High-speed rail roadbed hidden danger intelligent identification system based on infrared image analysis

Abstract

The invention discloses a high-speed railway roadbed hidden danger intelligent identification system based on infrared image analysis, and belongs to the technical field of railway infrastructure detection. The system comprises an infrared image acquisition module, a visible light image acquisition module, an environment parameter acquisition module, an image fusion module, an environment correction module, a thermal anomaly detection module, a hidden danger classification module and a space positioning module. The system fuses infrared and visible light images based on an attention mechanism, eliminates the influence of environmental factors based on environmental parameter adaptive correction, extracts spatio-temporal features from an image sequence to detect temperature anomaly, and intelligently classifies hidden danger types and severity based on form and temperature features. And mapping the abnormal region to an actual spatial position and generating a visual distribution map. According to the invention, high-precision, high-efficiency and high-reliability intelligent identification of the hidden danger of the high-speed rail roadbed is realized, the detection rate reaches more than 95%, and technical guarantee is provided for safe operation of the high-speed rail.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of railway infrastructure detection, and particularly relates to a high-speed railway subgrade hidden danger intelligent identification system based on infrared image analysis. BACKGROUND

[0002] The high-speed railway subgrade is an important infrastructure for train safe operation, and its quality is directly related to the safety and comfort of train operation. In the long-term use process, due to the influence of factors such as geological condition change, groundwater erosion, and freeze-thaw cycle, the subgrade is prone to internal hidden dangers such as cavities, water accumulation, and material unevenness. If these hidden dangers are not discovered and treated in time, they may lead to subgrade settlement, deformation, and even collapse, which seriously threatens the safety of train operation.

[0003] Traditional subgrade detection methods mainly include seismic exploration, electrical exploration, ground penetrating radar, and other geophysical methods. Although these methods can detect the internal structure of the subgrade, they have problems such as low detection efficiency, high cost, and the need to interrupt traffic, which makes it difficult to meet the daily detection needs during high-speed rail operation. In recent years, infrared thermal imaging technology has started to be applied to subgrade defect detection due to its non-contact, fast, and intuitive advantages. When there are defects such as cavities and water accumulation inside the subgrade, due to the difference in thermal conductivity, it will cause abnormal surface temperature distribution, and through infrared thermal imaging, these temperature abnormalities can be detected.

[0004] The existing infrared thermal imaging detection method has the following technical problems. CN113203743A discloses a subgrade crack detection and identification and repair method based on infrared thermal imaging analysis. This method uses an infrared thermal imaging device and a visible light imaging device to capture the road surface respectively, exports the temperature data of the infrared image to an Excel file, and then uses Matlab programming to generate a temperature contour map. Finally, the infrared contour map and the visible light photo are compared manually to identify cracks. This method has the following shortcomings: first, the infrared image and the visible light image are collected separately and not intelligently fused, which cannot fully utilize the complementary information of the two images, and the recognition accuracy is limited; second, environmental factors such as solar radiation, ambient temperature, humidity, and wind speed are not considered, which affects the infrared imaging, and the detection results are not consistent under different weather and time conditions; third, only a single or small amount of infrared images are collected, and the time sequence evolution process of heat conduction is not analyzed, which makes it difficult to accurately judge the depth and nature of the defect; fourth, it relies on manual interpretation of the temperature contour map, lacks intelligent defect identification and classification ability, and has low detection efficiency and strong subjectivity; fifth, it cannot realize accurate spatial positioning and three-dimensional visualization of the defect, which is not conducive to subsequent maintenance decision-making.

[0005] With the development of deep learning and computer vision technologies, advanced technologies such as multimodal image fusion, environmental adaptive correction, and spatiotemporal sequence analysis have provided new approaches to solving the aforementioned problems. Therefore, there is an urgent need to develop an intelligent high-speed railway subgrade hazard identification system capable of fusing infrared and visible light image information, adaptively compensating for the influence of environmental factors, analyzing the temporal characteristics of heat conduction, and achieving automatic detection, classification, and precise location of subgrade hazards. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent identification system for hidden dangers in high-speed railway subgrades based on infrared image analysis, so as to solve the technical problems in the prior art, such as low identification accuracy due to the lack of intelligent fusion of infrared and visible light images, poor detection consistency due to the lack of compensation for the influence of environmental factors, difficulty in judging the depth and nature of hidden dangers due to the lack of time-series analysis, low efficiency and strong subjectivity due to reliance on manual interpretation, and inability to accurately locate and visualize spatially.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The high-speed railway subgrade hidden danger intelligent identification system based on infrared image analysis includes an infrared image acquisition module, a visible light image acquisition module, an environmental parameter acquisition module, an image fusion module, an environmental correction module, a thermal anomaly detection module, a hidden danger classification module, and a spatial positioning module.

[0009] The infrared image acquisition module is used to acquire infrared thermal imaging image sequences of the roadbed, capturing the thermal radiation distribution on the roadbed surface. The visible light image acquisition module is used to acquire visible light image sequences corresponding to the infrared thermal imaging image sequences, providing texture and structural information about the roadbed. The environmental parameter acquisition module is connected to the infrared and visible light image acquisition modules to acquire environmental temperature, humidity, solar radiation intensity, and wind speed parameters, providing data support for subsequent environmental correction.

[0010] The image fusion module connects to both the infrared image acquisition module and the visible light image acquisition module. Based on an attention mechanism, it performs feature-level fusion of the infrared thermal imaging image sequence and the visible light image sequence to generate a fused image sequence, fully utilizing the temperature information from the infrared image and the detailed information from the visible light image. The environmental correction module connects to both the environmental parameter acquisition module and the image fusion module. Based on environmental temperature, humidity, solar radiation intensity, and wind speed parameters, it adaptively corrects the fused image sequence to generate a corrected image sequence, eliminating the influence of environmental factors on the detection results.

[0011] The thermal anomaly detection module is connected to the environmental correction module. It extracts spatiotemporal features from the corrected image sequence to detect temperature anomaly areas within the roadbed, identifying thermal distribution anomalies caused by defects such as cavities and water accumulation. The hazard classification module, also connected to the thermal anomaly detection module, determines the type and severity of roadbed hazards based on the morphological characteristics, temperature distribution characteristics, and temporal evolution characteristics of the temperature anomaly areas. The spatial positioning module, connected to the hazard classification module, maps the temperature anomaly areas to the actual spatial location of the roadbed, generating a hazard distribution map and providing intuitive visual information for maintenance decisions.

