Slope rockfall movement intelligent detection method and device

Abstract

The invention discloses a slope rockfall movement intelligent detection method and device, and relates to the field of rockfall detection.The slope rockfall movement intelligent detection method mainly comprises the steps that an image of a slope area is obtained and preprocessed, a trained real-time target detection network model is used for conducting static detection on preprocessed image data, and a target detection result is obtained; and performing dynamic detection by using the optical flow model, and comprehensively analyzing static detection and dynamic detection results to obtain rockfall risk data and early warning information. By implementing the intelligent detection method and device for movement of the slope rockfall, high-speed, efficient and accurate real-time monitoring of the slope rockfall can be realized.

Description

Technical Field

[0001] The present invention relates to the field of rockfall detection, and more specifically, to a method and device for intelligently detecting rockfall movement on a slope. Background Art

[0002] Existing rockfall detection technology has many shortcomings. Manual inspections require a lot of manpower and time, and are subject to the limitations of human visual observation, which can easily lead to missed inspections and false alarms. There are safety risks for inspectors walking in complex terrain, especially at night or in bad weather conditions, when inspection efficiency and safety are even more difficult to guarantee. In addition, the frequency of manual inspections is limited, and it is impossible to achieve all-weather, real-time monitoring, which can easily miss the early signs of rockfall and delay the warning time.

[0003] In recent years, with the rapid development of artificial intelligence and computer vision technology, deep learning-based object detection algorithms have gradually been applied to various detection tasks. However, traditional deep learning detection models still face great challenges in terms of real-time performance and resource consumption, and are difficult to run efficiently on embedded devices with limited resources. Typical deep learning models such as SSD and RetinaNet, although they perform well in detection accuracy, are not satisfactory in terms of speed and computing resource consumption, and are not suitable for real-time requirements. In addition, these models have high requirements for hardware resources and are not easy to deploy on mobile devices or edge computing platforms.

[0004] Therefore, there is an urgent need for an intelligent detection system that can ensure high detection accuracy and good real-time performance, so as to realize real-time monitoring and early warning of slope rockfall. Summary of the invention

[0005] The purpose of the present invention is to provide a method and device for intelligent detection of falling rocks on slopes, which can realize high-speed, efficient and accurate real-time monitoring of falling rocks on slopes.

[0006] The present invention provides a method for intelligent detection of falling rocks on a slope, comprising the following steps: S1: acquiring an image of a slope area, and performing preprocessing to obtain preprocessed image data; S2: building a real-time target detection network model and performing training to obtain a trained real-time target detection network model; performing target detection on the preprocessed image data using the trained real-time target detection network model to obtain target position and category data; S3: obtaining the motion trajectory and speed data of falling rocks using an optical flow model based on the preprocessed image data; S4: obtaining rockfall risk data and early warning information based on the target position and category data and the motion trajectory and speed data of falling rocks.

[0007] Furthermore, the real-time target detection network model of the above-mentioned slope rockfall movement intelligent detection method is built based on the YOLO framework, and the real-time target detection network model includes a trunk network, a neck network and a head network; the trunk network includes a first convolutional network, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an SPPF module connected in sequence; the neck network includes a first Upsampl network, a first splicing module, a first attention module, a second Upsampl network, a second splicing module, a second attention module, a second convolutional network, a third splicing module, a third attention module, a third convolutional network, a fourth splicing module and a fourth attention module connected in sequence ; The head network includes a first detection module, a second detection module and a third detection module; the first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure and are collectively referred to as convolution modules; the convolution module includes a connected convolution network and a feature extraction module; the first splicing module, the second splicing module, the third splicing module and the fourth splicing module have the same structure and are collectively referred to as splicing modules; the splicing module includes a connected splicing network and a feature extraction module; the input of the first splicing module is also connected to the output of the convolution network of the third convolution module, the input of the second Upsampl network is also connected to the output of the second convolution module, and the input of the fourth splicing module is also connected to the output of the SPPF module.

[0008] Furthermore, the first attention module, the second attention module, the third attention module and the fourth attention module of the above-mentioned slope rockfall movement intelligent detection method are constructed using the following formula:

[0009] AttentionMap=σ(W1·ReLU(W2·X))

[0010] Among them, AttentionMap is the output of the attention module, X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit.

