Traffic accident scene restoration analysis method and system based on three-dimensional reconstruction technology

Abstract

The present application relates to the technical field of three-dimensional reconstruction of traffic accident scene, and especially relates to a traffic accident scene restoration analysis method and system based on three-dimensional reconstruction technology. The method collects multi-view and multi-time sequence images of the accident scene, estimates the initial transmittance and the meteorological degradation intensity index through dark color prior, constructs a multi-scale feature pyramid according to the meteorological degradation intensity index, and extracts meteorological degradation perception features through dynamic gating mask and multi-scale feature pyramid at each scale. The three-dimensional space points are projected onto the image plane, the meteorological degradation perception features obtain continuous feature vectors through bilinear interpolation, the initial transmittance is refined by using joint bilateral filtering, the clean color and the atmospheric attenuation coefficient are calculated and analyzed, the observed color image is synthesized through volume rendering, and a multi-task joint loss function is constructed for optimization. The present application restores and reconstructs the accident scene through the degradation of meteorology and the transmittance of light.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology of traffic accident scenes, and in particular to a method and system for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology. Background Technology

[0002] Rapid and accurate 3D reconstruction of traffic accident scenes is crucial for accident liability determination, forensic identification, and traffic safety research. However, 3D reconstruction of traffic accident scenes occurring under adverse weather conditions such as rain and fog faces severe technical challenges: First, rain and fog significantly enhance atmospheric scattering, causing strong attenuation and scattering of light during propagation. This results in severely reduced contrast, color distortion, and blurred details in the acquired scene images. Traditional 3D reconstruction methods based on structured light and multi-view stereo vision heavily rely on the extraction and matching of image feature points. In degraded images due to rain and fog, the number of feature points extracted decreases sharply, and the matching error rate increases significantly. Second, the rainfall process... Raindrops create dynamic occlusion in front of the lens, appearing as transient and randomly distributed stripes and spots in the image sequence. Existing de-raining algorithms are mostly optimized for single images, lacking a systematic consideration of the accuracy of 3D reconstruction. The local information loss caused by raindrop occlusion will propagate into global geometric distortion during the 3D reconstruction process. Thirdly, depth estimation in rain and fog scenes has systematic biases. Atmospheric scattering causes image contrast to decrease exponentially with scene depth. Vision-based depth estimation networks will produce estimation errors that increase with depth under rain and fog conditions, resulting in significant inaccuracies in scene reconstruction. Therefore, there is an urgent need to develop a traffic accident scene reconstruction and analysis method and system based on 3D reconstruction technology. Summary of the Invention

[0003] To overcome the shortcomings of low accuracy and large depth estimation deviation in 3D reconstruction under adverse weather conditions such as rain and fog, this invention provides a method and system for reconstructing and analyzing traffic accident scenes based on 3D reconstruction technology.

[0004] The technical solution of this invention is: a method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology, comprising the following steps: S1: Acquire static images of the accident site and construct a multi-view image sequence; calculate the atmospheric light vector of the current accident site based on the multi-view image set; calculate the initial transmittance based on the atmospheric light vector; and calculate the meteorological degradation intensity index of the current accident site based on the initial transmittance. S2: Preset scale levels, construct a feature pyramid based on the initial transmittance, and obtain the result feature vector of the accident site through the feature pyramid; calculate the spatial attention map of the current accident site based on the result feature vector, and obtain the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index and the result feature vector; S3: Obtain continuous feature vectors based on real-space coordinates and meteorological degradation perception feature vectors; obtain refined transmittance based on continuous feature vectors; calculate atmospheric attenuation coefficient and clean color image based on refined transmittance. S4: Construct a 3D reconstruction scene based on the volume density and clean color image, obtain the accident meteorological scene based on the atmospheric attenuation coefficient, and generate a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene.

[0005] Preferably, the step of acquiring static image data of the accident scene and constructing a multi-view image sequence, and calculating the atmospheric light vector of the current accident scene based on the multi-view image set, includes: acquiring static image data of the accident scene based on synchronously triggered cameras and monitoring cameras deployed at the accident scene, wherein the static image data is the image pixel intensity value of the accident scene in multiple time frames; constructing a multi-view image set of the current accident scene based on the static image data according to different viewpoints and different time frames, wherein the multi-view image set is the normalized image pixel intensity value of the accident scene in different viewpoints and different time frames; setting a preset brightness threshold; acquiring a dark color map based on the multi-view image set, wherein the dark color map is the fog density in the current image; and calculating the atmospheric light vector of the current accident scene based on the dark color map and the preset brightness threshold.

[0006] Preferably, the step of calculating the initial transmittance based on the atmospheric light vector and calculating the meteorological degradation intensity index of the current accident site based on the initial transmittance includes: calculating the atmospheric light vector of the current accident site based on the dark spectrum and a preset brightness threshold; calculating the initial transmittance of the current accident site based on a multi-view image set and the atmospheric light vector, wherein the initial transmittance is the light transmission situation of the accident site under the current weather conditions; and calculating the meteorological degradation intensity index of the accident site based on the initial transmittance, wherein the meteorological degradation intensity index is the degree of meteorological degradation during the accident occurrence stage, wherein the calculation formulas for the meteorological degradation intensity index and the atmospheric light vector are: ; ; ; ; ; in, The meteorological degradation intensity index. This represents the total number of x-coordinates of the image pixels. This represents the total number of pixels on the y-coordinate of the image. Let be the set of all pixel coordinates. pixel position In time frame The initial transmittance, For camera indexing, used to represent different viewpoints. For adjusting the defogging intensity parameters, In pixel position The central local area, for The pixel position in This is the index for the three color channels. For the first The camera, the first Frame image in pixels passage The pixel intensity value on For the first The camera, the first Frame image in channel Atmospheric light vector, For the first The camera, the first Frame image in pixels The three-channel pixel intensity vector at that location. This is the set of candidate pixels for atmospheric light. This is a dark primary color image. The brightness threshold. The number of elements in the candidate pixel set for atmospheric light. The first pixel in the candidate pixel set for atmospheric light Each element.