[0012] The beneficial effects of this invention are as follows:

[0013] By setting up an image fusion module to perform feature-level fusion of infrared thermal imaging image sequences and visible light image sequences based on an attention mechanism, the temperature sensitivity of infrared images and the high-resolution detail information of visible light images can be fully utilized. Compared with the existing technology of separately acquiring infrared and visible light images without fusion, the fusion method of this invention significantly improves the accuracy of hazard identification. The fused image can simultaneously present temperature distribution and texture structure, making the correspondence between temperature anomaly areas and actual defect locations clearer.

[0014] By setting an environmental correction module to adaptively correct the fused image sequence based on environmental temperature, humidity, solar radiation intensity, and wind speed parameters, the influence of environmental factors such as solar radiation, changes in ambient temperature, humidity, and wind speed on infrared imaging can be effectively eliminated. This allows the system to obtain consistent and reliable results under different weather conditions and time periods. Compared with existing technologies that do not consider environmental factor compensation, this invention significantly improves the stability and repeatability of detection results, and the detection results are not limited by weather or time.

[0015] By setting up a thermal anomaly detection module to extract spatiotemporal features from the corrected image sequence to detect temperature anomaly areas, it is possible to analyze the temporal evolution of heat conduction and determine the depth and nature of potential hazards based on the temperature change over time. Compared with existing technologies that only collect single frames or a small number of images, this invention can more accurately identify internal defects in the roadbed through temporal analysis, distinguish between surface and deep hazards, estimate the depth and location of hazards, and provide more detailed information for the formulation of maintenance plans.

[0016] By setting up a hazard classification module, the type and severity of roadbed hazards are automatically determined based on the morphological characteristics, temperature distribution characteristics, and temporal evolution characteristics of temperature anomaly areas. This achieves intelligent hazard identification and classification. Compared with the existing technology that relies on manual interpretation of temperature contour maps, this invention significantly improves detection efficiency and objectivity. It can automatically distinguish different types of hazards such as cavities, water accumulation, and uneven materials, and assess their severity, providing a scientific basis for maintenance decisions.

[0017] By setting up a spatial positioning module to map the temperature anomaly area to the actual spatial location of the roadbed and generate a hazard distribution map, the precise spatial positioning and three-dimensional visualization of hazards are achieved. Compared with existing technologies that cannot accurately locate hazards, the hazard distribution map generated by this invention can intuitively display the location, type, depth and severity of each hazard, which greatly facilitates the on-site work of maintenance personnel and improves maintenance efficiency and pertinence.

[0018] This invention organically combines several innovative technologies, including intelligent fusion of infrared and visible light images, adaptive correction of environmental factors, temporal analysis of heat conduction, intelligent classification of hidden dangers, and precise spatial positioning, to construct a complete intelligent identification system for hidden dangers in high-speed railway subgrades. It can achieve high-precision, high-efficiency, and high-reliability detection of subgrade hidden dangers, with a detection rate that is more than 75% higher than that of existing technologies, providing strong technical support for the safe operation of high-speed railways. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall structure of the system of the present invention.

[0020] Figure 2 This is a schematic diagram of the infrared image acquisition module of the present invention.

[0021] Figure 3 This is a schematic diagram of the image fusion module of the present invention.

[0022] Figure 4 This is a schematic diagram of the environmental correction module of the present invention.

[0023] Figure 5 This is a schematic diagram of the thermal anomaly detection module of the present invention.

[0024] Figure 6 This is a schematic diagram of the hazard classification module of the present invention.

[0025] Figure 7 This is a schematic diagram of the spatial positioning module of the present invention. Detailed Implementation

[0026] Please refer to the attached document. Figures 1-7 The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Reference Figure 1 The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis of the present invention includes an infrared image acquisition module 1, a visible light image acquisition module 2, an environmental parameter acquisition module 3, an image fusion module 4, an environmental correction module 5, a thermal anomaly detection module 6, a hidden danger classification module 7, and a spatial positioning module 8.

[0028] Reference Figure 2The infrared image acquisition module 1 includes a thermal infrared imaging unit and an image sequence generation unit. The thermal infrared imaging unit is used to acquire thermal radiation information of the roadbed surface in the long-wave infrared band of 8μm to 14μm. This band corresponds to the peak thermal radiation within the surface temperature range and has good atmospheric transmittance and temperature sensitivity. The thermal infrared imaging unit uses an uncooled focal plane detector with a resolution of 640×480 pixels or higher and a temperature resolution better than 0.05℃, enabling it to accurately capture minute temperature differences on the roadbed surface. The thermal infrared imaging unit is mounted on a mobile platform, maintaining a vertical distance of 1.0m to 2.0m from the roadbed surface. This distance range ensures sufficient field-of-view coverage while achieving good spatial resolution, resulting in a roadbed surface size of 2mm to 5mm corresponding to a single pixel, meeting the accuracy requirements for hazard detection.

[0029] The image sequence generation unit is connected to the thermal infrared imaging unit, continuously acquiring multiple frames of infrared thermal imaging images at preset time intervals to generate an infrared thermal imaging image sequence. The preset time interval is set to 5s to 30s, preferably 10s, which captures the dynamic changes in heat conduction without generating excessive redundant data. The number of image frames acquired in a single detection is 10 to 50 frames, preferably 30 frames, with a total acquisition time of 5min to 15min. By acquiring image sequences rather than single frames, the evolution of temperature over time can be analyzed, which is crucial for determining the depth and nature of potential hazards. Defects such as cavities and water accumulation within the roadbed alter local heat conduction characteristics. Under external thermal excitation, the temperature response curve of the defective area differs significantly from that of the normal area. These differences can be identified by analyzing the time-series temperature curves.

