Abnormal detection and early warning method and system for curve driving

Abstract

The invention relates to the technical field of anomaly detection of curve driving, in particular to an anomaly detection and early warning method and system for curve driving, and the method comprises the steps: S1, inputting a preprocessed image into an encoder-decoder model of a convolutional neural network, and obtaining a curve lane detection result; s2, an adjacent curve speed limit sensing early warning algorithm is adopted, a curve merging method is adopted in the algorithm, real-time Beidou position data and vehicle speed information are utilized, and warning can be given out when a series of curves are approached; s3, driving of the vehicle is assisted according to the lane detection result of the curve and the early warning of the curve; by means of the method, it is ensured that the vehicle stably passes through the continuous curve, and the vehicle is prevented from colliding with the curve.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormality detection for curve driving, and in particular to an abnormality detection and early warning method and system for curve driving. Background Art

[0002] The road conditions in mountainous areas are poor and the terrain is complex. There are not only continuous curves, but also broken lane lines, no lane markings, and no separation of two lanes. Although the existing abnormal detection and warning of driving on curves can effectively avoid collisions, in the face of poor road conditions and complex terrain in mountainous areas, there are not only continuous curves, but also broken lane lines, no lane markings, and no separation of two lanes. The existing abnormal detection and warning system for driving on curves does not take into account the special conditions in mountainous areas. Under mountainous road conditions, there is a problem of being unable to accurately identify the lane lines of curves and unable to issue timely and effective warnings when approaching a series of curves. Therefore, even if the vehicle is equipped with the existing abnormal detection and warning system for driving on curves, it still faces a high risk of collision in mountainous areas, which brings great safety hazards to vehicles that often drive in mountainous areas.

[0003] Therefore, the present invention provides an abnormality detection and early warning method for curved driving to solve the above-mentioned problem. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an abnormality detection and early warning method for driving on a curved road, and the technical solution adopted is:

[0005] S1. Input the preprocessed image into the encoder-decoder model of the convolutional neural network to obtain the curved lane detection result;

[0006] S2, using the adjacent curve speed limit perception warning algorithm, which uses the merged curve method, using real-time Beidou position data and vehicle speed information to issue warnings when approaching a series of curves;

[0007] S3. Assisting the vehicle to pass through continuous curves based on the lane detection result of the curve and the warning of the curve.

[0008] The model consists of an encoder, an enhanced convolutional block attention module, and a decoder.

[0009] The encoder extracts basic features from the preprocessed image and builds low-level and high-level feature maps. The features created by the encoder part are passed through the attention module, which focuses on the main features such as lane markings, lane transitions, and lane curvature to extract these features in the road area. The decoder module reconstructs the feature maps collected from the unique attention module to provide a predicted image with the same resolution as the input image.

[0010] The encoder module consists of four residual blocks and four stride convolution layers. Each residual block of the encoder module consists of three convolution layers, three normalization layers, and three activation functions. The four stride convolution layers of the encoder module are responsible for downsampling and focusing on the preprocessed image features with minimal loss. The input feature map is normalized using attention normalization, and the Gaussian error linear unit is used as the activation function of the convolution layer. The activation function of the Gaussian error linear unit is shown in formula (1):

[0011]

[0012] Where l is the input of the Gaussian error linear unit function, tanh is the approximation of the Gaussian cumulative distribution function, a is a constant factor, preferably a=0.043.

[0013] The enhanced convolutional block attention modules perform key processing steps on the refined activation maps, which are the same size as the intermediate activation maps obtained from the encoder stage. These intermediate activation maps come from four stages, each representing a specific stage of the model's feature extraction process. In order to improve the characteristics of the input feature maps, the enhanced convolutional block attention modules must be integrated with these feature maps. Figure 1 The refined activation map obtained after associating with the enhanced convolutional block attention module retains the spatial and channel information of the intermediate activation map.