[0011] Furthermore, the second convolutional network and the third convolutional network of the above-mentioned slope rockfall movement intelligent detection method are constructed using the following formula:

[0012]

[0013] Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group.

[0014] The present invention also provides a device for detecting mobile rockfall on a slope, comprising the following modules: an image acquisition module, configured to acquire an image of a slope area, and perform preprocessing to obtain preprocessed image data; a static detection module, configured to build a real-time target detection network model and perform training to obtain a trained real-time target detection network model; perform target detection on the preprocessed image data using the trained real-time target detection network model to obtain target position and category data; a dynamic detection module, configured to obtain the motion trajectory and speed data of rockfall based on the preprocessed image data using an optical flow model; and a comprehensive analysis module, configured to obtain rockfall risk data and early warning information based on the target position and category data and the motion trajectory and speed data of rockfall.

[0015] Furthermore, the real-time target detection network model of the above-mentioned slope rockfall moving intelligent detection device is built based on the YOLO framework, and the real-time target detection network model includes a trunk network, a neck network and a head network; the trunk network includes a first convolutional network, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an SPPF module connected in sequence; the neck network includes a first Upsampl network, a first splicing module, a first attention module, a second Upsampl network, a second splicing module, a second attention module, a second convolutional network, a third splicing module, a third attention module, a third convolutional network, a fourth splicing module and a fourth attention module connected in sequence ; The head network includes a first detection module, a second detection module and a third detection module; the first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure and are collectively referred to as convolution modules; the convolution module includes a connected convolution network and a feature extraction module; the first splicing module, the second splicing module, the third splicing module and the fourth splicing module have the same structure and are collectively referred to as splicing modules; the splicing module includes a connected splicing network and a feature extraction module; the input of the first splicing module is also connected to the output of the convolution network of the third convolution module, the input of the second Upsampl network is also connected to the output of the second convolution module, and the input of the fourth splicing module is also connected to the output of the SPPF module.

[0016] Furthermore, the first attention module, the second attention module, the third attention module and the fourth attention module of the above-mentioned intelligent detection device for falling rocks on slopes are constructed using the following formula:

[0017] AttentionMap=σ(W1·ReLU(W2·X))

[0018] Among them, AttentionMap is the output of the attention module, X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit.

[0019] Furthermore, the second convolutional network and the third convolutional network of the above-mentioned intelligent detection device for falling rocks on slopes are constructed using the following formula:

[0020]

[0021] Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group.

[0022] Furthermore, the above-mentioned mobile intelligent detection device for falling rocks on slopes also includes a data transmission module, an information interaction module and a self-inspection and operation and maintenance module; the data transmission module is used to transmit the pre-processed image data to the embedded computing platform; the information interaction module is used to obtain monitoring and early warning interaction information according to the target position and category data, the movement trajectory and speed data of falling rocks, the rockfall risk data and early warning information, and display the monitoring and early warning interaction information in a visual manner; the self-inspection and operation and maintenance module is used to regularly self-check the operating status of the device, and perform remote maintenance and updates according to the operating status.

[0023] Furthermore, the above-mentioned mobile intelligent detection device for falling rocks on slopes also includes a power management module, which is used to manage the power of various components of the mobile intelligent detection device for falling rocks on slopes; the image acquisition module includes a high-resolution camera, the static detection module includes an embedded computing platform, and the data transmission module uses wireless transmission to transmit data.