[0030] The visible light image acquisition module 2 is used to acquire a visible light image sequence corresponding to the infrared thermal imaging image sequence. The visible light image acquisition module 2 uses a high-resolution color camera with a resolution of no less than 1920×1080 pixels, and is triggered synchronously with the infrared image acquisition module 1 to ensure strict temporal and spatial correspondence between the two images. The visible light images provide visible feature information such as roadbed texture, cracks, and stains. This information complements the infrared temperature information, improving the accuracy of hazard identification.

[0031] The environmental parameter acquisition module 3 is connected to the infrared image acquisition module 1 and the visible light image acquisition module 2 to collect ambient temperature, humidity, solar radiation intensity, and wind speed parameters in real time. The environmental parameter acquisition module 3 includes a temperature sensor, a humidity sensor, a solar radiation intensity sensor, and a wind speed sensor. The temperature sensor has a measurement range of -40℃ to 80℃ with an accuracy of ±0.2℃; the humidity sensor has a measurement range of 0% to 100%RH with an accuracy of ±2%RH; the solar radiation intensity sensor has a measurement range of 0 to 2000W / m² with an accuracy of ±5%; and the wind speed sensor has a measurement range of 0 to 30m / s with an accuracy of ±0.3m / s. These environmental parameters have a significant impact on the infrared imaging results. Solar radiation causes the roadbed surface temperature to rise and become unevenly distributed. Ambient temperature affects the driving force of heat conduction, humidity affects the thermal properties of materials, and wind speed affects surface convective heat transfer. These effects must be eliminated through environmental correction.

[0032] Reference Figure 3 The image fusion module 4 is connected to the infrared image acquisition module 1 and the visible light image acquisition module 2, and is used to perform feature-level fusion of the infrared thermal imaging image sequence and the visible light image sequence based on the attention mechanism to generate a fused image sequence. The image fusion module 4 includes a feature extraction unit, an attention weight calculation unit, and a feature fusion unit.

[0033] The feature extraction unit performs multi-scale feature extraction on both infrared thermal imaging image sequences and visible light image sequences. Feature extraction employs a convolutional neural network, with a network structure including multiple convolutional layers, pooling layers, and batch normalization layers. For infrared images, the extracted features primarily reflect temperature distribution, temperature gradients, and hotspot features; for visible light images, the extracted features primarily reflect edge, texture, and structural information. Multi-scale feature extraction is achieved through convolutional kernels of different sizes and network layers of different depths, enabling the simultaneous capture of local details and global structural information. Specifically, the first-scale feature extraction uses a 3×3 convolutional kernel with a stride of 1 to extract high-resolution detail features; the second-scale feature extraction uses a 5×5 convolutional kernel with a stride of 2 to extract medium-resolution structural features; and the third-scale feature extraction uses a 7×7 convolutional kernel with a stride of 4 to extract low-resolution global features.

[0034] The attention weight calculation unit is connected to the feature extraction unit, calculating infrared and visible light feature weights based on multi-scale features. The core idea of ​​the attention mechanism is to allow the network to automatically learn to focus on important features while ignoring unimportant ones. For roadbed hazard detection, areas with abnormal temperatures are the most important, therefore infrared features should receive higher weights in these areas; while in normal areas, visible light features provide richer texture information and should receive higher weights. The attention weight calculation employs a combination of channel attention and spatial attention mechanisms.

[0035] The channel attention mechanism calculates the importance of each feature channel. First, global average pooling and global max pooling are performed on the feature map, resulting in two one-dimensional vectors. These two vectors are then input into a shared fully connected layer, and the outputs are summed and passed through a sigmoid activation function to obtain the channel attention weight vector. Each element of the channel attention weight vector corresponds to the weight of a feature channel, with a value ranging from 0 to 1. The channel attention weight vector is then multiplied channel-by-channel with the original feature map to achieve channel-level feature weighting.

[0036] The spatial attention mechanism calculates the importance of each spatial location in the feature map. First, max pooling and average pooling are performed along the channel dimension of the feature map to obtain two single-channel feature maps. These two feature maps are then concatenated and fed into a 7×7 convolutional layer, followed by a sigmoid activation function to obtain the spatial attention weight map. Each pixel in the spatial attention weight map corresponds to a weight at the corresponding location in the original feature map, with values ​​ranging from 0 to 1. The spatial attention weight map is then multiplied pixel-by-pixel with the original feature map to achieve spatial-level feature weighting.

[0037] For infrared and visible light features, their channel attention weights and spatial attention weights are calculated separately, and then the two weights are multiplied to obtain the comprehensive attention weight. In regions with abnormal temperatures, the attention weight of infrared features is generally higher than that of visible light features; in normal regions, the attention weight of visible light features is generally higher than that of infrared features. This adaptive weight allocation mechanism can fully utilize the most valuable information in different regions.

[0038] The feature fusion unit is connected to the attention weight calculation unit. It performs weighted fusion of multi-scale features based on infrared and visible light feature weights to generate a fused image sequence. Feature fusion uses a weighted summation method; for each spatial location and each feature channel, the fused feature equals the infrared feature multiplied by the infrared weight plus the visible light feature multiplied by the visible light weight. To maintain consistency across different scales, feature maps at different scales need to be upsampled or downsampled to the same spatial resolution before fusion. The fused feature map is then passed through a convolutional layer and an activation function to generate the final fused image. The fused image contains both temperature information from the infrared image and texture information from the visible light image, providing a richer information foundation for subsequent hazard detection.

[0039] Reference Figure 4 The environmental correction module 5 is connected to the environmental parameter acquisition module 3 and the image fusion module 4. It is used to adaptively correct the fused image sequence based on environmental temperature, humidity, solar radiation intensity, and wind speed parameters to generate a corrected image sequence. The environmental correction module 5 includes a radiation correction unit, a temperature and humidity correction unit, and a convection correction unit.

[0040] The radiometric correction unit is used to perform radiometric compensation on the fused image sequence based on solar radiation intensity. Solar radiation is a major factor affecting the temperature distribution of the roadbed surface, especially under daytime illumination conditions. Direct and reflected sunlight can cause a significant increase in roadbed surface temperature, and due to differences in the geometry and materials of the roadbed surface, the temperature distribution is extremely uneven. This temperature distribution caused by solar radiation can mask temperature anomalies caused by internal hazards, and the influence of solar radiation must be eliminated through radiometric correction.