[0014] The multi-layer perceptron layer is removed from the channel attention module and replaced with a one-dimensional convolutional layer to learn and locate important channel features immediately after average pooling. Subsequently, the output is processed using the sigmoid function and multiplied with the intermediate features to generate the channel-modified output features, as shown in formula (2):

[0015]

[0016] in, is the Cth channel attention map, σ represents the Sigmoid function, which maps the input to a value between 0 and 1. 1 D k represents a one-dimensional convolution operation with a kernel of size k. This convolution operation is performed on the feature map after average pooling to capture patterns and relationships between channels. represents the average intermediate feature map from channel part C.

[0017] In the enhanced spatial attention module, the spatial dimension information of the input feature map is focused and compressed using maximum pooling and average pooling. It is then passed through a sharpening filter to enhance the information of the edge features. In the context of road lane detection, the "enhanced spatial attention" method uses a Sobel sharpening filter to improve spatial attention. The convolutional layer reduces the number of channels in the feature map and passes through a Sigmoid activation function. The Sigmoid function is multiplied by the channel refined features to obtain a spatially refined feature map. This process enables the system to improve and adjust the highway lane edge data. This is mathematically expressed by formula (3):

[0018]

[0019] Among them, σ represents the Sigmoid activation function, f nXn represents a sharpening filter, which is used to improve the spatial details of the spliced ​​feature map using a filter of size nxn, P 1X1 It is a 1*1 point-wise convolution applied to sharpen the feature map. and are the average intermediate feature map and the maximum intermediate feature map from the spatial part, respectively. Represents the concatenation of average pooling and maximum pooling feature maps along the channel dimension.

[0020] The decoder unit in this model is a key component, which includes a concatenation operation, an activation function, a deconvolution block similar to the encoder, and an upsampling layer. Each deconvolution block consists of three consecutive 2D transposed convolution layers, all of which contain the same number of filters. After these convolution layers, attention normalization is applied, and three Gaussian error linear unit activation functions are used in sequence to introduce nonlinearity to the processed data.

[0021] Analyze the current curve and the next curve. It compares the distance cND between them with the predefined merging threshold cMg. If cND is less than cMg, it means the curves are close enough to merge.

[0022] If cND is greater than cMg, an additional check is performed to determine whether cND is still less than the safe distance safe_dist. If cND is less than the calculated safe distance, the current curve is merged with the next curve.

[0023] Merge the current curve and the next curve, collect information about the two curves, and calculate the new total distance tempCdist, which is the sum of the curve distance cD, the next curve distance ncD, and the distance cND from the current curve to the next curve. At the same time, calculate the new average speed limit tempCspeedLimit based on the current curve speed limit cSL and the next curve speed limit ncSL. The new starting point and end point tempCstart and tempCend are determined by the starting point of the current curve and the end point of the next curve, respectively. In addition, the new distance tempCnextCdist to the next curve is calculated. Finally, mark the merged curve as a reverse type.

[0024] For simple or compound curves, the safety distance is directly calculated and a warning is generated. For reverse curves, a different processing method is adopted. First, the average speed limit of the current curve and the next curve is calculated, and then the average speed is used to calculate the safety distance and generate a warning.

[0025] In a second aspect, a system for abnormality detection and early warning for driving on a curve is provided, and the system for abnormality detection and early warning for driving on a curve is used to execute the above-mentioned abnormality detection and early warning method for driving on a curve.

[0026] In a third aspect, a computer device is provided, wherein a processor implements the steps of the method according to any one of claims 1 to 9 when executing the computer program.

