A driving early warning method and system based on water film rolling fog and image visibility

Abstract

The application relates to a driving early warning method based on water film rolling fog and image visibility, which captures and analyzes a driving vehicle on a front adjacent lane in real time through a rear video capture, comprises the following steps: applying a visibility detection model to detect a forward visibility distance corresponding to a rear video capture position, simultaneously detecting a water film thickness of a road section where a front driving vehicle is located through a road water film thickness detection model, calculating a rolling fog equivalent influence distance of the front driving vehicle, and finally comprehensively mapping the visibility distance and the rolling fog equivalent influence distance under a preset classification design to determine a driving early warning of the rear video capture position, and further designing a corresponding system to realize the landing application of the design, modularly building and implementing the design method, determining accurate and specific early warning information, realizing the driving early warning in the vehicle driving process, and greatly improving the safety of vehicle driving in rainy days.

Description

Technical Field

[0001] This invention relates to a driving warning method and system based on water film fogging and image visibility, belonging to the field of rainy weather safe traffic driving analysis technology. Background Technology

[0002] In rainy weather, when vehicles drive on flooded roads, the high-speed rotation of the wheels will cause the water on the road surface to be rolled up, forming water mist and water curtain, which significantly reduces the visibility of following vehicles and can lead to rear-end collisions in severe cases.

[0003] In existing technologies, visibility warnings for driving in rainy weather mainly rely on macroscopic environmental data collected by road weather stations, but this approach has the following drawbacks:

[0004] 1. Insufficient spatial resolution: The visibility data from weather stations reflects the average conditions within a few hundred meters around the station, and cannot capture localized changes in visibility at the vehicle's location;

[0005] 2. Lack of vehicle-fog causal relationship: Existing methods cannot identify the impact of rain and fog kicked up by a specific large vehicle on a specific vehicle behind it, and cannot achieve targeted early warning;

[0006] 3. Insufficient foresight in forecasting: Early warnings based on historical statistics cannot reflect the dynamic coupling effect of real-time traffic flow, road conditions and rainfall, resulting in insufficient advance warning. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a driving warning method based on water film fogging and image visibility. The method constructs a two-dimensional analysis based on the visibility observed in the rear direction of the vehicle in front and the equivalent influence distance of the fogging behind the vehicle in front, so as to accurately realize driving warning.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention designs a driving warning method based on water film fogging and image visibility. Based on video capture of a vehicle traveling in front from the position of the adjacent lane behind, the present invention performs the following steps A to D in real time for each video frame to provide a driving warning in real time.

[0009] Step A. For each video frame, apply a pre-trained visibility detection model to obtain the visibility distance corresponding to the video capture position behind, and then proceed to Step B;

[0010] Step B. Based on the current preset rainfall characteristics and current preset road characteristics corresponding to the road segment where the vehicle ahead is located, apply the pre-trained road surface water film thickness detection model to obtain the current water film thickness of the road segment where the vehicle ahead is located, and then proceed to step C.

[0011] Step C. Based on the current water film thickness of the road section where the vehicle ahead is located, and considering the driving parameter characteristics of the vehicle ahead, calculate the amount of fog caused by the vehicle's movement, and then calculate the equivalent impact distance of the fog behind the vehicle ahead. Then proceed to step D.

[0012] Step D. Based on the visibility distance corresponding to the rear video capture location and the equivalent impact distance of the fog on the vehicle in front, obtain the driving warning for the rear video capture location.

[0013] As a preferred technical solution of the present invention: the visibility detection model includes an image encoding module, a feature enhancement module, and a multimodal fusion and regression head module connected in series from the input end to the output end. The input end of the image encoding module constitutes the input end of the visibility detection model, and the output end of the multimodal fusion and regression head module constitutes the output end of the visibility detection model.

[0014] The image coding module consists of a sequentially connected input stem and four residual stages from input to output. The input of the input stem constitutes the input of the image coding module, and the output of the fourth residual stage constitutes the output of the image coding module. The input stem consists of a sequentially connected convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. The input of the convolutional layer constitutes the input of the input stem, and the output of the max pooling layer constitutes the output of the input stem. Each residual stage consists of 3, 4, 6, and 3 residual blocks respectively, sequentially connected from input to output. The input of the first residual block in each residual stage constitutes the input of the residual stage, and the output of the last residual block constitutes the output of the residual stage.