[0041] Radiation correction employs a physical model-based approach. The solar radiation received by the roadbed surface includes direct radiation and sky-scattered radiation; the radiative component of the surface temperature can be expressed as:

[0042] ,

[0043] in, The temperature increment caused by radiation (in °C). The reference ambient temperature (in °C) is used. The solar radiation absorptivity of the roadbed surface (dimensionless, typical value is 0.6 to 0.9). Solar radiation intensity (unit: W / m²) 2 ), Surface emissivity (dimensionless, typically 0.85 to 0.95). The Stefan-Boltzmann constant (with a value of 5.67 × 10⁻⁶) is given. -8 W / (m 2 ·K 4 ), where 273.15 is the conversion constant between Celsius and Kelvin temperatures. This formula is established based on the principle of energy balance. The left side represents the temperature increment caused by radiation, and the first term on the right side is the reference ambient temperature, while the second term is the contribution of solar radiation energy to temperature conversion.

[0044] During radiometric correction, the temperature value of each pixel in the fused image is subtracted. The corrected temperature value is obtained. Since the solar radiation intensity may vary at different locations on the roadbed surface (e.g., shaded and illuminated areas), the actual radiation intensity at each location needs to be calculated based on the solar altitude and azimuth angles. Furthermore, differences in the roadbed surface material can lead to variations in absorptivity. and emissivity Unlike other systems, this one adaptively adjusts these parameters based on the material type identified by the visible light image. Preferably, in practical applications, for common concrete subgrades, The value is 0.65. The value is 0.90; for crushed stone track bed, The value is 0.75. The value is 0.85.

[0045] The temperature and humidity correction unit is connected to the radiation correction unit, and performs temperature and humidity effect elimination on the radiation-compensated image sequence based on ambient temperature and humidity parameters. Ambient temperature and humidity mainly affect the heat conduction process and the thermal properties of materials. Ambient temperature determines the temperature gradient between the roadbed surface and interior, and is the driving force for heat conduction; humidity affects the moisture content of materials, and changes in moisture content will significantly change the thermal conductivity and specific heat capacity of materials.

[0046] Temperature and humidity correction employs a normalization method. First, standard reference conditions are established, defining the standard ambient temperature. 20℃, standard relative humidity It is 50%. Then, based on the actual ambient temperature... and relative humidity Calculate the correction factor:

[0047] ,

[0048] ,

[0049] in, This is the temperature correction factor (dimensionless). Humidity correction factor (dimensionless). This is the temperature effect coefficient (unit: °C⁻¹, typical value: 0.005 to 0.015). These are humidity influence coefficients (in %⁻¹, typically ranging from 0.002 to 0.008). These coefficients were obtained by fitting a large amount of experimental data and reflect the combined effects of changes in ambient temperature and humidity on heat conduction and thermal imaging.

[0050] During calibration, the radiation-compensated temperature value is divided by... and The product of these values ​​yields a temperature value normalized to standard conditions. This normalization process ensures comparability of test results under different environmental conditions. Preferably, for concrete subgrades, The value is 0.010. The value is 0.005; for soil subgrade, The value is 0.012. The value is 0.006.

[0051] The convection correction unit is connected to the temperature and humidity correction unit. Based on wind speed parameters, it performs convective heat transfer correction on the image sequence after eliminating the influence of temperature and humidity, generating a corrected image sequence. Wind speed affects the convective heat transfer coefficient of the roadbed surface; the higher the wind speed, the stronger the convective heat transfer, and the closer the surface temperature is to the ambient temperature. Under windy conditions, the temperature distribution on the roadbed surface is smoothed, reducing temperature differences caused by internal hazards and lowering detection sensitivity.

[0052] Convection correction is based on the calculation of the convective heat transfer coefficient. The convective heat transfer coefficient of the roadbed surface... With wind speed The relationship can be represented as:

[0053] ,

[0054] in, The convective heat transfer coefficient is expressed in W / (m²·K). Wind speed (unit: m / s) The natural convection heat transfer coefficient (dimensionless, typically 5 to 10). The forced convection coefficient (dimensionless, typically 3 to 6) is represented by an exponent of 0.8, which is an empirical value reflecting the characteristics of the turbulent boundary layer. When the wind speed is 0, equal This represents natural convection; when wind speed increases, Increased volume leads to enhanced convective heat transfer.

[0055] Convection heat transfer reduces the anomalous magnitude of surface temperature; the reduction can be expressed as:

[0056] ,

[0057] in, The temperature anomaly amplitude (in °C) is taken into account after considering the effects of convection. The temperature anomaly amplitude (in °C) under non-convection conditions. The equivalent thermal conductivity coefficient inside the roadbed (unit: W / (m²·K), typical value: 1 to 3) reflects the controlling effect of internal heat conduction on surface temperature. This formula indicates that when... Very light (no wind or a gentle breeze). near Surface temperature anomalies are mainly determined by internal heat conduction; when When it is very strong (strong wind) much smaller The abnormal surface temperature was smoothed out by convection heat transfer.

[0058] During convection correction, the temperature outlier after temperature and humidity correction is multiplied by... This yields the temperature anomaly value restored to a non-convective condition. This eliminates the influence of wind speed variations on detection sensitivity, making detection results under different wind speed conditions comparable. Preferably, in practical applications, The value is 8. The value is 4.5. The value is set to 2. After three steps of radiation correction, temperature and humidity correction, and convection correction, the corrected image sequence eliminates the influence of environmental factors. The temperature distribution mainly reflects the differences in thermal properties inside the roadbed, providing a reliable data basis for accurately detecting hidden dangers.

[0059] Reference Figure 5 The thermal anomaly detection module 6 is connected to the environmental correction module 5 and is used to extract spatiotemporal features from the corrected image sequence to detect temperature anomaly areas inside the roadbed. The thermal anomaly detection module 6 includes a spatial feature extraction unit, a temporal feature extraction unit, and an anomaly area localization unit.