[0024] The implementation of the slope rockfall movement intelligent detection method and device provided by the present invention has the following beneficial effects:

[0025] The present invention designs a complete algorithm framework based on the background of rockfall detection, and adds point-by-point spatial attention and grouped spatial convolution modules to the framework, thereby improving the accuracy and anti-interference ability of the model, simplifying the model, saving computing costs, and making it more suitable for embedded platform operation; the present invention integrates the FlowNet optical flow model for motion detection, and accurately captures the motion trajectory and speed of the object by calculating the optical flow information between consecutive image frames, so that the system can analyze the rockfall motion in real time, thereby improving the timeliness and accuracy of detection; the combination of the optical flow model and the YOLO model can not only identify static rockfalls, but also track their motion trajectories, timely judge the rockfall risks, send out early warning signals, and store relevant data for subsequent analysis and verification, providing more comprehensive and accurate monitoring and early warning services; the present invention combines the rockfall detection algorithm framework with hardware such as embedded development boards, which can automatically identify, classify and process a large amount of data, improve the monitoring ability and quality of rockfall safety hazards, and realize high-speed, efficient and accurate real-time monitoring of slope rockfall on the basis of visualization and interactivity. The slope rockfall mobile intelligent detection device provided by the present invention can also operate stably in complex environments, providing reliable support for slope safety protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0027] Figure 1 It is a flow chart of the intelligent detection method for falling rocks on slopes provided by the present invention;

[0028] Figure 2 It is a schematic diagram of the structure of the real-time target detection network model provided by the present invention;

[0029] Figure 3 It is a schematic diagram of the GSConv convolution structure in the real-time target detection network model provided by the present invention;

[0030] Figure 4 It is a structural schematic diagram of the attention mechanism PSA module provided by the present invention;

[0031] Figure 5 It is a schematic diagram of the structure of the FlowNet optical flow model provided by the present invention;

[0032] Figure 6 The present invention is a schematic diagram of the detection process of the intelligent detection device for falling rocks on slopes. DETAILED DESCRIPTION

[0033] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0034] Figure 1A schematic diagram of the intelligent detection method for rockfall on a slope according to the present embodiment is shown. In the present embodiment, the intelligent detection method for rockfall on a slope comprises the following steps:

[0035] S1: Acquire an image of the slope area and perform preprocessing to obtain preprocessed image data;

[0036] S2: Build a real-time target detection network model and train it to obtain a trained real-time target detection network model; use the trained real-time target detection network model to perform target detection on the preprocessed image data to obtain target position and category data;

[0037] Specifically, the above-mentioned real-time target detection network model is built based on the YOLO framework, and the real-time target detection network model includes a trunk network, a neck network and a head network; the trunk network includes a first convolutional network, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an SPPF module connected in sequence; the neck network includes a first Upsampl network, a first splicing module, a first attention module, a second Upsampl network, a second splicing module, a second attention module, a second convolutional network, a third splicing module, a third attention module, a third convolutional network, a fourth splicing module and a fourth attention module connected in sequence; the head network includes The first detection module, the second detection module and the third detection module; the first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure and are collectively referred to as convolution modules; the convolution module includes a connected convolution network and a feature extraction module; the first splicing module, the second splicing module, the third splicing module and the fourth splicing module have the same structure and are collectively referred to as splicing modules; the splicing module includes a connected splicing network and a feature extraction module; the input of the first splicing module is also connected to the output of the convolution network of the third convolution module, the input of the second Upsampl network is also connected to the output of the second convolution module, and the input of the fourth splicing module is also connected to the output of the SPPF module;

[0038] Specifically, Figure 2 As shown in the figure, Backbone is the trunk network, Neck is the neck network, Head is the head network, the first convolutional network is Conv, attention is the attention module, C2f is the feature extraction module, Concat is the splicing network, Detect is the detection module, the second convolutional network is CSConv and the third convolutional network is CSConv;

[0039] Specifically, the first attention module attention, the second attention module attention, the third attention module attention and the fourth attention module attention of the above-mentioned slope rockfall movement intelligent detection method are constructed using the following formula:

[0040] AttentionMap=σ(W1·ReLU(W2·X))

[0041] Among them, AttentionMap is the output of the attention module, X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit;

[0042] Specifically, the second convolutional network and the third convolutional network of the above-mentioned slope rockfall movement intelligent detection method are constructed using the following formula:

[0043]

[0044] Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group;

[0045] S3: Based on the preprocessed image data, the motion trajectory and speed data of the falling rock are obtained using the optical flow model;

[0046] S4: Obtain rockfall risk data and warning information based on target location and category data and rockfall trajectory and speed data.