[0027] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 9 are implemented.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] An abnormality detection and warning method for curved driving is proposed. The encoder-decoder model based on convolutional neural network is used for accurate lane detection based on attention mechanism. It can improve the accuracy of lane detection when the lane line is damaged, there is no lane marking, and the two lanes are not separated. At the same time, the merged curve method is adopted to use real-time Beidou position data and vehicle speed information to ensure that warnings can be issued in a timely and effective manner when the vehicle approaches continuous curves. The merged curve method uses real-time Beidou position data and vehicle speed information to ensure that warnings can be issued in a timely and effective manner when the vehicle approaches continuous curves, and calculates the speed value of the curve in advance according to the adjacent curve speed limit perception warning algorithm. This method takes into account the speed limit of the upcoming curve and solves a key defect of the existing system. According to the lane detection results of the curve and the warning of the curve, the vehicle drives smoothly through the continuous curves to prevent the vehicle from colliding in the curve. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The present invention is a schematic flow chart of an abnormality detection and early warning method for curved driving. DETAILED DESCRIPTION

[0031] The following will refer to the attached Figure 1 The embodiments of the present invention are described in detail. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0032] The present invention provides an abnormality detection and early warning method for driving on a curved road, and the technical solution adopted is:

[0033] S1. Input the preprocessed image into the encoder-decoder model of the convolutional neural network to obtain the curved lane detection result;

[0034] S2, using the adjacent curve speed limit perception warning algorithm, which uses the merged curve method, using real-time Beidou position data and vehicle speed information to issue warnings when approaching a series of curves;

[0035] S3. Assisting the vehicle to pass through continuous curves based on the lane detection result of the curve and the warning of the curve.

[0036] This paper proposes an encoder-decoder model based on convolutional neural networks for accurate lane detection based on attention mechanism. Since the attention module collects accurate data, the proposed strategy achieves effective and accurate lane recognition and prevents collisions on curves, especially in challenging situations.

[0037] A method for accurate lane detection on roads with poor road conditions, especially those with curves, broken lane lines, no lane markings, and extreme weather conditions, is proposed. To achieve this goal, a lane detection model based on convolutional attention mechanism is proposed. The model consists of an encoder, a boosted convolutional block attention module, and a decoder. The encoder unit extracts features of the input image, while the boosted convolutional block attention module focuses on optimizing the quality of the input image feature map extracted from the encoder, and the decoder provides the output without losing any information of the original image.

[0038] The accuracy of lane detection is extremely important. However, existing technologies focus on high accuracy under structured road conditions, ignore unstructured roads such as cracked lane lines and curved roads, and usually have complex architectures.

[0039] This paper proposes a model for lane detection on unstructured roads. The preliminary preprocessing module aims to reduce the size of the input image, thereby reducing memory usage, and applies linear interpolation to estimate the missing data in the dataset. This helps to speed up the processing. In addition, the RGB color image is converted to a grayscale image.

[0040] The preprocessed image is then fed into the unstructured road lane detection model, which consists of an encoder, enhanced convolutional block attention module, and decoder. The encoder extracts basic features from the preprocessed image and builds low-level (edges, corners, textures) and high-level (lane boundaries, lane markings, lane curvature, lane transitions, and road geometry) feature maps. The features created in the encoder part are passed through the attention module, which focuses on the main features such as lane markings, lane transitions, and lane curvature to extract these features in the road area. The decoder module reconstructs the feature maps collected from the unique attention module to provide a predicted image with the same resolution as the input image.

[0041] The encoder module consists of four residual blocks and four strided convolutional layers. Each residual block of the encoder module consists of three convolutional layers, three normalization layers, and three activation functions, which ensures reliable feature extraction and transformation. The four strided convolutional layers of the encoder module are responsible for downsampling and focusing on the preprocessed image features with minimal loss. The input feature maps are normalized using attention normalization, which is an important step to stabilize and improve the training process. Gaussian error linear units are used as the activation function of the convolutional layers, which combines the advantages of rectified linear units with random dropout. The problem with rectified linear units is that when the input lane image of the neurons in the network has negative values, it is treated as zero. Therefore, to overcome this problem, the Gaussian error linear unit activation function was selected in this study to achieve a smoother flow and avoid overfitting problems. In the residual block, the first convolutional layer uses a filter size of 1×1 and generates a feature map of dimension "n", where "n" is the number of feature maps generated by this convolutional layer in this residual block. In this block, the 1×1 convolution acts as a point-by-point convolution, enabling the model to modify channel-level data and represent complex relationships between different features. The main advantage of this filter is that it focuses on the main features, making it easier for the network to process. Next, attention normalization adaptively modifies the normalization parameters such as scaling and resizing, which focus on relevant features such as lanes. With the help of the attention mechanism, the network can dynamically scale and change the normalization based on the importance of different features. This leads to more efficient learning and improved model performance. Gaussian error linear unit is used as an activation function to improve the performance of deep learning models without falling into the gradient vanishing problem. The activation function of the Gaussian error linear unit is shown in formula (1):