[0060] The spatial feature extraction unit extracts spatial domain features from single-frame images of the corrected image sequence. These features reflect the spatial distribution pattern of temperature, including temperature gradients, texture features, and edge features. Spatial feature extraction employs a convolutional neural network (CNN), with the input being a single-frame temperature image and the output being a high-dimensional feature vector. The network consists of multiple convolutional layers, pooling layers, and fully connected layers. The first convolutional layer uses 64 3×3 kernels to extract basic temperature gradient features; the second uses 128 3×3 kernels to extract combined temperature pattern features; and the third uses 256 3×3 kernels to extract higher-level abstract features. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, then a 2×2 max-pooling layer. Finally, a fully connected layer flattens the feature map into a 1024-dimensional feature vector.

[0061] For temperature anomaly detection, important spatial features include local temperature extrema, the direction and magnitude of temperature gradients, and the uniformity of temperature distribution. Voids within the roadbed will form circular or elliptical regions with higher or lower temperatures on the surface, with distinct boundaries; water accumulation will create irregularly shaped regions with lower temperatures; and material inhomogeneity will create strip-shaped regions with varying temperature gradients. Convolutional neural networks can automatically learn these pattern features without the need for manually designed feature extraction algorithms.

[0062] The temporal feature extraction unit is connected to the spatial feature extraction unit to extract temporal evolution features from multiple frames of the corrected image sequence. These temporal evolution features reflect the temperature variation over time and are crucial for determining the depth and nature of potential hazards. Temporal feature extraction employs a Long Short-Term Memory (LSTM) network, a recurrent neural network specifically designed for processing sequential data, capable of learning long-term dependencies.

[0063] The input to an LSTM network is a sequence of feature vectors output by a spatial feature extraction unit, with one feature vector corresponding to each frame of the image. An LSTM network includes an input gate, a forget gate, and an output gate, enabling it to selectively remember or forget information from the sequence. For roadbed hazard detection, the LSTM network learns the temporal evolution pattern of temperature, such as the heating rate, cooling rate, and the time of temperature peak occurrence. These temporal features are closely related to the depth and thermal properties of the hazard.

[0064] The depth of internal hazards in the roadbed can be estimated using the phase delay of the temperature time series curve. Surface hazards have a fast temperature response and an early temperature peak; deep hazards have a slow temperature response and a late temperature peak. The relationship between phase delay and depth can be described using heat conduction theory.

[0065] ,

[0066] in, The depth of the hazard (in meters). Thermal diffusivity of the roadbed material (unit: m² / s, typical value: 5 × 10⁻) 7 Up to 2×10⁻ 6 ), This represents the time delay (in seconds) between the occurrence of the temperature peak and the start of surface heating. This formula is based on the analytical solution of periodic thermal excitation in a semi-infinite solid and reflects the phase delay characteristics of heat wave propagation in the material.

[0067] The LSTM network outputs a high-dimensional temporal feature vector with a dimension of 512. This feature vector integrates temporal temperature variation information, providing input for subsequent anomaly detection and classification.

[0068] The anomaly region localization unit is connected to the spatial feature extraction unit and the temporal feature extraction unit to determine the location and range of temperature anomaly regions based on spatial domain features and temporal evolution features. The anomaly region localization employs a fully convolutional network (FCN), an end-to-end semantic segmentation network capable of classifying each pixel in an image.

[0069] The input to the FCN network is a concatenation of spatial and temporal feature vectors, and the output is a binary segmentation map of the same size as the input image, where pixels in abnormal regions are labeled as 1 and pixels in normal regions are labeled as 0. The FCN network restores spatial resolution through upsampling and skip connections. Upsampling is achieved using transposed convolution, and skip connections fuse low-level features with high-level features, thus preserving both detailed and semantic information.

[0070] Anomaly region localization also employs post-processing steps to improve detection accuracy. First, morphological operations are performed on the segmentation image, including opening and closing operations, to remove isolated noise points and fill small holes. Then, connected regions are calculated, clustering adjacent anomaly pixels into a single anomaly region. Finally, anomaly regions are filtered based on their area; regions smaller than a preset threshold are considered noise and discarded. The threshold is set between 100 and 500 pixels, with 300 pixels being preferred. This post-processing results in highly accurate anomaly regions with clear boundaries, providing reliable input for subsequent classification and localization.

[0071] Reference Figure 6 The hazard classification module 7 is connected to the thermal anomaly detection module 6 and is used to determine the type and severity of roadbed hazards based on the morphological characteristics, temperature distribution characteristics, and temporal evolution characteristics of the temperature anomaly area. The hazard classification module 7 includes a morphological characteristic calculation unit, a temperature characteristic calculation unit, a hazard type determination unit, and a severity assessment unit.

[0072] The morphological feature calculation unit calculates the area, aspect ratio, and connectivity of temperature anomaly regions. Area reflects the scale of the hazard, aspect ratio reflects the shape of the hazard, and connectivity reflects the continuity of the hazard. For each anomaly region, the area is calculated by multiplying the number of pixels contained in the region by the actual area corresponding to a single pixel. The aspect ratio is calculated as the ratio of the longer side to the shorter side of the smallest bounding rectangle of the anomaly region. Connectivity is evaluated by calculating the number of connected components; a higher number of connected components indicates a more dispersed hazard.

[0073] The temperature feature calculation unit calculates the mean temperature, variance temperature, and gradient temperature of the temperature anomaly region. The mean temperature reflects the overall intensity of the anomaly, the variance temperature reflects the uniformity of the temperature distribution, and the gradient temperature reflects the steepness of the anomaly boundary. The mean temperature is calculated as the arithmetic mean of the temperature values ​​of all pixels in the anomaly region; the variance temperature is calculated as the variance of the temperature values ​​in the anomaly region; and the gradient temperature is calculated as the average gradient magnitude at the boundary of the anomaly region, using the Sobel operator.

[0074] The hazard type determination unit is connected to the morphological feature calculation unit and the temperature feature calculation unit to determine the type of roadbed hazard based on morphological and temperature features. Common types of roadbed hazards include cavities, water accumulation, and material inhomogeneity, and different types of hazards have different characteristic patterns.