[0047] This embodiment provides an intelligent detection device for rockfall movement on a slope, comprising the following modules:

[0048] The image acquisition module is configured to: acquire an image of the slope area and perform preprocessing to obtain preprocessed image data;

[0049] Specifically, the image acquisition module uses a high-resolution camera to collect images of the slope area in real time and perform preprocessing. The camera has a wide-angle field of view and night vision capabilities, and can capture images clearly under various lighting conditions. The collected original images undergo preprocessing steps, including image scaling, normalization, denoising and other operations, to adjust the image size and pixel value range to meet the input requirements of the subsequent deep learning model. Preprocessing also includes contrast enhancement and sharpening to improve image quality and feature extraction effects.

[0050] The static detection module is configured as follows: building a real-time target detection network model and training it to obtain a trained real-time target detection network model; using the trained real-time target detection network model to perform target detection on the preprocessed image data to obtain target position and category data;

[0051] Specifically, the above-mentioned real-time target detection network model is built based on the YOLO framework, and the real-time target detection network model includes a trunk network, a neck network and a head network; the trunk network includes a first convolutional network, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an SPPF module connected in sequence; the neck network includes a first Upsampl network, a first splicing module, a first attention module, a second Upsampl network, a second splicing module, a second attention module, a second convolutional network, a third splicing module, a third attention module, a third convolutional network, a fourth splicing module and a fourth attention module connected in sequence; the head network includes The first detection module, the second detection module and the third detection module; the first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure and are collectively referred to as convolution modules; the convolution module includes a connected convolution network and a feature extraction module; the first splicing module, the second splicing module, the third splicing module and the fourth splicing module have the same structure and are collectively referred to as splicing modules; the splicing module includes a connected splicing network and a feature extraction module; the input of the first splicing module is also connected to the output of the convolution network of the third convolution module, the input of the second Upsampl network is also connected to the output of the second convolution module, and the input of the fourth splicing module is also connected to the output of the SPPF module;

[0052] Specifically, Figure 2 As shown in the figure, Backbone is the trunk network, Neck is the neck network, Head is the head network, the first convolutional network is Conv, attention is the attention module, C2f is the feature extraction module, Concat is the splicing network, Detect is the detection module, the second convolutional network is CSConv and the third convolutional network is CSConv;

[0053] Specifically, the first attention module attention, the second attention module attention, the third attention module attention and the fourth attention module attention of the above-mentioned slope rockfall movement intelligent detection method are constructed using the following formula:

[0054] AttentionMap=σ(W1·ReLU(W2·X))

[0055] Among them, AttentionMap is the output of the attention module, X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit;

[0056] Specifically, the second convolutional network and the third convolutional network of the above-mentioned slope rockfall movement intelligent detection method are constructed using the following formula:

[0057]

[0058] Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group;

[0059] The dynamic state detection module is configured to: obtain the motion trajectory and speed data of the falling rock by using the optical flow model according to the preprocessed image data;

[0060] The comprehensive analysis module is configured to obtain rockfall risk data and warning information based on target location and category data and rockfall movement trajectory and speed data.

[0061] Specifically, the above-mentioned mobile intelligent detection device for falling rocks on slopes also includes a data transmission module, an information interaction module and a self-inspection and operation and maintenance module; the data transmission module is used to transmit the pre-processed image data to the embedded computing platform; the information interaction module is used to obtain monitoring and early warning interaction information based on target position and category data, movement trajectory and speed data of falling rocks, rockfall risk data and early warning information, and display the monitoring and early warning interaction information in a visual manner; the self-inspection and operation and maintenance module is used to regularly self-check the operating status of the device, and perform remote maintenance and updates according to the operating status.

[0062] Specifically, the above-mentioned intelligent mobile rockfall detection device for detecting falling rocks on slopes also includes a power management module, which is used to perform power management on the various components of the intelligent mobile rockfall detection device for detecting falling rocks on slopes; the image acquisition module includes a high-resolution camera, the static detection module includes an embedded computing platform, and the data transmission module uses wireless transmission to transmit data.