[0007] Preferably, the preset scale level, which constructs a feature pyramid based on the initial transmittance and obtains the result feature vector of the accident scene through the feature pyramid, includes: a preset maximum scale level; halving the resolution of the current multi-view pixel set and initial transmittance based on the maximum scale level; constructing a feature pyramid of the current multi-view pixel set and initial transmittance in ascending order of resolution; the feature pyramid consists of multi-view pixel sets and initial transmittance at different scale levels; obtaining the feature vector of the current scale level as the input vector of the current convolutional layer based on the feature pyramid; obtaining the output vector of the current scale level; performing dynamic gating mask calculation based on the output vector of the current scale level and the initial transmittance to obtain the feature vector of the current scale level; using the feature vector of the current scale level as the input vector of the convolutional layer at the next scale level; finally obtaining the feature vector of the maximum scale level in the current feature pyramid; and performing dynamic gating mask calculation based on the feature vector of the maximum scale level to obtain the result feature vector of the maximum scale level; the result feature vector is a low-resolution image vector of the original multi-view image set after convolution and dynamic gating at different scale levels, wherein the calculation formula for the feature vector is: ; ; ; ; in, As the initial input vector, for Output vector of scale-level convolutional layers for Scale-level feature vectors for Scale-level dynamic gating units For the first The camera, the first The three-channel pixel intensity vector map of the frame image. For the first The camera, the first The initial transmittance of the frame image. It is a ReLU nonlinear activation function. For the first The convolutional kernel weight tensor of the layer, For spatial downsampling with a step size of 2, For bilinear upsampling, This is a splicing operation at the channel dimension. For the first The convolution bias vector of the layer, For element-wise multiplication, It is a sigmoid activation function. It is a 1*1 pointwise convolution.

[0008] Preferably, the step of calculating the spatial attention map of the current accident site based on the result feature vector, and obtaining the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index, and the result feature vector includes: obtaining the spatial attention map of the current accident site by point-by-point convolution and transmittance weighting based on the result feature vector, wherein the spatial attention map is the attention weight of different pixel positions in the image; obtaining a secondary result feature vector by point-by-point convolution based on the spatial attention map and the result feature vector, wherein the secondary result feature vector is the feature vector of the result feature vector under the weighting of the spatial attention map; and calculating the meteorological degradation perception feature vector based on the secondary result feature vector and the meteorological degradation intensity index, wherein the meteorological degradation perception feature vector is the feature vector of the secondary result feature vector after adding the meteorological degradation factor, wherein the calculation formula for the meteorological degradation perception feature vector is: ; ; ; in, For the first The camera, the first Spatial attention map of frame image, For element-wise multiplication, It is a sigmoid activation function. It is a 1*1 pointwise convolution. For the resulting feature vector, For the largest scale level, The initial transmittance at the largest scale level. The feature vector of the quadratic result, For the dimension expansion function, Expanded to three dimensions This is a feature vector for sensing meteorological degradation. For scaling modulation coefficients, For translation modulation coefficients, It is a meteorological degradation intensity index.

[0009] Preferably, the step of obtaining continuous feature vectors based on real-world coordinates and meteorological degradation perception feature vectors includes: obtaining the camera's intrinsic and extrinsic parameter matrices according to the camera's user manual; expanding the coordinates of the current pixel position to homogeneous coordinates; projecting the real-world coordinates onto the current camera's pixel coordinate system based on the camera's intrinsic, extrinsic, and homogeneous coordinates and scaling the coordinates according to the maximum scale level of the feature pyramid to obtain the projected coordinates of the current real-world coordinates in the pixel coordinate system; performing bilinear interpolation on the projected coordinates in the meteorological degradation perception feature vector to obtain continuous feature vectors of the projected coordinates, wherein the continuous feature vectors are the vector values ​​of the projected coordinates in the meteorological degradation perception feature vectors; and concatenating the embedded vectors and continuous feature vectors along the channel dimension to obtain enhanced scene encoding, wherein the enhanced scene encoding is a scene characteristic description of the current projection position, and the scene characteristic description is the embedded vector and continuous feature vector of the current position, wherein the calculation formula for the continuous feature vector is: ; ; in, Let be the projection coordinates of the current real-world coordinates in the pixel coordinate system, and be , For the first The camera's intrinsic parameter matrix, For the first The camera's extrinsic matrix, Position in real-world coordinates homogeneous coordinates These are continuous eigenvectors of the projected coordinates. It is a bilinear interpolation function. This is a feature vector for sensing meteorological degradation.

[0010] Preferably, the step of obtaining refined transmittance based on continuous feature vectors and calculating the atmospheric attenuation coefficient and clean color image of the image based on the refined transmittance includes: obtaining a three-dimensional guiding vector by performing a fixed projection compression matrix on the continuous feature vectors and multiplying the matrix; the fixed projection compression matrix is ​​a matrix used to compress the principal components of the continuous feature vectors; the three-dimensional guiding vector is a three-dimensional vector that retains the basic features of the image from the continuous feature vectors; performing linear interpolation on the three-dimensional guiding vector to obtain a three-dimensional guiding vector for all spatial points; performing bilateral filtering on the initial transmittance based on the three-dimensional guiding vector to obtain refined transmittance; the refined transmittance is the detailed transmittance of the current accident scene under the initial transmittance; and calculating the atmospheric attenuation coefficient and clean color image in the current multi-view image set based on the refined transmittance; the atmospheric attenuation coefficient is the degree of light attenuation caused by rain and fog; and the clean color image is the true color of the object surface at the current spatial point under ideal conditions of no rain or fog obstruction and uniform illumination; wherein the formula for calculating the refined transmittance is: ; ; = ; ; in, Position in real-world coordinates The three-dimensional guiding vector, To refine the transmittance, The local region centered on the projected coordinates of the real-world coordinates. The initial transmittance, These are the projection coordinates of real-world coordinates onto the pixel coordinate system. Position in real-world coordinates The three-dimensional guiding vector, The first variance, The second variance, Position in real-world coordinates Atmospheric attenuation coefficient, For the first The camera in the Camera spot coordinates of a time frame, Position in real-world coordinates Clean color image, Atmospheric light vector, The three-channel pixel intensity vector is the projection coordinate. To prevent zero constant, It is a fixed projection compression matrix.

[0011] Preferably, the step of constructing a 3D restored scene based on volume density and clean color image, obtaining an accident meteorological scene based on atmospheric attenuation coefficient, and generating a traffic accident restoration report based on the 3D restored scene and the accident meteorological scene includes: calculating the observed color image of the accident scene based on the clean color image and refined transmittance, wherein the observed color image is the fused scene color after the accident scene restoration; calculating the observation weight of each real space coordinate of the observed color image based on refined transmittance, wherein the observation weight is the probability that light will stop at the current real space coordinate and reflect back to the camera; obtaining the final rendered color image by performing a point-by-point weighted summation based on the observation weight and the observed color image, and using the rendered color image as the 3D restored scene; calculating the average visibility of the scene based on atmospheric attenuation coefficient, and constructing an accident meteorological scene of the current accident scene based on the average visibility of the scene; calculating the loss based on the 3D restored scene and the accident meteorological scene, determining the loss of the current 3D restored scene, and generating a traffic accident restoration report based on the 3D restored scene and the accident meteorological scene; wherein the calculation formulas for the rendered color image, the average visibility of the scene, and the equivalent rainfall intensity are: ; ; ; ; in, Position in real-world coordinates The observed color image, Position in real-world coordinates Clean color image, To refine the transmittance, Atmospheric light vector, For observation weights, The sampling point number, For the first The real-world coordinates of each sampling point To render a color image, The total number of sampling points. The average visibility of the scene. This represents the average atmospheric attenuation coefficient in real-world coordinates.