[0075] The characteristics of cavities are approximately circular or elliptical shapes with an aspect ratio close to 1, an average temperature higher (daytime) or lower (nighttime) than the surrounding area, a relatively uniform temperature distribution, and a larger temperature gradient at the boundaries. This is because the cavity is filled with air, whose thermal conductivity is much lower than that of soil or concrete, resulting in impeded heat conduction. During the day, the surface temperature rises when heated by the sun, and at night, the surface temperature decreases when cooled. The criteria for detection are: aspect ratio between 0.5 and 2.0, temperature variance less than 2.0℃², and temperature gradient greater than 0.5℃ / pixel.

[0076] The characteristics of water accumulation hazards include irregular shape, potentially large aspect ratio, lower average temperature than the surrounding area, uneven temperature distribution, and a moderate temperature gradient at the boundaries. This is because water has a high specific heat capacity and thermal inertia, heats up slowly during the day and cools down slowly at night, and water infiltration leads to irregular shapes. The criteria for detection are: aspect ratio greater than 2.0, temperature variance greater than 3.0℃², and temperature gradient within the range of 0.2℃ / pixel to 0.5℃ / pixel.

[0077] The characteristics of material inhomogeneity include striped or irregular shapes, and a gradient temperature distribution with a larger temperature gradient within the region. This is because different material densities or types lead to different thermophysical parameters. The criteria for determination are: aspect ratio greater than 3.0 or high shape complexity, temperature variance greater than 4.0℃², and temperature gradient within the region greater than 0.3℃ / pixel.

[0078] The hazard type determination employs a decision tree method, making step-by-step judgments based on the aforementioned characteristics and thresholds. First, hazards are categorized into near-circular and non-circular types based on aspect ratio. For near-circular hazards, they are classified as voids based on temperature variance and temperature gradient; for non-circular hazards, they are classified as water accumulation or material inhomogeneity based on the mean temperature. The parameters of the decision tree are obtained through training on a large number of labeled samples and can be adjusted empirically in practical applications.

[0079] The severity assessment unit is connected to the hazard type determination unit, and the severity of roadbed hazards is determined based on the area proportion of temperature anomalies and the magnitude of temperature differences. Severity is divided into three levels: mild, moderate, and severe. The area proportion is calculated by dividing the area of ​​the anomaly region by the total area of ​​the monitored area; the magnitude of temperature differences is calculated as the absolute value of the difference between the average temperature of the anomaly region and the average temperature of the normal region.

[0080] The criteria for classifying a minor hazard are: an area of ​​less than 5% and a temperature difference of less than 2.0℃, or an area of ​​5% to 10% and a temperature difference of less than 1.0℃. The criteria for classifying a moderate hazard are: an area of ​​5% to 15% and a temperature difference of 2.0℃ to 4.0℃, or an area of ​​10% to 20% and a temperature difference of 1.0℃ to 2.0℃. The criteria for classifying a severe hazard are: an area of ​​more than 15% and a temperature difference of more than 4.0℃, or an area of ​​more than 20% and a temperature difference of more than 2.0℃. These thresholds are based on extensive field measurement data and maintenance case statistics, taking into account both the scale and intensity of the hazard.

[0081] Reference Figure 7 The spatial positioning module 8 is connected to the hazard classification module 7 and is used to map temperature anomaly areas to the actual spatial location of the roadbed, generating a hazard distribution map. The spatial positioning module 8 includes a coordinate mapping unit, a depth estimation unit, and a visualization unit.

[0082] The coordinate mapping unit maps the location of the temperature anomaly area in the image coordinate system to the roadbed spatial coordinate system. The image coordinate system is a two-dimensional pixel coordinate system with the top left corner of the image as the origin; the roadbed spatial coordinate system is a three-dimensional real coordinate system with a reference point as the origin, and the unit is meters. Coordinate mapping requires knowledge of the camera's intrinsic parameters (focal length, principal point coordinates), extrinsic parameters (position, attitude), and the geometric relationship between the camera and the roadbed.

[0083] Coordinate mapping employs a perspective transformation model. For pixels in the image... The corresponding actual coordinates of the roadbed surface It can be calculated using the following system of equations:

[0084] ,

[0085] in, This is the camera intrinsic parameter matrix (3×3 matrix). It is a rotation matrix (3×3 matrix). It is a translation vector (3×1 vector). Let be the camera extrinsic parameter matrix (3×4 matrix). Since the roadbed surface can be approximated as a plane, let be... Then the above formula simplifies to:

[0086] ,

[0087] in, It is a 3×3 homography matrix. By solving... The inverse matrix of the homography matrix can be used to calculate the actual coordinates from the image coordinates. The calibration process is predetermined. During calibration, a calibration plate at a known location is placed on the roadbed surface, and images of the calibration plate are taken. The coordinates of the image are then matched with the actual coordinates to obtain the desired result. Calibration accuracy directly affects positioning accuracy. During calibration, it is essential to ensure that the calibration plate covers different locations within the detection area, with at least four non-collinear calibration points.

[0088] The depth estimation unit is connected to the coordinate mapping unit, and estimates the depth location of potential hazards based on the temperature time-series curve of the temperature anomaly region. Depth estimation employs the aforementioned phase delay method, calculating the depth by analyzing the occurrence time of temperature peaks. Specifically, the peak moment is first identified from the time-series temperature curve. Peak identification uses a combination of first and second derivatives; points where the first derivative is zero and the second derivative is negative are identified as peak points. Then, the delay between the peak moment and the heating start moment is calculated, and this value is substituted into the depth calculation formula to obtain the depth value.

[0089] Due to the thermal diffusivity of the roadbed material They may differ; the system selects the appropriate method based on the roadbed type. Value. For concrete subgrade, The value is 8 × 10⁻ 7 m² / s; for crushed stone track bed The value is 1.2 × 10⁻ 6 m² / s; for soil subgrade The value is 6 × 10⁻ 7 m² / s. These parameters were determined through literature review and field measurements. The accuracy of depth estimation is affected by various factors, including the accuracy of temperature measurement, the sampling frequency of the time series curve, and the accuracy of material parameters. Typical depth estimation errors range from ±0.1m to ±0.3m.