[0042]

[0043] Where l is the input of the Gaussian error linear unit function, tanh is the approximation of the Gaussian cumulative distribution function, a is a constant factor, preferably a=0.043.

[0044] The Gaussian Error Linear Unit is a smooth, differentiable function that is useful for gradient-based optimization and training. Due to its zero-centered property, the vanishing gradient problem is alleviated, which helps the convergence of the neural network. Unlike the hyperbolic tangent or sigmoid function, the Gaussian Error Linear Unit does not reach saturation at higher input values. In addition, a filter size of 3×3 is used, which is the second convolutional layer, to transfer the features extracted from the Gaussian Error Linear Unit. Unlike the first 1×1 convolution, which mainly focuses on modifying channel-level information and collecting complex relationships, the 3×3 convolutional layer captures hierarchical features in the input data. The second convolutional layer enables the model to capture fine-grained features and local relationships in the data by using a 3×3 filter. Due to its large filter size, the residual block is able to recognize and consider patterns that contain multiple pixels, thereby enhancing the recognition of deeper patterns in the input. In order to strike a balance between collecting local and global features, multiple convolutional layers with different filter sizes are usually used in the residual block. Within the residual block, the last convolutional layer uses a 1×1 filter of size n*2. This method of applying 1×1 convolutions helps control the computational complexity while retaining the ability to detect complex patterns in the data.

[0045] Skip connections are proposed to solve the problem of vanishing gradients. The output of residual block 1 after the Gaussian error linear unit layer is integrated through a skip connection and sent as input to the second residual block, and so on. For the gradient during backpropagation, this creates a direct channel for the skip connection to alleviate the gradient vanishing problem, ensuring that the model can effectively learn from the input data. The skip connection effectively acts as a shortcut to make the gradient flow smoother in deep learning models. Once the output of the residual block passes through the strided convolution layer, it is downsampled to half the size of the input feature map. In the final encoder block, the downsampled and transformed representation of the input data is reflected in a feature map of dimension . Finally, it extracts low-level and high-level features and inputs them into the enhanced convolution block attention module.

[0046] The enhanced convolutional block attention modules perform key processing steps on the refined activation maps, which are the same size as the intermediate activation maps obtained from the encoder stage. These intermediate activation maps come from four stages, each representing a specific stage of the model's feature extraction process. In order to improve the characteristics of the input feature maps, the enhanced convolutional block attention modules must be integrated with these feature maps. Figure 1The refined activation map obtained after association with the enhanced convolutional block attention module retains the spatial and channel information of the intermediate activation map. In this way, the model is able to selectively focus on and emphasize important features, thereby improving the overall quality of feature representation.

[0047] The main goal of the enhanced channel attention module is to dynamically enhance features that convey important information while suppressing those that are less important. The channel attention module uses global data from the intermediate map to guide the network architecture and helps to identify the importance of different channels of the feature map in the intermediate features. When lane markings disappear completely from the road or lane boundaries are invisible, the existing technology cannot identify edge information or road boundaries. Sometimes, typical channel attention modules waste computational resources and lead to parameter accumulation. In this study, the multi-layer perceptron layer is removed from the channel attention module and replaced with a one-dimensional convolutional layer to learn and locate important channel features immediately after average pooling. Subsequently, the output is processed using the sigmoid function and multiplied with the intermediate features to generate the channel-improved output features as shown in formula (2):

[0048]

[0049] in, is the Cth channel attention map, σ represents the Sigmoid function, which maps the input to a value between 0 and 1. 1 D k represents a one-dimensional convolution operation with a kernel of size k. This convolution operation is performed on the feature map after average pooling to capture patterns and relationships between channels. represents the average intermediate feature map from channel part C.

