Feffol multi-target detection model, a loss function of obstacles is established; The obstacle feature is obtained according to the loss function of the obstacle.

5. The obstacle detection method according to claim 4, characterized in that: The loss function includes a classification loss function, and obtaining the obstacle feature according to the loss function of the obstacle includes: Establish the cross entropy function; According to the cross entropy function, multiple types of obstacle features are obtained.

6. The obstacle detection method according to claim 5, characterized in that: The cross entropy function expression is: Among them, y is the value of whether the obstacle is the foreground, and p is the probability that the obstacle is the foreground.

7. The obstacle detection method according to claim 4, characterized in that: The loss function includes a border loss function, and obtaining the obstacle feature according to the loss function of the obstacle includes: Establish a bounding box loss function; Obstacle features of the predicted frame and the real frame are obtained according to the frame loss function.

8. The obstacle detection method according to claim 7, characterized in that: The expression of the border loss function is: Among them, b is the center point of the predicted box, bgt is the center point of the real box, ρ is the shortest distance between the two center points, c is the diagonal distance of the minimum closure area that can contain both the predicted box and the real box, α is the weight function, and v is the similarity function that measures the aspect ratio.

9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the obstacle detection method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the obstacle detection method according to any one of claims 1 to 8 are implemented.

Images