[0090] The visualization unit connects with the coordinate mapping unit and the depth estimation unit to generate a hazard distribution map containing information on the location, type, and depth of hazards. The visualization unit uses a 3D graphics engine to generate plan views, cross-sectional views, and 3D stereoscopic views. The plan view displays the hazard distribution on the roadbed surface from a top-down perspective. Each hazard is marked with a different colored graphic, where color represents the hazard type (e.g., red for cavities, blue for water accumulation, and yellow for uneven material distribution), and the graphic size represents the hazard scale. Depth values ​​and severity are also labeled within the graphic. The cross-sectional view displays the depth distribution of hazards along the longitudinal or transverse direction of the roadbed, with the horizontal axis representing location and the vertical axis representing depth. Hazards are represented by rectangles or ellipses. The 3D stereoscopic view models the roadbed as a 3D solid, with hazards embedded within the roadbed using semi-transparent geometry. Users can rotate and zoom the viewpoint using the mouse to observe the hazard distribution from different angles.

[0091] The hazard distribution map supports exporting to multiple formats, including PNG images, PDF documents, and CAD formats, facilitating integration with other engineering software. The hazard distribution map can also be integrated with the roadbed's BIM (Building Information Modeling) data, overlaying hazard information onto the roadbed's 3D model, providing digital support for maintenance decisions and construction management.

[0092] The system of this invention operates according to the following procedure in practical applications. First, the operator mounts the infrared image acquisition module 1 and the visible light image acquisition module 2 onto a mobile platform, which can be a rail vehicle, a drone, or a ground vehicle. An environmental parameter acquisition module 3 is mounted on the platform and works synchronously with the image acquisition modules. During detection, the mobile platform moves along the roadbed, the image acquisition modules continuously acquire image sequences, and the environmental parameter acquisition module records environmental parameters in real time.

[0093] After acquisition, image data and environmental data are transmitted to the backend processing system. Image fusion module 4 fuses the infrared and visible light image sequences to generate a fused image sequence. Environmental correction module 5 performs environmental correction on the fused image sequence to eliminate the effects of solar radiation, temperature, humidity, and wind speed, generating a corrected image sequence. Thermal anomaly detection module 6 detects temperature anomaly areas from the corrected image sequence and extracts spatiotemporal features. Hazard classification module 7 classifies the detected anomaly areas to determine the hazard type and severity. Spatial positioning module 8 maps the anomaly areas to their actual spatial locations, estimates their depth, and generates a hazard distribution map.

[0094] The entire process is highly automated, typically generating a hazard distribution map within 10 to 30 minutes after data acquisition. The results are displayed through a user interface, allowing users to view detailed information for each hazard, including location, type, depth, and severity. For severe hazards, the system automatically generates an early warning report to alert maintenance personnel for timely intervention.

[0095] This invention achieves high-precision, high-efficiency, and high-reliability intelligent identification of hidden dangers in high-speed railway subgrades through the organic combination of multiple innovative technologies. Image fusion technology fully utilizes the complementary information of infrared and visible light images, improving the accuracy of hidden danger identification; environmental correction technology eliminates the influence of environmental factors, ensuring the consistency and repeatability of detection results; time-series analysis technology estimates the depth and nature of hidden dangers by analyzing the dynamic process of heat conduction; intelligent classification technology enables automatic determination of hidden danger type and severity, improving detection efficiency; and precise positioning technology generates intuitive three-dimensional visualization results, providing a scientific basis for maintenance decisions.

[0096] Compared with existing technologies, this invention has significant technical advantages. First, it boasts high detection accuracy, achieving a detection rate of over 95% for hazards with a diameter greater than 30cm and a depth within 1.5m, far exceeding the 60% to 70% of existing technologies. Second, it offers high detection efficiency; the mobile platform travels at speeds of 5km / h to 10km / h, enabling continuous roadbed detection with a detection range of 5km to 10km per hour. Third, it has a wide applicability; the system can perform detection under various weather conditions (sunny, cloudy, overcast) and at different times (daytime, nighttime), with consistent results. Fourth, it features a high degree of automation; the entire process from data acquisition to result output is automated, reducing manual intervention and improving the objectivity and reliability of the detection. Fifth, it provides excellent visualization; the generated hazard distribution map is intuitive and clear, facilitating understanding and use by maintenance personnel.

[0097] This invention is particularly suitable for the acceptance testing of newly built high-speed railway lines and the periodic inspection of operational lines, and has unique advantages for subgrade inspection in seasonal freeze-thaw zones and karst areas. In seasonal freeze-thaw zones, subgrades are prone to internal cracks and cavities due to freeze-thaw cycles; the time-series analysis technology of this invention can detect the development and changes of these hidden dangers. In karst areas, there may be sinkholes or underground rivers beneath the subgrade, leading to a decrease in subgrade stability; the depth estimation technology of this invention can detect these deep-seated hidden dangers. Furthermore, this invention can also be applied to the inspection of other transportation infrastructure, such as highway subgrades, airport runways, and bridges, and has broad application prospects.

[0098] A high-speed railway operating company used the system of this invention to inspect a 20km long section of roadbed. The inspection took place on a sunny morning from 9:00 AM to 11:00 AM, with an ambient temperature of 25℃, relative humidity of 45%, solar radiation intensity of 800W / m², and wind speed of 3m / s. The mobile platform was a rail vehicle traveling at 8km / h. The infrared image acquisition module and the visible light image acquisition module worked synchronously, with the image sequence generation unit acquiring one frame every 10 seconds. 30 frames were acquired at each inspection point, for a total acquisition time of 5 minutes.

[0099] After data acquisition, the system automatically processes the data. The image fusion module fuses infrared and visible light image sequences based on an attention mechanism, resulting in a fused image that includes both temperature and texture information. The environmental correction module performs radiation, temperature and humidity, and convection corrections based on environmental parameters, eliminating the influence of environmental factors. The thermal anomaly detection module detects 135 temperature anomaly areas from the corrected image sequence. The hazard classification module categorizes these anomaly areas, identifying 48 cavities, 32 water accumulation hazards, and 55 material inhomogeneity hazards. Among these, 12 are severe hazards, 45 are moderate hazards, and 78 are minor hazards. The spatial positioning module generates a hazard distribution map, marking the location, type, depth, and severity of each hazard.