[0015] Each residual block includes a fusion module, an activation module, and two branches. In each residual block structure, the inputs of the two branches are connected to form the input of the residual block. One branch, from its input to its output, includes two convolutional normalization activation modules and one convolutional normalization module connected in series. The input of the first convolutional normalization activation module forms the input of one branch, and the output of the convolutional normalization module forms the output of one branch. The outputs of the two branches are connected to the input of the fusion module, and the output of the fusion module is connected to the input of the activation module. The output of the activation module forms the output of the residual block. In the first residual block of each residual stage structure, the other branch is a convolutional normalization module, and the number of input channels and the number of output channels of the two branches are equal. In the remaining residual blocks of each residual stage structure, the other branch is a direct identity mapping, used to realize the direct connection between the input of the residual block and the input of the fusion module.

[0016] As a preferred technical solution of the present invention: the feature enhancement module includes a channel attention module, a spatial attention module, a discrete cosine transform module, and a multi-scale feature fusion module, wherein the input ends of the channel attention module, the spatial attention module, and the discrete cosine transform module are connected to form the input end of the feature enhancement module, and the output ends of the channel attention module, the spatial attention module, and the discrete cosine transform module are connected to and connected to the input end of the multi-scale feature fusion module, and the output end of the multi-scale feature fusion module constitutes the output end of the feature enhancement module;

[0017] The multimodal fusion and regression head module consists of a global average pooling layer, a fully connected layer FC1, a ReLU activation layer, a Dropout layer, a fully connected layer FC2, and a ReLU activation layer connected in series from the input to the output. The input of the multimodal fusion and regression head module is formed by the input of the global average pooling layer, and the output of the last ReLU activation layer is formed by the output of the multimodal fusion and regression head module.

[0018] As a preferred embodiment of the present invention: the road surface water film thickness detection model comprises an input layer, a hidden layer, and an output layer connected in series from the input end to the output end. The input end of the input layer constitutes the input end of the road surface water film thickness detection model, and the output end of the output layer constitutes the output end of the road surface water film thickness detection model. The number of neurons in the input layer is equal to the number of preset rainfall features and preset road features received by the road surface water film thickness detection model, and they correspond one-to-one. The output layer includes one neuron for outputting the water film thickness.

[0019] As a preferred technical solution of the present invention: a water film thickness sensor is distributed along the test road direction and located on the roadside ground; two vehicles traveling at the same speed in adjacent lanes in the test road direction are separated by front and rear; the rear vehicle has a built-in satellite positioning module; and an image capture device with the shooting direction pointing towards the front vehicle.

[0020] Based on the exhaustive enumeration of three types of data—preset rainfall levels, preset distances between workshops, and preset driving speeds—the following steps a to b are performed for each spatiotemporal combination to obtain individual samples corresponding to each time point under the spatiotemporal combination application. Then, the individual samples corresponding to each time point under each spatiotemporal combination are combined to form a sample set, and step d is performed to achieve model training.

[0021] Step a. Two vehicles travel on the test road according to the requirements of spatiotemporal combination. The vehicle behind captures video of the vehicle in front through an image capture device. At the same time, the vehicle behind travels through the test road segment corresponding to its positioning information, and determines the positioning information of the vehicle in front in real time by considering the distance between vehicles. It also obtains the preset road features of the test road segment where the vehicle in front is located in real time, and obtains the water film thickness on the test road segment where the vehicle in front is located in real time through the water film thickness sensors on the roadside.

[0022] Step b. For each video frame, based on the annotations of the boundary polygons of the dense fog area on the video frame and the annotations of the distance from the shooting position to the preset type of reference object, and considering the distance between the two vehicles in front and behind, obtain the standard visibility distance corresponding to the video frame through perspective transformation; according to the time point corresponding to each video frame, obtain the video frame, standard visibility distance, various rainfall characteristics under the spatiotemporal combination, water film thickness of the test road section where the vehicle in front is located, and preset road characteristics corresponding to each time point, thus forming a single sample corresponding to each time point; and thus forming a sample set.

[0023] Step d. Based on the sample set, using video frames in the samples as input and corresponding visibility distance as output, train the visibility detection model to obtain a trained visibility detection model; based on the sample set, using various rainfall features and road features in the samples as input and corresponding water film thickness as output, train the road surface water film thickness detection model to obtain a trained road surface water film thickness detection model.