[0063] Specifically, the data transmission module of the above-mentioned intelligent mobile detection device for falling rocks on slopes is configured as follows: the pre-processed image data is transmitted to the embedded computing platform in real time through the wireless transmission module; the wireless transmission module adopts Wi-Fi, 4G / 5G and other technologies to ensure the efficiency and stability of data transmission; in order to ensure the integrity and low latency of data during transmission, compression and encryption technologies are adopted to reduce the amount of transmitted data and prevent data loss and information leakage; the real-time transmitted image data is further processed and analyzed on the embedded platform;

[0064] Specifically, the above-mentioned real-time target detection network model is mainly constructed as follows: running the optimized YOLO model on the embedded computing platform for target detection, outputting the location and category of the target, and introducing PSA and GSConv modules to improve accuracy and anti-interference ability;

[0065] In this embodiment, the above-mentioned slope rockfall mobile intelligent detection device runs the optimized YOLO series model on the embedded computing platform to perform real-time target detection on the received image; the YOLO model completes the detection task of the entire image through one forward propagation and outputs the position and category of each target; in order to improve the detection accuracy and anti-interference ability, while reducing the computing cost, the model introduces point-wise spatial attention (PSA) and grouped spatial convolution (GSConv) modules; these optimizations enable the model to run efficiently on the embedded platform and realize real-time detection; the PSA module improves the accuracy of target detection by a point-by-point attention mechanism on the feature map; specifically, PSA strengthens the key area by calculating the importance weight of each point on the feature map, such as the formula:

[0066] Attention Map=σ(W1·ReLU(W2·X))

[0067] Where X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit;

[0068] The GSConv module reduces the amount of computation and improves the efficiency of the model through grouped convolution. Grouped convolution divides the feature map into multiple groups, and each group performs an independent convolution operation. The formula is expressed as:

[0069]

[0070] Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group;

[0071] Specifically, Figure 2 As shown in the figure, for the processed slope rockfall image data input, the image first passes through a series of convolutional layers to extract preliminary features. After passing through the first convolutional layer, it enters the C2f module for further feature extraction and compression, passes through the convolutional layer again, and then another C2f module. The feature extraction process is repeated 4 times, and then the multi-scale features are further extracted through the fast spatial pyramid pooling layer SPPF module to generate the final feature map; the feature map is fused and extracted through upsampling (Upsample) and C2f modules; the upsampled feature map and the feature map output by the backbone network are concatenated (Concat), and the concatenated feature map is passed Figure 4The attention module shown in the figure (the model introduces point-wise spatial attention, Point-wise Spatial Attention, PSA) enhances important features and then passes Figure 3 The GSConv module shown in the figure performs feature compression; the compressed feature map is concatenated again (Concat) and passes through another C2f module and upsampling operation, repeating the previous process, passing through multiple attention and CSC modules, and finally generating a feature map for detection, and finally inputting it into the detection module (Detect) for target recognition and positioning, and outputting the location and category information of the rockfall target in the image;

[0072] Specifically, the above method uses the optical flow model based on the preprocessed image data to obtain the motion trajectory and speed data of the falling rock. The main contents are: combining the FlowNet optical flow model to perform motion detection on continuous image frames, and by calculating the optical flow information between image sequences, identify and track fast-moving falling rocks, and determine their motion trajectory and speed; the optical flow model can accurately capture the changes of moving targets, provide dynamic monitoring data, make up for the shortcomings of static target detection, and further improve the detection accuracy and real-time performance of the system; dynamic analysis includes determining the motion trajectory and speed of falling rocks to provide support for comprehensive monitoring; FlowNet detects motion by calculating the optical flow information between continuous frames, and the optical flow information is expressed as:

[0073]

[0074] Among them, u and v represent the optical flow components in the horizontal and vertical directions respectively; the loss function of FlowNet includes image reconstruction error and optical flow smoothing term:

[0075]

[0076] in, is the reconstruction error, is the smoothing term, λ is the balance coefficient;

[0077] Specifically, Figure 5 As shown in the figure, the two input images are stacked together. The optical flow model recognizes and tracks fast-moving falling rocks by calculating the optical flow information between consecutive image frames. The images are sequentially passed through multiple convolutional layers (Conv1 to Conv6) for feature extraction. The size of the feature map is gradually reduced while the number of channels is increased. High-level features are extracted through this series of convolution operations. Subsequently, the feature map enters the refinement module for further processing and optimization to obtain the final output.