[0012] Preferably, the step of calculating the loss based on the 3D reconstructed scene and the accident meteorological scene respectively, and determining the loss for the current 3D reconstructed scene, includes: calculating the square of the Euclidean distance between the observed color image and the real color image for each pixel position based on the pixel intensity values ​​of the 3D reconstructed scene and the original image to obtain the color mean square error for each pixel position; and taking the average of the color mean square errors for all pixels, all viewpoints, and all time frames as the first loss function; calculating the square of the Euclidean distance between the refined transmittance and the initial transmittance for each pixel position based on the refined transmittance and the initial transmittance to obtain the transmission mean square error for each pixel position; and taking the average of the color mean square errors for all pixels, all viewpoints, and all time frames as the first loss function; and taking the average of the color mean square errors for all pixels, all viewpoints, and all time frames as the first loss function. The average of the transmission mean square error of each frame is used as the second loss function. The theoretical atmospheric attenuation index is calculated by multiplying the meteorological intensity degradation index and the preset maximum atmospheric attenuation coefficient. The atmospheric mean square error of each pixel is obtained by calculating the square of the Euclidean distance between the average atmospheric attenuation index and the theoretical atmospheric attenuation index. The average of the transmission mean square error of all cameras and all time-series frames is used as the third loss function. The first, second, and third loss functions are added together to obtain the total loss function. The total loss value of the current 3D reconstructed scene is obtained based on the total loss function. A traffic accident scene reconstruction report is generated based on the total loss value, the 3D reconstructed scene, and the accident meteorological scene.

[0013] Preferably, the traffic accident scene reconstruction and analysis system based on three-dimensional reconstruction technology further includes: The meteorological transmission module acquires static images of the accident site and constructs a multi-view image sequence. It calculates the atmospheric light vector of the current accident site based on the multi-view image set, calculates the initial transmittance based on the atmospheric light vector, and calculates the meteorological degradation intensity index of the current accident site based on the initial transmittance. The meteorological degradation module has a preset scale level. It constructs a feature pyramid based on the initial transmittance and obtains the result feature vector of the accident site through the feature pyramid. It calculates the spatial attention map of the current accident site based on the result feature vector and obtains the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index and the result feature vector. The image cleaning module obtains continuous feature vectors based on real-world spatial coordinates and meteorological degradation perception feature vectors; it then obtains refined transmittance based on the continuous feature vectors, and calculates the atmospheric attenuation coefficient and clean color image of the image based on the refined transmittance. The accident reconstruction module constructs a 3D reconstruction scene based on volume density and clean color images, obtains the accident meteorological scene based on the atmospheric attenuation coefficient, and generates a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene.

[0014] The beneficial effects of this invention are: 1. By incorporating the physical laws of atmospheric scattering into the reconstruction process, this invention can effectively distinguish between the intrinsic information of the scene and environmental degradation interference, making the recovered 3D scene closer to the real physical state at the time of the accident, and reducing the defect of traditional defogging algorithms that simultaneously erase physical evidence details when clearing fog; 2. This invention guides the reconstruction process of traffic accident scenes by using the degradation distribution information of the image itself. Under adverse conditions such as significantly reduced visibility and difficulty in extracting effective feature points by traditional visual methods, it can adaptively adjust the processing strategy according to the spatial distribution of degradation degree. It focuses on restoring scene structure in areas with heavy fog and finely preserves texture details in relatively clear areas, thereby improving the applicability and robustness of the method under complex weather conditions. 3. This invention reduces the huge number of parameters and long training time caused by large-scale multilayer perceptrons in traditional neural radiation field methods by deterministic analytical computation and lightweight differentiable filtering operations. It significantly reduces the dependence on high-performance computing hardware, meets the actual business needs of rapid investigation and instant reconstruction of traffic accident scenes, and effectively shortens the time interval between accident scene investigation and subsequent liability determination. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the traffic accident scene reconstruction and analysis method based on the three-dimensional reconstruction technology of this invention. Figure 2 This is a schematic diagram of the traffic accident scene reconstruction and analysis system based on the three-dimensional reconstruction technology of this invention. Detailed Implementation

[0016] The embodiments of the present invention will be described below with reference to the accompanying drawings.

[0017] Example 1: A method for reconstructing and analyzing traffic accident scenes based on 3D reconstruction technology, such as... Figure 1 As shown, it includes the following steps: S1: Acquire static images of the accident site and construct a multi-view image sequence; calculate the atmospheric light vector of the current accident site based on the multi-view image set; calculate the initial transmittance based on the atmospheric light vector; and calculate the meteorological degradation intensity index of the current accident site based on the initial transmittance. S2: Preset scale levels, construct a feature pyramid based on the initial transmittance, and obtain the result feature vector of the accident site through the feature pyramid; calculate the spatial attention map of the current accident site based on the result feature vector, and obtain the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index and the result feature vector; S3: Obtain continuous feature vectors based on real-space coordinates and meteorological degradation perception feature vectors; obtain refined transmittance based on continuous feature vectors; calculate atmospheric attenuation coefficient and clean color image based on refined transmittance. S4: Construct a 3D reconstruction scene based on the volume density and clean color image, obtain the accident meteorological scene based on the atmospheric attenuation coefficient, and generate a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene.

[0018] Based on the synchronously triggered cameras and monitoring cameras deployed at the accident site, static image data of the accident site is collected. The static image data consists of the pixel intensity values ​​of the accident site at multiple time frames. Based on the static image data, a multi-view image set of the current accident site is constructed according to different perspectives and different time frames. The multi-view image set consists of the normalized pixel intensity values ​​of the accident site at different perspectives and different time frames. A preset brightness threshold is set, and a dark color map is obtained from the multi-view image set. The dark color map represents the fog density in the current image. The atmospheric light vector of the current accident site is calculated based on the dark color map and the preset brightness threshold.

[0019] It should be explained that pixel intensity value is the light intensity response value of a pixel on a specified color channel. Each pixel position has a pixel intensity value on the red, green, and blue channels. The overall pixel intensity value of the current pixel position is the average of the pixel intensity values ​​of the three channels. Based on static image data, a multi-view image set of the current accident scene is constructed according to different viewpoints and different time frames. The multi-view image set is the normalized image pixel intensity value of the accident scene from different viewpoints and different time frames. The specific normalization method is: divide the pixel intensity value of each channel of each pixel in all images by 255 to obtain the normalized image pixel intensity value. A preset length size threshold is used as a square window of the length size threshold as a local neighborhood window. Each pixel value in the dark primary color image is in the range of zero to one.