[0024] As a preferred embodiment of the present invention: In step C, the current water film thickness of the road segment where the vehicle ahead is located is first determined. This involves considering the driving parameters of vehicles ahead, including their speed. The tire contact patch width of the vehicle ahead According to the following formula:

[0025] ;

[0026] Calculate the amount of fog on a single tire of a vehicle traveling ahead. ,in, Indicates the preset water film entrainment coefficient;

[0027] Then follow the formula: The total fog amount of vehicles traveling ahead is calculated. ,in, This indicates the preset multi-round occlusion correction coefficient;

[0028] Finally, follow the formula: Calculate the equivalent distance of the fog effect on the vehicle ahead. ,in, This represents the preset diffusion attenuation coefficient. Indicates the current wind speed.

[0029] As a preferred technical solution of the present invention: step D includes the following steps D1 to D2;

[0030] Step D1. Based on the scores corresponding to each preset visibility level and the corresponding visibility distance range, determine the visibility distance corresponding to the rear video capture position. Further, the corresponding visibility level score ;

[0031] Based on the scores corresponding to each preset impact distance level and the corresponding equivalent impact distance range of fog, the equivalent impact distance of fog blowing backward from a vehicle traveling ahead is determined. The score corresponding to the distance level is further affected. ;

[0032] Then proceed to step D2;

[0033] Step D2. Combine scores according to preset weights. With score The weighted values ​​are then calculated to obtain the weighted result. Based on the preset score range corresponding to each warning level, the warning level corresponding to the weighted result is determined according to the following formula:

[0034] ;

[0035] Obtain the safe rearward distance of a vehicle traveling in front. Based on the warning level and safe distance The combination constitutes a driving warning system based on the rear video capture location, wherein, Indicates the braking distance of the vehicle ahead. This indicates the preset safety distance fluctuation amount that is greater than 0.

[0036] Corresponding to the above, the technical problem that this invention also needs to solve is to provide a system for realizing a driving warning method based on water film fogging and image visibility, a modular decomposition implementation design method, and a systematic and accurate realization of driving warning.

[0037] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention designs a driving warning method based on water film fogging and image visibility, which is based on an image capture device installed on the vehicle for video capture of vehicles traveling in adjacent lanes ahead along the road driving direction, and also includes a satellite positioning module, a visibility detection module, an impact distance detection module, and a graded warning module deployed on the vehicle; the vehicle obtains its real-time positioning information according to the satellite positioning module;

[0038] For each video frame in the captured video, the visibility detection module performs step A to obtain the visibility distance in front of the vehicle; the influence distance detection module performs steps B to C to obtain the current water film thickness of the road segment where the vehicle is located; and the graded warning module performs step D based on the visibility distance in front of the vehicle and the current water film thickness of the road segment where the vehicle is located to obtain the driving warning corresponding to the vehicle.

[0039] The driving warning method and system based on water film fogging and image visibility described in this invention have the following technical advantages compared with the prior art:

[0040] This invention presents a driving warning method based on water film fogging and image visibility. The method involves a rear-end system capturing and analyzing video of vehicles in adjacent lanes ahead. This includes using a visibility detection model to determine the forward visibility distance corresponding to the video capture location, and using a road surface water film thickness detection model to determine the water film thickness of the road segment where the vehicle is located, calculating the equivalent fogging effect distance of the vehicle behind. Finally, by combining the visibility distance and the equivalent fogging effect distance and mapping under a pre-defined classification design, a driving warning is determined for the rear video capture location. A corresponding system is further designed to achieve practical application. The modular construction and implementation of the design method capture sudden changes in visibility in local road sections, improving the perception accuracy from "hundred-meter-level regional average" to "vehicle-level real-time local," avoiding ineffective wide-area warnings. Through precise, forward-looking, targeted, and intelligent warnings, the method effectively alleviates the problem of sudden visibility drops caused by water fogging from vehicles, reducing the occurrence of rear-end collisions and other accidents. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the visibility detection model in the design of this invention. Detailed Implementation

[0042] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] The present invention designs a driving warning method based on water film fogging and image visibility. In practical application, for a moving vehicle, the vehicle in the adjacent lane behind it captures video of the vehicle in front of it. That is, the vehicle in the adjacent lane behind it captures video of the vehicle in front of it through its on-board image capture device, and performs the following steps A to D in real time for each video frame to provide a driving warning.