[0050] The output of the channel refined features from the enhanced channel attention module is sent to the enhanced spatial attention module. In the enhanced spatial attention module, the spatial dimension information of the input feature map is focused and compressed using maximum pooling and average pooling. It is then passed through a sharpening filter to enhance the information of the edge features. In the context of road lane detection, the enhanced spatial attention method uses a Sobel sharpening filter to improve spatial attention. The enhanced channel attention of the convolutional block attention module is similar to the enhanced spatial attention module, which focuses on determining the optimal location and intensity of the road lane boundary. The goal is to improve the detection accuracy of the road lane edge. Once the mapping feature channel is narrowed, the pooled output is passed through a sharpening filter with a weighted assignment. In addition, the convolution layer reduces the number of channels of the feature map and passes through a Sigmoid activation function. The Sigmoid function is multiplied by the channel refined features to obtain a spatially refined feature map. This process enables the system to improve and adjust the highway lane edge data. This is mathematically represented by formula (3):

[0051]

[0052] Among them, σ represents the Sigmoid activation function, f nXn represents a sharpening filter, which is used to improve the spatial details of the spliced ​​feature map using a filter of size nxn, P 1X1 It is a 1*1 point-wise convolution applied to sharpen the feature map. and are the average intermediate feature map and the maximum intermediate feature map from the spatial part, respectively. Represents the concatenation of average pooling and maximum pooling feature maps along the channel dimension.

[0053] The average pooling operation of the channel attention module makes the process scale-invariant by capturing contextual information regardless of the input size. The one-dimensional convolutional layer uses a fixed-size pooled feature vector, which is flexible enough to work at any resolution. Due to its weight sharing and local receptive field, the convolutional filters applied to the spatial map in the spatial attention module can adapt to different spatial dimensions. Through element-wise rescaling multiplication, the convolutional block attention module is able to successfully improve feature representations at various scales and resolutions by matching the attention map with the input feature map, regardless of the resolution. Finally, the features refined by the enhanced convolutional block attention module and the features from the last set of strided convolutional layers of the encoder module are transferred to the decoder module, which will be discussed in the next section. In this work, the edge data of the extracted features is strengthened using 3×3 first-order Sobel sharpening.

[0054] The decoder unit in this architecture is a key component, which includes a concatenation operation, an activation function, a deconvolution block similar to the encoder, and an upsampling layer. Each deconvolution block consists of three consecutive 2D transposed convolution layers, all of which contain the same number of filters (1×1, 3×3, and 1×1).

[0055] After these convolutional layers, attention normalization is applied and three Gaussian error linear unit activation functions are used in sequence to introduce nonlinearity to the processed data. The process first collects the outputs from each enhanced convolutional block attention module, including enhanced convolutional block attention module 1, the second enhanced convolutional block attention module 2, enhanced convolutional block attention module 3 and enhanced convolutional block attention module 4, which are combined with a 2×2 upsampling layer in the decoder. The concatenation process specifically pairs the outputs of enhanced convolutional block attention module 1 and enhanced convolutional block attention module 2, and the outputs of enhanced convolutional block attention module 3 and enhanced convolutional block attention module 4 to ensure that no features are lost in this integration step. In order to construct two activation maps, the concatenated outputs are then carefully combined, emphasizing the effective integration of low-level and high-level features. It is worth noting that enhanced convolutional block attention module 2 and enhanced convolutional block attention module 4 undergo a 2×2 upsampling process before concatenation to match the form of enhanced convolutional block attention module 1 and enhanced convolutional block attention module 3. After this, three parallel deconvolution blocks are used, whose inputs are the 2×2 upsampling result, the enhanced convolution block attention module 1, and the enhanced convolution block attention module 2. The driving force behind the use of 2×2 upsampling at this stage is the need to recover the original features from each encoder block. The deconvolution block further receives inputs upsampled to 2×2, 4×4, and 16×16 to maintain consistency with the original input image scale. This consistency ensures that the reconstructed features are aligned with the initial input structure. The last stage is to create a single map that combines the features of the lowest level, the best level, and the refined level. This result is obtained by concatenating these three feature maps. Next, a convolution with a single filter is used to create a predicted image. The entire architecture is designed to collect and aggregate data at multiple scales to provide deeper and more accurate predictions.