[0020] The atmospheric light vector at the current accident site is calculated based on the dark spectrum and a preset brightness threshold. The initial transmittance at the current accident site is calculated based on a multi-view image set and the atmospheric light vector. The initial transmittance represents the light transmission at the accident site under current weather conditions. The meteorological degradation intensity index at the accident site is then calculated based on the initial transmittance. The meteorological degradation intensity index represents the degree of meteorological degradation during the accident occurrence phase. The formulas for calculating the meteorological degradation intensity index and the atmospheric light vector are as follows: ; ; ; ; ; in, The meteorological degradation intensity index. This represents the total number of x-coordinates of the image pixels. This represents the total number of pixels on the y-coordinate of the image. Let be the set of all pixel coordinates. pixel position In time frame The initial transmittance, For camera indexing, used to represent different viewpoints. For adjusting the defogging intensity parameters, In pixel position The central local area, for The pixel position in This is the index for the three color channels. For the first The camera, the first Frame image in pixels passage The pixel intensity value on For the first The camera, the first Frame image in channel Atmospheric light vector, For the first The camera, the first Frame image in pixels The three-channel pixel intensity vector at that location. This is the set of candidate pixels for atmospheric light. This is a dark primary color image. The brightness threshold. The number of elements in the candidate pixel set for atmospheric light. The first pixel in the candidate pixel set for atmospheric light Each element.

[0021] It needs to be explained that the specific method for obtaining the preset brightness threshold is as follows: all pixels in the dark color image are sorted in descending order of value, and the pixels whose pixel values ​​rank in the top 0.1% of all pixels in the dark color image are selected; the initial transmittance values ​​are distributed in the range of zero to one, and the local window is a square window centered on the target pixel with a length threshold; the dimension of the three-channel pixel intensity vector is three-dimensional, corresponding to the types of color channels; the specific method for obtaining the dehazing intensity adjustment parameters is as follows: a large number of foggy images and corresponding clear weather reference images covering different fog concentrations and different scene types were collected to form a test dataset, and the dehazing intensity adjustment parameters were obtained from... The values ​​are taken one by one in increments of 0.05 within the range. Dehazing is performed on each foggy image in the test dataset, and the dehazing results are quantitatively compared with the corresponding clear reference image to obtain the closest dehazing intensity adjustment parameter. The numerical meaning of the meteorological degradation intensity index is as follows: When the index approaches zero, it means that the initial transmittance of most pixels in the entire frame is close to one, indicating that the accident site is in clear and fog-free or almost fog-free meteorological conditions, and the propagation of light is not significantly affected by atmospheric scattering. When the index approaches one, it means that the initial transmittance of most pixels in the entire frame is close to zero, indicating that the accident site is in extremely dense fog meteorological conditions, and the reflected light from objects in the scene is almost completely scattered.

[0022] A maximum scale level is preset. Based on this maximum scale level, the resolution of the current multi-view pixel set and initial transmissivity is halved. A feature pyramid is constructed based on the current multi-view pixel set and initial transmissivity in ascending order of resolution. This feature pyramid represents the multi-view pixel set and initial transmissivity at different scale levels. The feature vector of the current scale level is obtained from the feature pyramid and used as the input vector of the current convolutional layer. The output vector of the current scale level is then obtained. Dynamic gating masking is performed based on the output vector and initial transmissivity of the current scale level to obtain the feature vector of the current scale level. This feature vector is then used as the input vector of the next scale level convolutional layer. Finally, the feature vector of the maximum scale level in the current feature pyramid is obtained. Dynamic gating masking is then performed based on the feature vector of the maximum scale level to obtain the result feature vector of the maximum scale level. This result feature vector is the low-resolution image vector of the original multi-view image set after convolution and dynamic gating at different scale levels. The formula for calculating the feature vector is as follows: ; ; ; ; in, As the initial input vector, for Output vector of scale-level convolutional layers for Scale-level feature vectors for Scale-level dynamic gating units For the first The camera, the first The three-channel pixel intensity vector map of the frame image. For the first The camera, the first The initial transmittance of the frame image. It is a ReLU nonlinear activation function. For the first The convolutional kernel weight tensor of the layer, For spatial downsampling with a step size of 2, For bilinear upsampling, This is a splicing operation at the channel dimension. For the first The convolution bias vector of the layer, For element-wise multiplication, It is a sigmoid activation function. It is a 1*1 pointwise convolution.

[0023] It needs to be explained that the specific method for obtaining the preset maximum scale level is to increment the preset maximum scale level by one based on the number of color channels. The original resolution image is used as the input to the zeroth level of the feature pyramid. The zeroth level retains the complete resolution and all detail information of the original image. Starting from the zeroth level, with each upward layer, the spatial resolution of the image is halved in both height and width dimensions. A feature pyramid is constructed based on the current multi-view pixel set and initial transmissivity in ascending order of resolution. The feature pyramid represents the multi-view pixel set and initial transmissivity at different scale levels. Each level simultaneously contains downsampled multi-view image data and corresponding initial transmissivity data. The convolutional kernel of each convolutional layer uses a size of... The convolution kernel is 3x3 with a stride of 1. The dimension of the three-channel pixel intensity vector map is the total pixel height multiplied by the total pixel width multiplied by the number of color channels. The dimension of the initial transmittance is the total pixel height multiplied by the total pixel width multiplied by 1. The dimension of the initial input vector is the total pixel height multiplied by the total pixel width multiplied by 4. The zeroth layer has four output channels. The dimension of the output vector of each convolutional layer is the total pixel height multiplied by the total pixel width multiplied by the number of output channels of each convolutional layer. The total pixel height multiplied by the same level is calculated by multiplying the total pixel height according to the current convolutional layer number. The number of output channels of each layer is the same level increase of the number of output channels of the zeroth layer. The number of output channels for a layer is twice the number of output channels for the zeroth layer. The resulting feature vector has a dimension that is the total pixel height minus four times the total pixel width minus four times the number of output channels of the fourth layer. The specific method for obtaining the convolutional kernel weight tensor is as follows: initial values ​​are assigned using the Kaiming He initialization method. During training, the gradient of the total loss function with respect to each convolutional kernel weight tensor is calculated, and the weight values ​​are gradually updated along the gradient descent direction to obtain the weight tensor for each convolutional kernel. The dimension of each layer's convolutional kernel weight tensor is the kernel size multiplied by the number of output channels of the previous layer multiplied by the number of output channels of the current layer. The dimension of each layer's convolutional bias vector is the number of output channels of that layer. The specific method for obtaining the convolutional bias vector is as follows: all bias vectors are uniformly set to all zeros during the initialization phase. The gradient value of the total loss function with respect to each bias vector element is calculated. Based on the preset learning rate, the value of the current bias vector element is subtracted from the learning rate multiplied by the corresponding gradient value to obtain the updated bias vector element value. The number of iterations for training the convolutional kernel weight tensor and convolutional bias vector is 10,000.