[0044] Step A. For each video frame, apply a pre-trained visibility detection model to obtain the visibility distance corresponding to the vehicles behind, and then proceed to Step B.

[0045] Step B. Based on the current preset rainfall characteristics and current preset road characteristics corresponding to the road segment where the vehicle ahead is located, apply the pre-trained road surface water film thickness detection model to obtain the current water film thickness of the road segment where the vehicle ahead is located, and then proceed to step C.

[0046] Step C. Based on the current water film thickness of the road section where the vehicle ahead is located, and considering the driving parameter characteristics of the vehicle ahead, calculate the amount of fog caused by the vehicle's movement, and then calculate the equivalent impact distance of the fog behind the vehicle ahead. Then proceed to step D.

[0047] Step D. Based on the visibility distance of the vehicles behind and the equivalent impact distance of the fog on the vehicles in front, obtain the driving warning of the rear video capture position.

[0048] In actual implementation, the above design scheme, such as Figure 1 As shown, the structure of the visibility detection model is specifically designed. From the input end to the output end, it includes an image encoding module, a feature enhancement module, and a multimodal fusion and regression head module connected in series. The input end of the image encoding module constitutes the input end of the visibility detection model, and the output end of the multimodal fusion and regression head module constitutes the output end of the visibility detection model.

[0049] like Figure 1As shown, the image coding module includes a sequentially connected input stem and four residual stages from the input end to the output end. The input end of the input stem constitutes the input end of the image coding module, and the output end of the fourth residual stage constitutes the output end of the image coding module. The input stem includes a sequentially connected convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. The input end of the convolutional layer constitutes the input end of the input stem, and the output end of the max pooling layer constitutes the output end of the input stem. Each residual stage includes, respectively, three, four, six, and three residual blocks sequentially connected from the input end to the output end. The input end of the first residual block in each residual stage constitutes the input end of the residual stage, and the output end of the last residual block constitutes the output end of the residual stage.

[0050] Each residual block includes a fusion module, an activation module, and two branches. In each residual block structure, the inputs of the two branches are connected to form the input of the residual block. One branch, from its input to its output, includes two convolutional normalization activation modules and one convolutional normalization module connected in series. The input of the first convolutional normalization activation module forms the input of one branch, and the output of the convolutional normalization module forms the output of one branch. The outputs of the two branches are connected to the input of the fusion module, and the output of the fusion module is connected to the input of the activation module. The output of the activation module forms the output of the residual block. In the first residual block of each residual stage structure, the other branch is a convolutional normalization module, and the number of input channels and the number of output channels of the two branches are equal. In the remaining residual blocks of each residual stage structure, the other branch is a direct identity mapping, used to realize the direct connection between the input of the residual block and the input of the fusion module.

[0051] Regarding the feature enhancement module, such as Figure 1 As shown, the specific design includes a channel attention module, a spatial attention module, a discrete cosine transform module, and a multi-scale feature fusion module. The inputs of the channel attention module, spatial attention module, and discrete cosine transform module are connected to form the input of the feature enhancement module. The outputs of the channel attention module, spatial attention module, and discrete cosine transform module are connected and connected to the input of the multi-scale feature fusion module. The output of the multi-scale feature fusion module constitutes the output of the feature enhancement module. Here, the three attention mechanisms are fused in parallel to achieve dual feature enhancement for the "blurred edges + fog texture" of the water mist image.

[0052] Regarding the multimodal fusion and regression head module, such as Figure 1As shown, the specific design from the input end to the output end includes a global average pooling layer, a fully connected layer FC1, a ReLU activation layer, a Dropout layer, a fully connected layer FC2, and a ReLU activation layer connected in series. The input end of the global average pooling layer constitutes the input end of the multimodal fusion and regression head module, and the output end of the last ReLU activation layer constitutes the output end of the multimodal fusion and regression head module.