[0056] Datasets were used to train the proposed model, such as image datasets of different curved lanes and damaged road images in road images. The proposed method was trained to retain important spatial information for accurate prediction.

[0057] Lane detection with convolutional attention mechanism is able to effectively reduce false positives and false negatives while accurately detecting key features. Due to its high precision and recall, the model is the most effective among the tested models, while also demonstrating its excellent performance in maintaining a good balance. The proposed method outperforms other methods in all evaluation metrics. This shows that the proposed model will be more adaptable when used to create real self-driving vehicles that can identify lanes in real time to avoid accidents.

[0058] The user inputs the start and end locations, and the system uses the BeiDou positioning system to retrieve the shortest route from a commercial map database. Through a continuous curve detection algorithm, the system extracts information about horizontal curves on the retrieved route. This information includes curve attributes (radius, start / end point) and associated speed limits. The extracted curve attributes and speed limits are transmitted to this stage, and the system continuously monitors the vehicle's current speed and the speed limit of the approaching curve. Based on the predetermined trigger distance and real-time measurements, a pre-recorded warning message designed specifically for the driver is activated before reaching the curve. The core function of the proposed system is the adjacent curve speed limit perception warning; this method, which is sensitive to the speed limit of adjacent curves, plays a vital role. It adopts a "merged curve method" that uses real-time BeiDou position data and vehicle speed information to ensure timely and effective warnings when approaching a series of curves.

[0059] The current speed (Vc) and position (Vl) of the vehicle are obtained. In addition, it considers a reference value (cMg) that specifies the distance threshold for merging adjacent curves.

[0060] Distance-based merging: Analyzes the current and next curves. It compares the distance between them (cND) with a predefined merging threshold (cMg). If cND is less than cMg, it means the curves are close enough to merge.

[0061] Merge based on safe distance: If cND is greater than cMg, an additional check is performed to determine whether cND is still less than the safe distance (safe_dist). If cND is less than the calculated safe distance, the current curve is merged with the next curve.

[0062] Merge curve attributes: Merge the current curve and the next curve, collect information about the two curves, and calculate the new total distance (tempCdist), which is the sum of the curve distance (cD), the next curve distance (ncD), and the distance from the current curve to the next curve (cND). At the same time, calculate the new average speed limit (tempCspeedLimit) based on the current curve speed limit (cSL) and the next curve speed limit (ncSL). The new start and end points are determined by the start point of the current curve and the end point of the next curve, respectively. In addition, the new distance to the next curve is calculated). Finally, mark the merged curve as a reverse type.

[0063] Warning generation for merging curves: For simple or compound curves, the safety distance is directly calculated and a warning is generated.

[0064] For reverse curves, a different approach is used: First, the average speed limit of the current curve and the next curve is calculated, and then the average speed is used to calculate the safe distance and generate an early warning.

[0065] Handling possible lower speed limits in merging curves: The speed limit in a sub-curve may be lower than the overall average speed limit used in the main curve. For example: the speed limit in curve 1 is 50km / h, the speed limit in curve 2 is 20km / h, and the average speed after merging is 35km / h. However, when the vehicle approaches curve 2, where the actual speed limit is 20km / h, the speed needs to be adjusted accordingly. This loop ensures that lower speed limits in merging curves are recognized and handled.

[0066] Speed ​​reduction: includes safe distance (safe_dist), curve speed limit (cSL), curve direction (cB), vehicle current speed (Vc) and position (Vl). It calculates the distance between the vehicle and the safety area, and if the vehicle speed exceeds the curve speed limit, it will start the deceleration operation. In the ADAS system, the driver will receive an audio warning to reduce speed, and also get information about the curve direction.