[0024] Based on the resulting feature vector, a spatial attention map of the current accident site is obtained through pointwise convolution and transmittance weighting. This spatial attention map represents the attention weights at different pixel locations in the image. A secondary result feature vector is obtained by pointwise convolution of the spatial attention map and the resulting feature vector. This secondary result feature vector is the feature vector of the resulting feature vector weighted by the spatial attention map. A meteorological degradation perception feature vector is calculated based on the secondary result feature vector and the meteorological degradation intensity index. This meteorological degradation perception feature vector is the feature vector of the secondary result feature vector after incorporating the meteorological degradation factor. The formula for calculating the meteorological degradation perception feature vector is as follows: ; ; ; in, For the first The camera, the first Spatial attention map of frame image, For element-wise multiplication, It is a sigmoid activation function. It is a 1*1 pointwise convolution. For the resulting feature vector, For the largest scale level, The initial transmittance at the largest scale level. The feature vector of the quadratic result, For the dimension expansion function, Expanded to three dimensions This is a feature vector for sensing meteorological degradation. For scaling modulation coefficients, For translation modulation coefficients, It is a meteorological degradation intensity index.

[0025] It needs to be explained that the dimensions of the spatial attention map and the meteorological degradation sensing feature vector are the same as the dimensions of the result feature vector. The specific method for obtaining the secondary result feature vector is as follows: the spatial attention map is expanded by copying it multiple times along the channel dimension, with the number of copies equal to the total number of channels in the result feature vector. This ensures that the expanded 3D tensor perfectly matches the result feature vector in terms of spatial size and number of channels. The expanded spatial attention map and the result feature vector are then multiplied element-wise along the spatial position and channel dimension to obtain the weighted feature response, which is the secondary result feature vector. The specific method for obtaining the scaling modulation coefficient is as follows: a batch of calibration image data containing different fog concentration levels is collected. Each frame of the calibration image data includes a corresponding meteorological degradation intensity index. The images are sequentially input into the feature pyramid and spatial attention calculation module to obtain the quadratic result feature vector corresponding to each frame. The overall amplitude of the quadratic result feature vector of each frame is quantized. By statistically analyzing the feature amplitude under different meteorological degradation indices, a statistical relationship function between the meteorological degradation indices and the feature amplitude is established. The compensation factor required to keep the feature amplitude stable under different meteorological degradation indices is calculated based on the statistical relationship function, which is the scaling modulation coefficient. The specific method for obtaining the translation modulation coefficient is as follows: the global average value of the quadratic result feature vector of each frame is statistically analyzed, and a functional relationship between the meteorological degradation indices and the feature response baseline offset is established through statistical analysis. The translation amount to compensate for the baseline offset is calculated based on the functional relationship, which is the translation modulation coefficient.

[0026] According to the camera's user manual, obtain the camera's intrinsic and extrinsic parameter matrices. Expand the coordinates of the current pixel position into homogeneous coordinates. Project the real-world coordinates onto the current camera's pixel coordinate system based on the camera's intrinsic and extrinsic parameter matrices and homogeneous coordinates, and scale the coordinates according to the maximum scale level of the feature pyramid to obtain the projected coordinates of the current real-world coordinates in the pixel coordinate system. Perform bilinear interpolation on the projected coordinates in the meteorological degradation sensing feature vector to obtain continuous feature vectors of the projected coordinates. The continuous feature vectors are the vector values ​​of the projected coordinates in the meteorological degradation sensing feature vector. Concatenate the embedding vector and the continuous feature vectors along the channel dimension to obtain the enhanced scene code. The enhanced scene code is a scene characteristic description of the current projection position. The scene characteristic description is the embedding vector and the continuous feature vector of the current position, where the calculation formula for the continuous feature vector is: ; ; in, Let be the projection coordinates of the current real-world coordinates in the pixel coordinate system, and be , For the first The camera's intrinsic parameter matrix, For the first The camera's extrinsic matrix, Position in real-world coordinates homogeneous coordinates These are continuous eigenvectors of the projected coordinates. It is a bilinear interpolation function. This is a feature vector for sensing meteorological degradation.

[0027] It should be explained that the intrinsic parameter matrix represents the camera's internal optical geometry properties, including the equivalent focal length in the horizontal direction, the equivalent focal length in the vertical direction, the horizontal and vertical coordinates of the principal point of the image plane, and the skew coefficient of the pixel array. The extrinsic parameter matrix represents the camera's spatial position and orientation in the world coordinate system, including a rotation matrix and a translation vector. The rotation matrix represents the rotation relationship between the three axes of the world coordinate system and the three axes of the camera coordinate system. The translation vector represents the offset of the origin of the world coordinate system relative to the optical center of the camera. The intrinsic and extrinsic parameter matrices are obtained from the camera's user manual. The specific method of homogeneous coordinate expansion is to add an extra component with a value of one to the end of the three-dimensional vector to form a four-dimensional vector as the homogeneous coordinates of the current coordinates. The dimension of the continuous feature vector is the number of output channels of the fourth layer.

[0028] A fixed projection compression matrix is ​​used to compress continuous feature vectors, and a three-dimensional guiding vector is obtained through matrix multiplication. The fixed projection compression matrix is ​​used to compress the principal components of the continuous feature vectors. The three-dimensional guiding vector is a three-dimensional vector that retains the basic features of the image from the continuous feature vectors. Linear interpolation is performed on the three-dimensional guiding vector to obtain the three-dimensional guiding vector for all spatial points. Bilateral filtering is applied to the initial transmittance based on the three-dimensional guiding vector to obtain a refined transmittance. The refined transmittance is the detailed transmittance of the current accident scene under the initial transmittance. The atmospheric attenuation coefficient and the clean color image in the current multi-view image set are calculated based on the refined transmittance. The atmospheric attenuation coefficient represents the degree of light attenuation due to rain and fog, and the clean color image is the true color of the object surface at the current spatial point under ideal conditions of no rain or fog obstruction and uniform illumination. The formula for calculating the refined transmittance is: ; ; = ; ; in, Position in real-world coordinates The three-dimensional guiding vector, To refine the transmittance, The local region centered on the projected coordinates of the real-world coordinates. The initial transmittance, These are the projection coordinates of real-world coordinates onto the pixel coordinate system. Position in real-world coordinates The three-dimensional guiding vector, The first variance, The second variance, Position in real-world coordinates Atmospheric attenuation coefficient, For the first The camera in the Camera spot coordinates of a time frame, Position in real-world coordinates Clean color image, Atmospheric light vector, The three-channel pixel intensity vector is the projection coordinate. To prevent zero constant, It is a fixed projection compression matrix.