[0053] The above design scheme also involves a road surface water film thickness detection model. In practical applications, the specific design structure, from input to output, includes an input layer, a hidden layer, and an output layer connected in series. The input end of the input layer constitutes the input end of the road surface water film thickness detection model, and the output end of the output layer constitutes the output end of the road surface water film thickness detection model. The number of neurons in the input layer is equal to the number of preset rainfall features and preset road features received by the road surface water film thickness detection model, and they correspond one-to-one. The hidden layer includes... The layer consists of one neuron, with the kernel function using a Gaussian radial basis function; the output layer includes one neuron for outputting the water film thickness.

[0054] The visibility detection model and the road surface water film thickness detection model involved in the design scheme are both applied in practice using the corresponding pre-trained models. Regarding the training of the above two models in practical applications, the specific training process is designed as follows: First, a data acquisition environment is built based on the test road. Various water film thickness sensors are distributed along the test road direction at the roadside ground position, and two vehicles are arranged in front and behind each other in the test road direction, with two adjacent lanes traveling at the same speed. The rear vehicle has a built-in satellite positioning module and an image capture device pointing in the direction of the front vehicle.

[0055] Before collecting training data, based on preset rainfall levels, preset distances between vehicles, and preset driving speeds, various spatiotemporal combinations of these three types of data were constructed through exhaustive enumeration. Among them, preset rainfall levels include light rain, moderate rain, heavy rain, and torrential rain; preset distances between vehicles include 10m, 20m, 30m, 50m, and 80m; and preset driving speeds include 20m / s, 25m / s, 30m / s, 35m / s, and 40m / s.

[0056] For each spatiotemporal combination, perform steps a to b to obtain individual samples corresponding to each time point under the spatiotemporal combination application. Then, combine the individual samples corresponding to each time point under each spatiotemporal combination to form a sample set.

[0057] Step a. Two vehicles travel on the test road according to the requirements of spatiotemporal combination. The vehicle behind captures video of the vehicle in front through an image capture device. At the same time, the vehicle behind travels through the test road segment corresponding to its positioning information, and determines the positioning information of the vehicle in front in real time by considering the distance between vehicles. It also obtains the preset road features of the test road segment where the vehicle in front is located in real time, and obtains the water film thickness on the test road segment where the vehicle in front is located in real time through the water film thickness sensors on the roadside.

[0058] Step b. For each video frame, based on the annotations of the boundary polygons of the dense fog area on the video frame and the annotations of the distance from the shooting position to the preset type of reference object, and considering the distance between the two vehicles in front and behind, obtain the standard visibility distance corresponding to the video frame through perspective transformation; according to the time point corresponding to each video frame, obtain the video frame, standard visibility distance, various rainfall characteristics under the spatiotemporal combination, water film thickness of the test road section where the vehicle in front is located, and preset road characteristics corresponding to each time point according to the synchronization method of the time point, and form a single sample corresponding to each time point; thus forming a sample set.

[0059] In practical applications, it is possible to further design the process of executing steps a to b at least five times for each spatiotemporal combination to eliminate random errors and obtain individual samples corresponding to each time point of each application under the spatiotemporal combination.

[0060] In practical applications, the construction of the sample set is required to have a scale of more than 5,000 samples, of which the proportion of video frames containing water mist in each sample is ≥30%. In actual implementation, based on the obtained sample set, the following step d is further performed to achieve model training.

[0061] Step d. Based on the sample set, with video frames in the sample as input and corresponding visibility distance as output, train the visibility detection model to obtain a trained visibility detection model. During this training process, the Smooth L1 Loss loss function is selected for training.

[0062] Furthermore, based on the sample set, the model for detecting road surface water film thickness is trained by taking the rainfall and road features in the sample as inputs and the corresponding water film thickness as outputs. In practical applications, the rainfall features, such as minute precipitation, atmospheric pressure, and road surface temperature, are obtained from meteorological monitoring websites; the road features, such as the longitudinal slope, slope length, and surface coarseness of the test road section where the vehicle ahead is located, are obtained by querying the road design parameter table.

[0063] The design scheme of this invention is executed according to the above steps A to B, that is, the pre-trained visibility detection model and the pre-trained road surface water film thickness detection model are applied respectively to obtain the visibility distance corresponding to the vehicle behind and the current water film thickness of the road segment where the vehicle in front is located. Then, steps C and D are executed in sequence according to the design to achieve the final driving warning about the vehicle behind.