[0078] Specifically, the rockfall risk data and warning information are obtained based on the target position and category data and the movement trajectory and speed data of the rockfall, and the content is mainly as follows: comprehensively analyze the static detection results and dynamic detection results, judge the rockfall risk, issue a warning signal, and store the detection data and results; combine the static detection results of the YOLO model with the dynamic detection results of the FlowNet optical flow model for comprehensive analysis; accurately judge whether there is a rockfall risk through comprehensive processing of the detection data, and issue a warning signal in time; all detection data and analysis results will be stored in the local storage device for subsequent analysis and verification; these data records can be used to improve the model and enhance the detection capability of the system;

[0079] Specifically, the above-mentioned information interaction module is configured as follows: to obtain monitoring and early warning interaction information based on target location and category data, rockfall movement trajectory and speed data, rockfall risk data and early warning information, and to display the monitoring and early warning interaction information in a visual manner; to display detection data, detection results and early warning information in real time on the user interface, and to provide user interaction functions to adjust parameters and view historical data; to display detection data, detection results and early warning information in real time on the user interface; the user interface intuitively displays the slope conditions monitored by the system, helping monitoring personnel to understand and deal with potential risks in a timely manner; to provide a touch screen or keyboard input function, allowing users to adjust monitoring parameters, view historical data and process alarm information; the interactive interface is designed to be friendly and easy to operate, thereby improving the operability and user experience of the system;

[0080] like Figure 6 As shown, the camera captures images of the slope surface and transmits the data to the hardware processing device, which then processes the images using a pre-trained model to detect falling rocks and other targets. The detection results are displayed on the interactive interface, including the number of categories, the number of targets, and the number of frames per second. The interface provides real-time monitoring, displays unprocessed and processed images, and annotates detected targets and confidence scores, thereby achieving efficient and accurate monitoring of falling rocks on slopes.

[0081] Specifically, the self-inspection and operation and maintenance module is configured as follows: regularly self-check the operating status of the device, and perform remote maintenance and updates based on the operating status; that is, regularly check the operating status of the hardware and software to ensure the normal operation of each component and prevent failures; the self-inspection function includes a comprehensive inspection of the camera, wireless transmission module, computing platform and power system; remote maintenance and updates are achieved through the wireless transmission module to facilitate system management and upgrades; the remote maintenance function allows technicians to monitor and debug the system at any time and any place, thereby improving the system's maintenance efficiency and reliability.

[0082] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for intelligent detection of rockfall on a slope, characterized in that: The following steps are involved: S1: Acquire an image of the slope area and perform preprocessing to obtain preprocessed image data; S2: Building a real-time target detection network model and training it to obtain a trained real-time target detection network model; using the trained real-time target detection network model to perform target detection on the preprocessed image data to obtain target position and category data; S3: obtaining the motion trajectory and speed data of the falling rock using an optical flow model according to the preprocessed image data; S4: Obtain rockfall risk data and warning information based on the target position and category data and the movement trajectory and speed data of the rockfall.

2. The intelligent detection method for falling rocks on slopes according to claim 1 is characterized in that: The real-time target detection network model is built based on the YOLO framework, and the real-time target detection network model includes a trunk network, a neck network and a head network; the trunk network includes a first convolutional network, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an SPPF module connected in sequence; the neck network includes a first Upsampl network, a first splicing module, a first attention module, a second Upsampl network, a second splicing module, a second attention module, a second convolutional network, a third splicing module, a third attention module, a third convolutional network, a fourth splicing module and a fourth attention module connected in sequence; the head network It includes a first detection module, a second detection module and a third detection module; the first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure and are collectively referred to as convolution modules; the convolution module includes a connected convolution network and a feature extraction module; the first splicing module, the second splicing module, the third splicing module and the fourth splicing module have the same structure and are collectively referred to as splicing modules; the splicing module includes a connected splicing network and a feature extraction module; the input of the first splicing module is also connected to the output of the convolution network of the third convolution module, the input of the second Upsampl network is also connected to the output of the second convolution module, and the input of the fourth splicing module is also connected to the output of the SPPF module.