[0029] It needs to be explained that the fixed projection compression matrix is ​​a matrix used to compress the principal components of continuous eigenvectors. The specific method for obtaining the fixed projection compression matrix is ​​as follows: A large amount of image data from the accident scene is collected, and these images are sequentially input into the aforementioned feature extraction process to obtain the continuous eigenvectors corresponding to each image, constructing a sample set of continuous eigenvectors. Then, principal component analysis is performed on the sample set: First, the mean vector of all sample continuous eigenvectors is calculated. The mean vector is subtracted from each continuous eigenvector in the sample set to obtain a zero-mean sample matrix. The covariance matrix of the sample matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix. The eigenvalues ​​are arranged in descending order, and the eigenvectors corresponding to the three largest eigenvalues ​​are selected. These three eigenvectors are arranged in rows to form a three-row matrix with the number of columns equal to the dimension of the continuous eigenvectors. This matrix is ​​the fixed projection compression matrix. The specific method for obtaining the first variance is: The initial candidate values ​​of the first variance are... The numerical range is set to be related to the size of the local neighborhood window. Grid search optimization is performed on the calibration dataset: with a step size of 0.5 pixels, different first variance values ​​are taken sequentially from three pixels to five pixels. A complete bilateral filtering operation is performed on each set of values. Then, the transmittance consistency loss in the total loss function and the accuracy index of subsequent 3D reconstruction are used as evaluation criteria to select the first variance value that optimizes the comprehensive evaluation index. The specific method for obtaining the second variance is as follows: calculate the Euclidean distance distribution between 3D guide vectors at different pixel positions within the same object surface region, and calculate the Euclidean distance distribution between 3D guide vectors of different object categories. The theoretical range of the second variance is the interval between the mean intra-class distance and the mean inter-class distance. The search is performed within this interval with a step size of 0.1. A complete bilateral filtering operation is performed on each set of candidate second variance values. The transmittance edge preservation capability is used as the evaluation criterion to select the second variance value that optimizes the transmittance edge preservation capability.

[0030] The observed color image of the accident scene is calculated based on the clean color image and the refined transmittance. This observed color image represents the fused scene color after the accident scene reconstruction. The observation weight of each real-space coordinate in the observed color image is calculated based on the refined transmittance. This observation weight represents the probability that light rays will stop at the current real-space coordinate and reflect back to the camera. The final rendered color image is obtained by weighted summation of the observation weights and the observed color image, and this rendered color image serves as the 3D reconstructed scene. The average visibility of the scene is calculated based on the atmospheric attenuation coefficient, and an accident meteorological scene is constructed based on this average visibility. Losses are calculated based on both the 3D reconstructed scene and the accident meteorological scene. A loss assessment is performed on the current 3D reconstructed scene, and traffic accident reconstruction reports are generated based on both the 3D reconstructed scene and the accident meteorological scene. The formulas for calculating the rendered color image, average scene visibility, and equivalent rainfall intensity are as follows: ; ; ; ; in, Position in real-world coordinates The observed color image, Position in real-world coordinates Clean color image, To refine the transmittance, Atmospheric light vector, For observation weights, The sampling point number, For the first The real-world coordinates of each sampling point To render a color image, The total number of sampling points. The average visibility of the scene. This represents the average atmospheric attenuation coefficient in real-world coordinates.

[0031] Based on the pixel intensity values ​​of the 3D reconstructed scene and the original image, the squared Euclidean distance between the observed color image and the true color image is calculated for each pixel location to obtain the color mean square error. The average of the color mean square errors for all pixels, all viewpoints, and all time frames is taken as the first loss function. Based on the refined transmittance and the initial transmittance, the squared Euclidean distance between the refined transmittance and the initial transmittance is calculated for each pixel location to obtain the transmission mean square error. The average of the transmission mean square errors for all pixels, all viewpoints, and all time frames is taken as the second loss function. Based on meteorological... The theoretical atmospheric attenuation index is obtained by multiplying the intensity degradation index and the preset maximum atmospheric attenuation coefficient. The mean square error of the atmosphere at each pixel is obtained by calculating the square of the Euclidean distance between the average atmospheric attenuation index and the theoretical atmospheric attenuation index. The average of the mean square error of transmission is taken for all cameras and all time frames as the third loss function. The first loss function, the second loss function and the third loss function are added to obtain the total loss function. The total loss value of the current 3D reconstructed scene is obtained according to the total loss function. A traffic accident scene reconstruction report is generated based on the total loss value, the 3D reconstructed scene and the accident meteorological scene.

[0032] It needs to be explained that the specific method for obtaining the preset maximum atmospheric attenuation coefficient is based on the standard definition in meteorology and atmospheric optics. When the atmospheric attenuation coefficient reaches 3.0 per kilometer, the corresponding atmospheric visibility is approximately 100 meters, meaning that the human eye can only clearly see the outline of objects within a distance of 100 meters in dense fog. The specific method for constructing the first loss function is as follows: calculate the square of the Euclidean distance between the three-channel color vector of the rendered color image at the current pixel position and the three-channel color vector of the real color image at the current pixel position, obtain the color mean square error at the current pixel position, sum the color mean square errors of all pixels, and then divide by the image size. The total number of pixels is used to obtain the average color mean square error of the current viewpoint in the current time frame. The same calculation is performed on all camera views and all time frames to obtain the average color mean square error of each viewpoint and each time frame. The average color mean square errors of all views and all time frames are summed again and divided by the product of the total number of cameras and the total number of time frames to obtain the global average color mean square error. The global average color mean square error is the first loss function. The full loss function is used as the optimization objective. The gradient of the full loss function is backpropagated layer by layer to the trainable parameters of each layer of the feature extraction network through the backpropagation algorithm, driving the parameters to be updated in the direction of reducing the full loss function.

[0033] Example 2: Based on Example 1, a traffic accident scene reconstruction and analysis system based on three-dimensional reconstruction technology, such as... Figure 2 As shown, it also includes: The meteorological transmission module acquires static images of the accident site and constructs a multi-view image sequence. It calculates the atmospheric light vector of the current accident site based on the multi-view image set, calculates the initial transmittance based on the atmospheric light vector, and calculates the meteorological degradation intensity index of the current accident site based on the initial transmittance. The meteorological degradation module has a preset scale level. It constructs a feature pyramid based on the initial transmittance and obtains the result feature vector of the accident site through the feature pyramid. It calculates the spatial attention map of the current accident site based on the result feature vector and obtains the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index and the result feature vector. The image cleaning module obtains continuous feature vectors based on real-world spatial coordinates and meteorological degradation perception feature vectors; it then obtains refined transmittance based on the continuous feature vectors, and calculates the atmospheric attenuation coefficient and clean color image of the image based on the refined transmittance. The accident reconstruction module constructs a 3D reconstruction scene based on volume density and clean color images, obtains the accident meteorological scene based on the atmospheric attenuation coefficient, and generates a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene.