[0064] In the practical application of step C, based on the current water film thickness of the road segment where the vehicle ahead is located, the first step is to determine the current water film thickness of the road segment where the vehicle ahead is located. This involves considering the driving parameters of vehicles ahead, including their speed. The tire contact patch width of the vehicle ahead According to the following formula:

[0065] ;

[0066] Calculate the amount of fog on a single tire of a vehicle traveling ahead. ,in, This indicates the preset water film entrainment coefficient.

[0067] Then follow the formula: The total fog amount of vehicles traveling ahead is calculated. ,in, This indicates a preset multi-round occlusion correction coefficient, which is used in practical applications such as setting... .

[0068] Finally, follow the formula: Calculate the equivalent distance of the fog effect on the vehicle ahead. ,in, This represents the preset diffusion attenuation coefficient. Indicates the current wind speed.

[0069] Finally, the actual implementation of step D involves executing steps D1 to D2 based on the visibility distance corresponding to the rear video capture location and the equivalent impact distance of the fog on the vehicle in front, to obtain a driving warning for the rear video capture location.

[0070] Step D1. Based on the scores corresponding to each preset visibility level and the corresponding visibility distance range, determine the visibility distance corresponding to the rear video capture position. Further, the corresponding visibility level score Based on the scores corresponding to each preset impact distance level and the corresponding equivalent impact distance range of fog, the equivalent impact distance of fog blowing backward from a vehicle traveling ahead is determined. The score corresponding to the distance level is further affected. Then proceed to step D2.

[0071] Step D2. Combine scores according to preset weights. With score The weighted values ​​are then calculated to obtain the weighted result. Based on the preset score range corresponding to each warning level, the warning level corresponding to the weighted result is determined according to the following formula:

[0072] ;

[0073] Obtain the safe rearward distance of a vehicle traveling in front. Based on the warning level and safe distance The combination constitutes a driving warning system based on the rear video capture location, wherein, Indicates the braking distance of the vehicle ahead. This indicates the preset safety distance fluctuation amount that is greater than 0.

[0074] To realize the practical application of the designed driving warning method, a corresponding system was further designed, including an image capture device, a satellite positioning module, a visibility detection module, an impact distance detection module, and a graded warning module deployed on the vehicle. Each vehicle operates according to the various modules of its onboard system, capturing and analyzing images of vehicles ahead during its journey to obtain driving warnings. This process is implemented by the various modules of the designed system. The image capture device is installed inside the vehicle's windshield, facing forward, and is powered by the vehicle's power supply. It is used to capture video of vehicles traveling in adjacent lanes ahead along the road's direction of travel. The satellite positioning module, such as an RTK-GPS module, is used to obtain the vehicle's positioning information.

[0075] For the acquired video, the visibility detection module performs step A to obtain the visibility distance ahead of the vehicle for each video frame in the captured video; the influence distance detection module performs steps B to C to obtain the current water film thickness of the road segment where the vehicle is located; and the graded warning module performs step D to obtain the driving warning corresponding to the vehicle based on the visibility distance ahead of the vehicle and the current water film thickness of the road segment where the vehicle is located.

[0076] This invention's design involves a rear-mounted system that captures and analyzes video of vehicles traveling in adjacent lanes ahead. This includes using a visibility detection model to determine the forward visibility distance corresponding to the video capture location, and using a road surface water film thickness detection model to determine the water film thickness of the road segment where the vehicle is located, calculating the equivalent fog impact distance of the vehicle behind. Finally, by combining the visibility distance and the equivalent fog impact distance and mapping under a pre-defined classification design, a driving warning for the rear video capture location is determined. A corresponding system is further designed to achieve practical application. The modular construction and implementation of the design method capture sudden changes in visibility in local road sections, improving perception accuracy from "hundred-meter-level regional average" to "vehicle-level real-time local," avoiding ineffective wide-area warnings. Through precise, forward-looking, targeted, and intelligent warnings, it effectively alleviates the problem of sudden visibility drops caused by water vapor kicked up by vehicles, reducing the occurrence of rear-end collisions and other accidents.