[0100] Comparison of the test results with subsequent manual verification and maintenance records revealed that the detection accuracy of the system was 96%, with all severe hazards detected, a 95% detection rate for moderate hazards, and a 92% detection rate for minor hazards. The false detection rate was only 4%, mainly due to some material joints being misjudged as minor material unevenness hazards. In terms of detection efficiency, the inspection of 20km of roadbed was completed within 2.5 hours, while traditional ground-penetrating radar inspection takes 2 days. Maintenance personnel developed a maintenance plan based on the hazard distribution map, grouting reinforcement was carried out on 12 severe hazards, monitoring was conducted on 45 moderate hazards, and 78 minor hazards were temporarily left untreated but included in subsequent inspection plans. The stability of the roadbed was significantly improved after the treatment, and no further settlement or deformation incidents occurred.

[0101] Practical applications have shown that the intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis of this invention has significant advantages such as high detection accuracy, high detection efficiency, wide applicability, high degree of automation, and good visualization effects. It can effectively identify hidden dangers such as cavities, water accumulation, and uneven material distribution within the subgrade, providing strong technical support for the safe operation of high-speed railways. The system has been applied on multiple high-speed railway lines, with a cumulative detection mileage exceeding 1000km, detecting more than 5000 hidden dangers of various types. It provides a scientific basis for preventive maintenance by railway operators and significantly reduces the risk of safety accidents caused by subgrade hidden dangers.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-speed railway subgrade hidden danger intelligent identification system based on infrared image analysis, characterized in that, include: Infrared image acquisition module, used to acquire infrared thermal imaging image sequences of the roadbed; The visible light image acquisition module is used to acquire a visible light image sequence corresponding to the infrared thermal imaging image sequence; An environmental parameter acquisition module is connected to the infrared image acquisition module and the visible light image acquisition module, and is used to acquire environmental temperature, humidity, solar radiation intensity and wind speed parameters. An image fusion module, connected to the infrared image acquisition module and the visible light image acquisition module, is used to perform feature-level fusion of the infrared thermal imaging image sequence and the visible light image sequence based on an attention mechanism to generate a fused image sequence. An environmental correction module, connected to the environmental parameter acquisition module and the image fusion module, is used to adaptively correct the fused image sequence based on the environmental temperature, humidity, solar radiation intensity and wind speed parameters, and generate a corrected image sequence. A thermal anomaly detection module, connected to the environmental correction module, is used to extract spatiotemporal features from the corrected image sequence and detect temperature anomaly areas inside the roadbed. The hazard classification module, connected to the thermal anomaly detection module, is used to determine the type and severity of roadbed hazards based on the morphological characteristics, temperature distribution characteristics, and temporal evolution characteristics of the temperature anomaly area. The spatial positioning module, connected to the hazard classification module, is used to map the temperature anomaly area to the actual spatial location of the roadbed and generate a hazard distribution map.

2. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The infrared image acquisition module includes: The thermal infrared imaging unit is used to acquire thermal radiation information of the roadbed surface in the 8μm to 14μm band. An image sequence generation unit, connected to the thermal infrared imaging unit, is used to continuously acquire multiple frames of infrared thermal imaging images at preset time intervals to generate the infrared thermal imaging image sequence.

3. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The image fusion module includes: A feature extraction unit is used to extract multi-scale features from the infrared thermal imaging image sequence and the visible light image sequence, respectively. An attention weight calculation unit, connected to the feature extraction unit, is used to calculate infrared feature weights and visible light feature weights based on the multi-scale features. The feature fusion unit, connected to the attention weight calculation unit, is used to perform weighted fusion of the multi-scale features according to the infrared feature weights and the visible light feature weights to generate the fused image sequence.

4. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The environmental correction module includes: A radiation correction unit is used to perform radiation compensation on the fused image sequence based on the solar radiation intensity. A temperature and humidity correction unit, connected to the radiation correction unit, is used to eliminate the influence of temperature and humidity on the radiation-compensated image sequence based on the ambient temperature and humidity parameters. A convection correction unit, connected to the temperature and humidity correction unit, is used to perform convective heat transfer correction on the image sequence after the influence of temperature and humidity has been eliminated based on the wind speed parameters, and generate the corrected image sequence.

5. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The thermal anomaly detection module includes: A spatial feature extraction unit is used to extract spatial domain features from single-frame images of the corrected image sequence; A temporal feature extraction unit, connected to the spatial feature extraction unit, is used to extract temporal evolution features from multiple frames of the corrected image sequence; An abnormal region localization unit, connected to the spatial feature extraction unit and the temporal feature extraction unit, is used to determine the location and range of the temperature anomaly region based on the spatial domain features and the temporal evolution features.

6. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The hazard classification module includes: A morphological feature calculation unit is used to calculate the area, aspect ratio, and connectivity of the temperature anomaly region. A temperature characteristic calculation unit is used to calculate the temperature mean, temperature variance, and temperature gradient of the temperature anomaly region. The hazard type determination unit is connected to the morphological feature calculation unit and the temperature feature calculation unit, and is used to determine the type of roadbed hazard based on the morphological features and the temperature features. The types include cavities, water accumulation and uneven material distribution.

7. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 6, characterized in that, The hazard classification module also includes: The severity assessment unit, connected to the hazard type determination unit, is used to determine the severity of roadbed hazards based on the area proportion of the temperature anomaly region and the temperature difference range. The severity includes mild, moderate and severe.

8. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The spatial positioning module includes: A coordinate mapping unit is used to map the position of the temperature anomaly area in the image coordinate system to the roadbed space coordinate system; A depth estimation unit, connected to the coordinate mapping unit, is used to estimate the depth location of potential hazards based on the temperature time-series curve of the temperature anomaly region. The visualization unit, connected to the coordinate mapping unit and the depth estimation unit, is used to generate the hazard distribution map containing hazard location, type, and depth information.

9. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The detector of the thermal infrared imaging unit has a resolution of 640×480 pixels or higher, and the distance between the thermal infrared imaging unit and the roadbed surface is 1.0m to 2.0m.

10. The intelligent identification system for hidden dangers in high-speed railway subgrade based on infrared image analysis according to claim 1, characterized in that, The preset time interval of the image sequence generation unit is 5s to 30s, and the number of image frames acquired in a single detection is 10 to 50 frames.

Images