[0034] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology, characterized by: Includes the following steps: S1: Acquire static images of the accident site and construct a multi-view image sequence; calculate the atmospheric light vector of the current accident site based on the multi-view image set; calculate the initial transmittance based on the atmospheric light vector; and calculate the meteorological degradation intensity index of the current accident site based on the initial transmittance. S2: Preset scale levels, construct a feature pyramid based on the initial transmittance, and obtain the result feature vector of the accident site through the feature pyramid; calculate the spatial attention map of the current accident site based on the result feature vector, and obtain the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index and the result feature vector; S3: Obtain continuous feature vectors based on real-space coordinates and meteorological degradation perception feature vectors; obtain refined transmittance based on continuous feature vectors; calculate atmospheric attenuation coefficient and clean color image based on refined transmittance. S4: Construct a 3D reconstruction scene based on the volume density and clean color image, obtain the accident meteorological scene based on the atmospheric attenuation coefficient, and generate a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene.

2. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 1, characterized in that, The process of acquiring static image data of the accident scene and constructing a multi-view image sequence, and calculating the atmospheric light vector of the current accident scene based on the multi-view image set, includes: acquiring static image data of the accident scene based on synchronously triggered cameras and monitoring cameras deployed at the accident scene, wherein the static image data is the image pixel intensity value of the accident scene at multiple time frames; constructing a multi-view image set of the current accident scene based on the static image data according to different viewpoints and different time frames, wherein the multi-view image set is the normalized image pixel intensity value of the accident scene at different viewpoints and different time frames; setting a preset brightness threshold; acquiring a dark chromaticity map based on the multi-view image set, wherein the dark chromaticity map is the fog density in the current image; and calculating the atmospheric light vector of the current accident scene based on the dark chromaticity map and the preset brightness threshold.

3. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 1, characterized in that, The process of calculating the initial transmittance based on the atmospheric light vector and then calculating the meteorological degradation intensity index of the current accident site based on the initial transmittance includes: calculating the atmospheric light vector of the current accident site based on the dark spectrum and a preset brightness threshold; calculating the initial transmittance of the current accident site based on a multi-view image set and the atmospheric light vector, wherein the initial transmittance is the light transmission situation of the accident site under the current weather conditions; and calculating the meteorological degradation intensity index of the accident site based on the initial transmittance, wherein the meteorological degradation intensity index represents the degree of meteorological degradation during the accident occurrence stage. The calculation formulas for the meteorological degradation intensity index and the atmospheric light vector are as follows: ； ； ； ； ； in, The meteorological degradation intensity index. This represents the total number of x-coordinates of the image pixels. This represents the total number of pixels on the y-coordinate of the image. Let be the set of all pixel coordinates. pixel position In time frame The initial transmittance, For camera indexing, used to represent different viewpoints. For adjusting the defogging intensity parameters, In pixel position The central local area, for The pixel position in This is the index for the three color channels. For the first The camera, the first Frame image in pixels passage The pixel intensity value on For the first The camera, the first Frame image in channel Atmospheric light vector, For the first The camera, the first Frame image in pixels The three-channel pixel intensity vector at that location. This is the set of candidate pixels for atmospheric light. This is a dark primary color image. The brightness threshold. The number of elements in the candidate pixel set for atmospheric light. The first pixel in the candidate pixel set for atmospheric light Each element.

4. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 1, characterized in that, The preset scale level constructs a feature pyramid based on the initial transmittance, and obtains the result feature vector of the accident scene through the feature pyramid. This includes: a preset maximum scale level; halving the resolution of the current multi-view pixel set and initial transmittance based on the maximum scale level; constructing a feature pyramid of the current multi-view pixel set and initial transmittance in ascending order of resolution; the feature pyramid represents the multi-view pixel set and initial transmittance at different scale levels; obtaining the feature vector of the current scale level as the input vector of the current convolutional layer based on the feature pyramid; obtaining the output vector of the current scale level; performing dynamic gating mask calculation based on the output vector of the current scale level and the initial transmittance to obtain the feature vector of the current scale level; using the feature vector of the current scale level as the input vector of the convolutional layer at the next scale level; finally obtaining the feature vector of the maximum scale level in the current feature pyramid; and performing dynamic gating mask calculation based on the feature vector of the maximum scale level to obtain the result feature vector of the maximum scale level. The result feature vector is a low-resolution image vector of the original multi-view image set after convolution and dynamic gating at different scale levels. The formula for calculating the feature vector is: ； ； ； ； in, As the initial input vector, for Output vector of scale-level convolutional layers for Scale-level feature vectors for Scale-level dynamic gating units For the first The camera, the first The three-channel pixel intensity vector map of the frame image. For the first The camera, the first The initial transmittance of the frame image. It is a ReLU nonlinear activation function. For the first The convolutional kernel weight tensor of the layer, For spatial downsampling with a step size of 2, For bilinear upsampling, This is a splicing operation at the channel dimension. For the first The convolution bias vector of the layer, For element-wise multiplication, It is a sigmoid activation function. It is a 1*1 pointwise convolution.

5. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 1, characterized in that, The step of calculating the spatial attention map of the current accident site based on the result feature vector, and obtaining the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index, and the result feature vector, includes: obtaining the spatial attention map of the current accident site through point-by-point convolution and transmittance weighting based on the result feature vector, wherein the spatial attention map represents the attention weights at different pixel positions in the image; obtaining a secondary result feature vector through point-by-point convolution based on the spatial attention map and the result feature vector, wherein the secondary result feature vector is the feature vector of the result feature vector weighted by the spatial attention map; and calculating the meteorological degradation perception feature vector based on the secondary result feature vector and the meteorological degradation intensity index, wherein the meteorological degradation perception feature vector is the feature vector of the secondary result feature vector after incorporating the meteorological degradation factor. The formula for calculating the meteorological degradation perception feature vector is as follows: ； ； ； in, For the first The camera, the first Spatial attention map of frame image, For element-wise multiplication, It is a sigmoid activation function. It is a 1*1 pointwise convolution. For the resulting feature vector, For the largest scale level, The initial transmittance at the largest scale level. The feature vector of the quadratic result, For the dimension expansion function, Expanded to three dimensions This is a feature vector for sensing meteorological degradation. For scaling modulation coefficients, For translation modulation coefficients, It is a meteorological degradation intensity index.

6. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 1, characterized in that, The process of obtaining continuous feature vectors based on real-world coordinates and meteorological degradation perception feature vectors includes: obtaining the camera's intrinsic and extrinsic parameter matrices according to the camera's user manual; expanding the coordinates of the current pixel position to homogeneous coordinates; projecting the real-world coordinates onto the current camera's pixel coordinate system based on the camera's intrinsic, extrinsic, and homogeneous coordinates, and scaling the coordinates according to the maximum scale level of the feature pyramid to obtain the projected coordinates of the current real-world coordinates in the pixel coordinate system; performing bilinear interpolation on the projected coordinates in the meteorological degradation perception feature vector to obtain continuous feature vectors of the projected coordinates, where the continuous feature vectors are the vector values ​​of the projected coordinates in the meteorological degradation perception feature vector; and concatenating the embedding vector and the continuous feature vectors along the channel dimension to obtain enhanced scene encoding, where the enhanced scene encoding is a scene characteristic description of the current projection position, and the scene characteristic description is the embedding vector and the continuous feature vector of the current position, wherein the calculation formula for the continuous feature vector is: ； ； in, Let be the projection coordinates of the current real-world coordinates in the pixel coordinate system, and be , For the first The camera's intrinsic parameter matrix, For the first The camera's extrinsic matrix, Position in real-world coordinates homogeneous coordinates These are continuous eigenvectors of the projected coordinates. It is a bilinear interpolation function. This is a feature vector for sensing meteorological degradation.

7. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 6, characterized in that, The process of obtaining refined transmittance based on continuous feature vectors and calculating the atmospheric attenuation coefficient and clean color image of the image based on the refined transmittance includes: obtaining a three-dimensional guiding vector by multiplying a fixed projection compression matrix based on the continuous feature vectors, wherein the fixed projection compression matrix is ​​a matrix used to compress the principal components of the continuous feature vectors, and the three-dimensional guiding vector is a three-dimensional vector that retains the basic features of the image from the continuous feature vectors; obtaining a three-dimensional guiding vector for all spatial points by linear interpolation of the three-dimensional guiding vector; obtaining a refined transmittance by bilateral filtering of the initial transmittance based on the three-dimensional guiding vector, wherein the refined transmittance is the detailed transmittance of the current accident scene under the initial transmittance; and calculating the atmospheric attenuation coefficient and clean color image in the current multi-view image set based on the refined transmittance; wherein the atmospheric attenuation coefficient is the degree of light attenuation due to rain and fog, and the clean color image is the true color of the object surface at the current spatial point under ideal conditions of no rain or fog obstruction and uniform illumination, wherein the formula for calculating the refined transmittance is: ； ； = ； ； in, Position in real-world coordinates The three-dimensional guiding vector, To refine the transmittance, The local region centered on the projected coordinates of the real-world coordinates. The initial transmittance, These are the projection coordinates of real-world coordinates onto the pixel coordinate system. Position in real-world coordinates The three-dimensional guiding vector, The first variance, The second variance, Position in real-world coordinates Atmospheric attenuation coefficient, For the first The camera in the Camera spot coordinates of a time frame, Position in real-world coordinates Clean color image, Atmospheric light vector, The three-channel pixel intensity vector is the projection coordinate. To prevent zero constant, It is a fixed projection compression matrix.

8. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 1, characterized in that, The process of constructing a 3D reconstruction scene based on volume density and clean color images, obtaining an accident meteorological scene based on atmospheric attenuation coefficients, and generating a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene includes: calculating the observed color image of the accident scene based on the clean color image and refined transmittance, wherein the observed color image is the fused scene color after the accident scene reconstruction; calculating the observation weight of each real-space coordinate of the observed color image based on refined transmittance, wherein the observation weight is the probability that light will stop at the current real-space coordinate and reflect back to the camera; obtaining the final rendered color image by performing a point-by-point weighted summation based on the observation weights and the observed color image, and using the rendered color image as the 3D reconstruction scene; calculating the average visibility of the scene based on the atmospheric attenuation coefficient, and constructing an accident meteorological scene of the current accident scene based on the average visibility of the scene; calculating the loss based on the 3D reconstruction scene and the accident meteorological scene, determining the loss of the current 3D reconstruction scene, and generating a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene respectively; wherein the calculation formulas for the rendered color image, the average visibility of the scene, and the equivalent rainfall intensity are: ； ； ； ； in, Position in real-world coordinates The observed color image, Position in real-world coordinates Clean color image, To refine the transmittance, Atmospheric light vector, For observation weights, The sampling point number, For the first The real-world coordinates of each sampling point To render a color image, The total number of sampling points. The average visibility of the scene. This represents the average atmospheric attenuation coefficient in real-world coordinates.

9. The method for reconstructing and analyzing traffic accident scenes based on three-dimensional reconstruction technology as described in claim 8, characterized in that, The step of calculating losses based on the 3D reconstructed scene and the accident weather scene, and determining the loss for the current 3D reconstructed scene, includes: calculating the square of the Euclidean distance between the observed color image and the real color image for each pixel location based on the pixel intensity values ​​of the 3D reconstructed scene and the original image, obtaining the color mean square error for each pixel location, and taking the average of the color mean square errors for all pixels, all viewpoints, and all time frames as the first loss function; calculating the square of the Euclidean distance between the refined transmittance and the initial transmittance for each pixel location, obtaining the transmission mean square error for each pixel location, and taking the average of the color mean square errors for all pixels, all viewpoints, and all time frames as the first loss function; and taking the average of the color mean square errors for all pixels, all viewpoints, and all time frames as the first loss function. The average of the transmission mean square error is used as the second loss function; the theoretical atmospheric attenuation index is calculated by multiplying the meteorological intensity degradation index and the preset maximum atmospheric attenuation coefficient; the atmospheric mean square error of each pixel is obtained by calculating the square of the Euclidean distance between the average atmospheric attenuation index and the theoretical atmospheric attenuation index; the average of the transmission mean square error of all cameras and all time frames is used as the third loss function; the first, second, and third loss functions are added together to obtain the total loss function; the total loss value of the current 3D reconstructed scene is obtained based on the total loss function; and a traffic accident scene reconstruction report is generated based on the total loss value, the 3D reconstructed scene, and the accident meteorological scene.

10. A traffic accident scene reconstruction and analysis system based on three-dimensional reconstruction technology, used to implement the traffic accident scene reconstruction and analysis method based on three-dimensional reconstruction technology as described in any one of claims 1-9, characterized in that it further... include: The meteorological transmission module acquires static images of the accident site and constructs a multi-view image sequence. Based on the multi-view image set, it calculates the atmospheric light vector of the current accident site. The initial transmittance is calculated based on the atmospheric light vector, and the meteorological degradation intensity index at the current accident site is calculated based on the initial transmittance. The meteorological degradation module has a preset scale level. It constructs a feature pyramid based on the initial transmittance and obtains the result feature vector of the accident site through the feature pyramid. It calculates the spatial attention map of the current accident site based on the result feature vector and obtains the meteorological degradation perception feature vector based on the spatial attention map, the meteorological degradation intensity index and the result feature vector. The image cleaning module obtains continuous feature vectors based on real-world spatial coordinates and meteorological degradation perception feature vectors; it then obtains refined transmittance based on the continuous feature vectors, and calculates the atmospheric attenuation coefficient and clean color image of the image based on the refined transmittance. The accident reconstruction module constructs a 3D reconstruction scene based on volume density and clean color images, obtains the accident meteorological scene based on the atmospheric attenuation coefficient, and generates a traffic accident reconstruction report based on the 3D reconstruction scene and the accident meteorological scene.

Images