[0077] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A driving warning method based on water film fogging and image visibility, characterized in that: Based on video capture of vehicles traveling ahead from the position of the adjacent lane behind, the following steps A to D are executed in real time for each video frame to provide real-time driving warnings. Step A. For each video frame, apply a pre-trained visibility detection model to obtain the visibility distance corresponding to the video capture position behind, and then proceed to Step B; Step B. Based on the current preset rainfall characteristics and current preset road characteristics corresponding to the road segment where the vehicle ahead is located, apply the pre-trained road surface water film thickness detection model to obtain the current water film thickness of the road segment where the vehicle ahead is located, and then proceed to step C. The road surface water film thickness detection model consists of an input layer, a hidden layer, and an output layer connected in series from the input end to the output end. The input end of the input layer constitutes the input end of the road surface water film thickness detection model, and the output end of the output layer constitutes the output end of the road surface water film thickness detection model. The number of neurons in the input layer is equal to the number of preset rainfall features and preset road features received by the road surface water film thickness detection model, and they correspond one-to-one. The output layer includes one neuron, which is used to output the water film thickness. Step C. Based on the current water film thickness of the road section where the vehicle ahead is located, and considering the driving parameter characteristics of the vehicle ahead, calculate the amount of fog caused by the vehicle's movement, and then calculate the equivalent impact distance of the fog behind the vehicle ahead. Then proceed to step D. In step C above, the first step is to determine the current water film thickness of the road section where the vehicle ahead is located. This involves considering the driving parameters of vehicles ahead, including their speed. The tire contact patch width of the vehicle ahead According to the following formula: ； Calculate the amount of fog on a single tire of a vehicle traveling ahead. ,in, Indicates the preset water film entrainment coefficient; Then follow the formula: The total fog amount of vehicles traveling ahead is calculated. ,in, This indicates the preset multi-round occlusion correction coefficient; Finally, follow the formula: Calculate the equivalent distance of the fog effect on the vehicle ahead. ,in, This represents the preset diffusion attenuation coefficient. Indicates the current wind speed; Step D. Based on the visibility distance corresponding to the rear video capture location and the equivalent impact distance of the fog on the vehicle in front, obtain the driving warning for the rear video capture location; Step D above includes steps D1 to D2 as follows; Step D1. Based on the scores corresponding to each preset visibility level and the corresponding visibility distance range, determine the visibility distance corresponding to the rear video capture position. Further, the corresponding visibility level score ; Based on the scores corresponding to each preset impact distance level and the corresponding equivalent impact distance range of fog, the equivalent impact distance of fog blowing backward from a vehicle traveling ahead is determined. Further, the corresponding score affects the distance level. ; Then proceed to step D2; Step D2. Combine scores according to preset weights. With score The weighted values ​​are then calculated to obtain the weighted result. Based on the preset score range corresponding to each warning level, the warning level corresponding to the weighted result is determined according to the following formula: ； Obtain the safe rearward distance of a vehicle traveling in front. Based on the warning level and safe distance The combination constitutes a driving warning system based on the rear video capture location, wherein, Indicates the braking distance of the vehicle ahead. This indicates the preset safety distance fluctuation amount that is greater than 0.

2. The driving warning method based on water film fogging and image visibility according to claim 1, characterized in that: The visibility detection model includes an image encoding module, a feature enhancement module, and a multimodal fusion and regression head module connected in series from the input end to the output end. The input end of the image encoding module constitutes the input end of the visibility detection model, and the output end of the multimodal fusion and regression head module constitutes the output end of the visibility detection model. The image coding module consists of a sequentially connected input stem and four residual stages from input to output. The input of the input stem constitutes the input of the image coding module, and the output of the fourth residual stage constitutes the output of the image coding module. The input stem consists of a sequentially connected convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. The input of the convolutional layer constitutes the input of the input stem, and the output of the max pooling layer constitutes the output of the input stem. Each residual stage consists of 3, 4, 6, and 3 residual blocks respectively, sequentially connected from input to output. The input of the first residual block in each residual stage constitutes the input of the residual stage, and the output of the last residual block constitutes the output of the residual stage. Each residual block includes a fusion module, an activation module, and two branches. In each residual block structure, the inputs of the two branches are connected to form the input of the residual block. One branch, from its input to its output, includes two convolutional normalization activation modules and one convolutional normalization module connected in series. The input of the first convolutional normalization activation module forms the input of one branch, and the output of the convolutional normalization module forms the output of one branch. The outputs of the two branches are connected to the input of the fusion module, and the output of the fusion module is connected to the input of the activation module. The output of the activation module forms the output of the residual block. In the first residual block of each residual stage structure, the other branch is a convolutional normalization module, and the number of input channels and the number of output channels of the two branches are equal. In the remaining residual blocks of each residual stage structure, the other branch is a direct identity mapping, used to realize the direct connection between the input of the residual block and the input of the fusion module.