[0067] The method proposed in the present invention uses map path data and existing curve detection systems to extract curvature information. Then, this information is used to warn the driver of the upcoming curve in the safest possible position, taking into account the speed limits of adjacent curves, and the lane lines can be accurately detected, greatly reducing the risk of collision of vehicles on curves in mountainous areas. Although existing curve warning systems can already warn of hidden curves on uphill roads, these systems often do not take into account the speed limits and distances of subsequent curves when issuing warnings, especially when encountering a series of curves in succession. In this case, it is necessary to quickly decelerate from a higher speed limit to a lower speed limit in a short distance, which makes it difficult to calculate the optimal warning position, which significantly increases the risk of accidents.

[0068] This limitation is addressed by dynamically calculating a safe position for driver warnings. This safe position takes into account both the speed limit of the current curve and the speed limit of the curve that immediately follows it. By incorporating this comprehensive assessment, the proposed system aims to improve driver safety and reduce risks when negotiating a series of curves on uphill roads. The "merging curve method" uses real-time Beidou position data and vehicle speed information to ensure that warnings are issued in a timely and effective manner when the vehicle approaches consecutive curves.

[0069] The adjacent curve speed limit perception and warning algorithm calculates the appropriate speed for the next curve in advance, allowing the car to pass through consecutive curves smoothly at the optimal speed.

[0070] Based on the lane detection results of the curve and the warning assistance of the curve, the vehicle can pass through continuous curves smoothly and prevent vehicle collisions on the curve.

[0071] The present invention provides a system for abnormality detection and early warning for driving on a curve. The system for abnormality detection and early warning for driving on a curve is used to execute the above-mentioned abnormality detection and early warning method for driving on a curve.

[0072] The above-mentioned unit modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above-mentioned modules.

[0073] This embodiment also provides a computer device, which can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC or other technologies. When the computer program is executed by the processor, a method for detecting power transmission corrosion defects based on semi-supervised learning is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0075] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0076] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0077] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0078] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0079] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0080] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for detecting and warning abnormalities when driving on a curve, characterized in that: include: S1. Input the preprocessed image into the encoder-decoder model of the convolutional neural network to obtain the curved lane detection result; S2, using the adjacent curve speed limit perception warning algorithm, which uses the merged curve method, using real-time Beidou position data and vehicle speed information to issue warnings when approaching a series of curves; S3. Assisting the vehicle to pass through continuous curves based on the lane detection result of the curve and the warning of the curve.

2. The abnormality detection and warning method for curved driving according to claim 1, characterized in that: The encoder extracts basic features from the preprocessed image and builds low-level and high-level feature maps. The features created by the encoder part are transmitted through the attention module, which focuses on the main features such as lane markings, lane transitions, and lane curvature to extract these features in the road area. The encoder module consists of four residual blocks and four strided convolutional layers. Each residual block of the encoder module consists of three convolutional layers, three normalization layers, and three activation functions. The four strided convolutional layers of the encoder module are responsible for downsampling and focusing on the preprocessed image features with minimal loss. The input feature map is normalized using attention normalization, and the Gaussian error linear unit is used as the activation function of the convolutional layer. The activation function of the Gaussian error linear unit is shown in formula (1): Where l is the input of the Gaussian error linear unit function, tanh is the approximation of the Gaussian cumulative distribution function, a is a constant factor, preferably a=0.

043.

3. The abnormality detection and warning method for curved driving according to claim 2, characterized in that: The enhanced convolutional block attention modules perform key processing steps on the refined activation maps, which are the same size as the intermediate activation maps obtained from the encoder stage. These intermediate activation maps come from four stages, each representing a specific stage of the model's feature extraction process. In order to improve the characteristics of the input feature maps, the enhanced convolutional block attention modules must be processed together with these feature maps. The refined activation maps obtained after associating with the enhanced convolutional block attention modules retain the spatial and channel information of the intermediate activation maps.