3. The intelligent detection method for falling rocks on slopes according to claim 2 is characterized in that: The first attention module, the second attention module, the third attention module and the fourth attention module are constructed using the following formula: AttentionMap=σ(W1·ReLU(W2·X)) Among them, AttentionMap is the output of the attention module, X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit.

4. The intelligent detection method for falling rocks on slopes according to claim 2 is characterized in that: The second convolutional network and the third convolutional network are constructed using the following formula: Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group.

5. An intelligent detection device for rockfall on a slope, characterized in that: The intelligent detection device for slope rockfall movement includes the following modules: The image acquisition module is configured to: acquire an image of the slope area and perform preprocessing to obtain preprocessed image data; The static detection module is configured to: build a real-time target detection network model and train it to obtain a trained real-time target detection network model; use the trained real-time target detection network model to perform target detection on the preprocessed image data to obtain target position and category data; The dynamic state detection module is configured to: obtain the motion trajectory and speed data of the falling rock by using the optical flow model according to the pre-processed image data; The comprehensive analysis module is configured to obtain rockfall risk data and warning information based on the target position and category data and the movement trajectory and speed data of the rockfall.

6. The intelligent detection device for rockfall on slope according to claim 5, characterized in that: The real-time target detection network model is built based on the YOLO framework, and the real-time target detection network model includes a trunk network, a neck network and a head network; the trunk network includes a first convolutional network, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and an SPPF module connected in sequence; the neck network includes a first Upsampl network, a first splicing module, a first attention module, a second Upsampl network, a second splicing module, a second attention module, a second convolutional network, a third splicing module, a third attention module, a third convolutional network, a fourth splicing module and a fourth attention module connected in sequence; the head network It includes a first detection module, a second detection module and a third detection module; the first convolution module, the second convolution module, the third convolution module and the fourth convolution module have the same structure and are collectively referred to as convolution modules; the convolution module includes a connected convolution network and a feature extraction module; the first splicing module, the second splicing module, the third splicing module and the fourth splicing module have the same structure and are collectively referred to as splicing modules; the splicing module includes a connected splicing network and a feature extraction module; the input of the first splicing module is also connected to the output of the convolution network of the third convolution module, the input of the second Upsampl network is also connected to the output of the second convolution module, and the input of the fourth splicing module is also connected to the output of the SPPF module.

7. The intelligent detection device for rockfall on slope according to claim 6, characterized in that: The first attention module, the second attention module, the third attention module and the fourth attention module are constructed using the following formula: AttentionMap=σ(W1·ReLU(W2·X)) Among them, AttentionMap is the output of the attention module, X is the input feature map, W1 and W2 are trainable weight matrices, σ represents the Sigmoid activation function, and ReLU is the rectified linear unit.

8. The intelligent detection device for rockfall on slope according to claim 6, characterized in that: The second convolutional network and the third convolutional network are constructed using the following formula: Among them, X ik represents the input features of the i-th group, W k represents the kth convolution kernel, Y i represents the output features of the i-th group.

9. The intelligent detection device for rockfall on slope according to claim 5, characterized in that: The intelligent detection device for falling rocks on slopes also includes a data transmission module, an information interaction module and a self-checking and operation and maintenance module; the data transmission module is used to transmit the pre-processed image data to the embedded computing platform; The information interaction module is used to obtain monitoring and early warning interaction information based on the target position and category data, the movement trajectory and speed data of the falling rocks, the falling rock risk data and early warning information, and display the monitoring and early warning interaction information in a visual manner; the self-inspection and operation and maintenance module is used to regularly self-inspect the operating status of the device, and perform remote maintenance and updates based on the operating status.

10. The intelligent detection device for rockfall on slope according to claim 9, characterized in that: The device for detecting falling rocks on slopes also includes a power management module, which is used to perform power management on the components of the device; the image acquisition module includes a high-resolution camera, the static detection module includes an embedded computing platform, and the data transmission module uses wireless transmission to transmit data.

Images