3. The driving warning method based on water film fogging and image visibility according to claim 2, characterized in that: The feature enhancement module includes a channel attention module, a spatial attention module, a discrete cosine transform module, and a multi-scale feature fusion module. The input ends of the channel attention module, the spatial attention module, and the discrete cosine transform module are connected to form the input end of the feature enhancement module. The output ends of the channel attention module, the spatial attention module, and the discrete cosine transform module are connected and connected to the input end of the multi-scale feature fusion module. The output end of the multi-scale feature fusion module constitutes the output end of the feature enhancement module. The multimodal fusion and regression head module consists of a global average pooling layer, a fully connected layer FC1, a ReLU activation layer, a Dropout layer, a fully connected layer FC2, and a ReLU activation layer connected in series from the input to the output. The input of the multimodal fusion and regression head module is formed by the input of the global average pooling layer, and the output of the last ReLU activation layer is formed by the output of the multimodal fusion and regression head module.

4. A driving warning method based on water film fogging and image visibility according to any one of claims 1 to 3, characterized in that: The test road consists of various water film thickness sensors located on the roadside ground along the test road direction, and two vehicles traveling at the same speed in adjacent lanes in the test road direction, one in front of the other. The rear vehicle has a built-in satellite positioning module and an image capture device pointing in the direction of the front vehicle. Based on the exhaustive enumeration of three types of data—preset rainfall levels, preset distances between workshops, and preset driving speeds—the following steps a to b are performed for each spatiotemporal combination to obtain individual samples corresponding to each time point under the spatiotemporal combination application. Then, the individual samples corresponding to each time point under each spatiotemporal combination are combined to form a sample set, and step d is performed to achieve model training. Step a. Two vehicles travel on the test road according to the requirements of spatiotemporal combination. The vehicle behind captures video of the vehicle in front through an image capture device. At the same time, the vehicle behind travels through the test road segment corresponding to its positioning information, and determines the positioning information of the vehicle in front in real time by considering the distance between vehicles. It also obtains the preset road features of the test road segment where the vehicle in front is located in real time, and obtains the water film thickness on the test road segment where the vehicle in front is located in real time through the water film thickness sensors on the roadside. Step b. For each video frame, based on the annotations of the boundary polygons of the dense fog area on the video frame and the annotations of the distance from the shooting position to the preset type of reference object, and considering the distance between the two vehicles in front and behind, obtain the standard visibility distance corresponding to the video frame through perspective transformation; according to the time point corresponding to each video frame, obtain the video frame, standard visibility distance, various rainfall characteristics under the spatiotemporal combination, water film thickness of the test road section where the vehicle in front is located, and preset road characteristics corresponding to each time point, thus forming a single sample corresponding to each time point; and thus forming a sample set. Step d. Based on the sample set, using video frames in the samples as input and corresponding visibility distance as output, train the visibility detection model to obtain a trained visibility detection model; based on the sample set, using various rainfall features and road features in the samples as input and corresponding water film thickness as output, train the road surface water film thickness detection model to obtain a trained road surface water film thickness detection model.

5. A system for implementing the driving warning method based on water film fogging and image visibility as described in any one of claims 1 to 4, characterized in that: Based on the image capture device installed on the vehicle, it is used to capture video of vehicles traveling in adjacent lanes ahead along the road direction. It also includes a satellite positioning module, a visibility detection module, an impact distance detection module, and a graded warning module deployed on the vehicle; the vehicle obtains its real-time positioning information based on the satellite positioning module. For each video frame in the captured video, the visibility detection module performs step A to obtain the visibility distance in front of the vehicle; the influence distance detection module performs steps B to C to obtain the current water film thickness of the road segment where the vehicle is located; and the graded warning module performs step D based on the visibility distance in front of the vehicle and the current water film thickness of the road segment where the vehicle is located to obtain the driving warning corresponding to the vehicle.

Images