4. The abnormality detection and warning method for curved driving according to claim 3, characterized in that: The multi-layer perceptron layer is removed from the channel attention module and replaced with a one-dimensional convolutional layer to learn and locate important channel features immediately after average pooling. Subsequently, the output is processed using the sigmoid function and multiplied with the intermediate features to generate the channel-modified output features, as shown in formula (2): in, is the Cth channel attention map, σ represents the Sigmoid function, which maps the input to a value between 0 and 1. 1D k represents a one-dimensional convolution operation with a kernel of size k. This convolution operation is performed on the feature map after average pooling to capture patterns and relationships between channels. represents the average intermediate feature map from channel part C.

5. The abnormality detection and warning method for curved driving according to claim 4, characterized in that: In the enhanced spatial attention module, the spatial dimension information of the input feature map is focused and compressed using maximum pooling and average pooling. It is then passed through a sharpening filter to enhance the information of the edge features. In the context of road lane detection, the enhanced spatial attention method uses a Sobel sharpening filter to improve spatial attention. The convolution layer reduces the number of channels in the feature map and passes through a Sigmoid activation function. The Sigmoid function is multiplied by the channel refined features to obtain a spatially refined feature map. This process enables the system to improve and adjust the highway lane edge data. This is mathematically represented by formula (3): Among them, σ represents the Sigmoid activation function, f nXn represents a sharpening filter, which is used to improve the spatial details of the spliced ​​feature map using a filter of size nxn, P 1X1 It is a 1*1 point-wise convolution applied to sharpen the feature map. and are the average intermediate feature map and the maximum intermediate feature map from the spatial part, respectively. Represents the concatenation of average pooling and maximum pooling feature maps along the channel dimension.

6. The abnormality detection and early warning method for driving on a curved road according to claim 5, characterized in that: The decoder module reconstructs the feature maps collected from the unique attention module to provide a predicted image with the same resolution as the input image. The decoder unit in this model is a key component, which includes a splicing operation, an activation function, a deconvolution block similar to the encoder, and an upsampling layer. Each deconvolution block consists of three consecutive 2D transposed convolution layers, all of which contain the same number of filters. After these convolution layers, attention normalization is applied, and three Gaussian error linear unit activation functions are used in sequence to introduce nonlinearity to the processed data.

7. The abnormality detection and early warning method for driving on a curve according to any one of claims 1 to 6, characterized in that: Analyze the current curve and the next curve. It compares the distance cND between them with the predefined merge threshold cMg. If cND is less than cMg, it means that the curves are close enough to merge; if cND is greater than cMg, an additional check will be performed to determine whether cND is still less than the safe distance safe_dist. If cND is less than the calculated safe distance, the current curve and the next curve are merged.

8. The abnormality detection and warning method for curved driving according to claim 7, characterized in that: Merge the current curve and the next curve, collect information about the two curves, and calculate the new total distance tempCdist, which is the sum of the curve distance cD, the next curve distance ncD, and the distance cND from the current curve to the next curve. At the same time, calculate the new average speed limit tempCspeedLimit based on the current curve speed limit cSL and the next curve speed limit ncSL. The new starting point and end point tempCstart and tempCend are determined by the starting point of the current curve and the end point of the next curve, respectively. In addition, the new distance tempCnextCdist to the next curve is calculated. Finally, mark the merged curve as a reverse type.

9. The abnormality detection and early warning method for driving on a curved road according to claims 7-8, characterized in that: For simple or complex curves, the safety distance is directly calculated and an early warning is generated; For reverse curves, a different approach is used. First, the average speed limit of the current curve and the next curve is calculated, and then the average speed is used to calculate the safe distance and generate an early warning.

10. A system for abnormal detection and early warning of driving on a curve, characterized in that: The system for abnormal detection and early warning of driving on a curve is used to execute the abnormal detection and early warning method for driving on a curve as described in any one of claims 1-9.

Images