Distraction driving detection method based on depth separable convolution and multi-spectrum attention

Abstract

The invention provides a distracted driving detection method based on depth separable convolution and multi-spectrum attention, and the method comprises the following steps: firstly, carrying out the efficient feature extraction of an input driver image through a feature extraction module based on depth separable convolution, so as to reduce the calculation complexity of a model and the number of parameters; secondly, a multi-spectrum attention mechanism is designed and integrated, and the capturing capability of the model on key features is enhanced through self-adaptive attention on different spectrum information, so that the accuracy of distraction driving detection is improved; secondly, carrying out diversified processing on training data by adopting a data enhancement technology so as to improve the generalization ability of the model; and finally, constructing a deep neural network architecture containing a maximum pooling layer and a classifier, and realizing real-time detection and classification of distraction driving behaviors. Compared with the prior art, the method has the following advantages: 1) by introducing the depth separable convolution, the calculation amount and the parameter scale of the model are significantly reduced, and the training and reasoning efficiency is improved; 2) a multi-spectrum attention mechanism enhances the attention capability of the model on key features, and improves the accuracy of distraction driving detection; 3) an optimized data enhancement strategy improves the generalization performance of the model and reduces the risk of overfitting; and 4) the whole model is simple in architecture and convenient to deploy and expand in practical application.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning, specifically a distracted driving detection method based on deep separable convolution and multi-spectral attention. This method is suitable for real-time monitoring and detection of driver distraction in intelligent driving systems, aiming to improve detection accuracy and system efficiency. Background Art

[0002] With the rapid development of intelligent transportation systems and autonomous driving technologies, driving safety has become a major concern. Distracted driving behaviors, such as using a mobile phone, eating, or talking to passengers, significantly increase the risk of traffic accidents. Therefore, developing efficient and accurate methods to detect distracted driving is of great practical significance.

[0003] Existing distracted driving detection methods are primarily based on traditional computer vision techniques and machine learning algorithms. These methods typically involve the following steps: first, capturing the driver's facial and body movements with a camera; second, extracting the driver's features using feature extraction algorithms (such as HOG and SIFT); and finally, employing classifiers (such as support vector machines and decision trees) to classify and identify distracted driving behaviors. While these methods can detect distracted driving to a certain extent, the following issues remain:

[0004] Low feature extraction efficiency: Traditional feature extraction methods have high computational complexity when processing high-dimensional data and are difficult to meet the needs of real-time detection.

[0005] Limited recognition accuracy: Due to the limitations of feature extraction methods, it is difficult to fully capture the subtle features of driver distraction behavior, resulting in low recognition accuracy.

[0006] Large model parameters: While existing deep learning models improve their accuracy, they are often accompanied by a sharp increase in model parameters, resulting in high consumption of computing resources during training and inference.

[0007] Insufficient generalization ability: Existing methods perform unstably under different lighting, viewing angles, and driving environments, have poor generalization capabilities, and are difficult to adapt to complex and changing practical application scenarios.

[0008] To overcome these challenges, deep learning techniques, particularly convolutional neural networks (CNNs), have achieved remarkable results in image classification and object detection. Distracted driving detection methods based on deep learning can automatically learn high-level features of driver behavior through end-to-end training, significantly improving detection accuracy and efficiency. However, traditional deep convolutional neural networks suffer from high computational complexity and a large number of parameters during feature extraction, limiting their application in resource-constrained real-time detection systems.

[0009] In recent years, depthwise separable convolution (DSC), as an efficient convolution operation, has been widely used in lightweight neural networks, such as the MobileNet series, significantly reducing the model's computational complexity and parameter count. Furthermore, the attention mechanism (ACM) enhances the model's ability to capture key features by adaptively adjusting the importance of different features, further improving the model's recognition performance.

[0010] Despite the success of depthwise separable convolution and attention mechanisms in their respective fields, their combined application in distracted driving detection remains limited. Existing methods still have room for improvement in achieving efficient feature extraction and enhancing key feature capture, particularly in the utilization of multi-spectral information and efficient feature processing. Therefore, a distracted driving detection method based on depthwise separable convolution and multi-spectral attention mechanisms is urgently needed to improve detection accuracy and real-time performance, thereby meeting the efficiency and safety requirements of intelligent driving systems.

[0011] It is in this context that the present invention is proposed. By introducing deep separable convolution and multi-spectral attention mechanism, an efficient and accurate distracted driving detection method is designed, which overcomes many shortcomings of the existing technology and has broad application prospects and significant technical advantages. Summary of the Invention

[0012] This paper aims to provide a distracted driving detection method based on depthwise separable convolution and multi-spectral attention, aiming to address the challenges of existing distracted driving detection technologies in terms of feature extraction efficiency, recognition accuracy, model parameter size, and generalization. By incorporating depthwise separable convolution and multi-spectral attention mechanisms, this method designs an efficient and accurate distracted driving detection system suitable for real-time monitoring and detection of driver distraction in intelligent driving systems.

[0013] The present invention's distracted driving detection method based on deep separable convolution and multi-spectral attention includes key steps such as image preprocessing, feature extraction, multi-spectral attention enhancement, feature downsampling, feature flattening, and classification detection, including the following steps:

[0014] Step 1. Image preprocessing: The input driver image is preprocessed, including resizing the image to 256×256 pixels, randomly cropping it to 224×224 pixels and padding it, randomly horizontally flipping it, and normalizing it to enhance the generalization ability and robustness of the model.

[0015] Step 2: Feature Extraction: The preprocessed image is fed into a feature extraction module based on depthwise separable convolution. This module consists of multiple depthwise separable convolution layers, which combine depthwise convolution with pointwise convolution to efficiently extract image features, significantly reducing the computational complexity and number of parameters of the model.

[0016] Step 3: Multi-spectral attention enhancement: The extracted features are passed to the multi-spectral attention layer, and the multi-spectral attention mechanism is used to adaptively weight the information of different spectra, enhancing the model's ability to capture key features and improving the accuracy of distracted driving detection.

[0017] Step 4: Feature downsampling: The feature map processed by the attention mechanism is downsampled through the maximum pooling layer, using a 28×28 pooling kernel and a stride of 1 to reduce the feature dimension while retaining important spatial information.

[0018] Step 5: Feature flattening: Flatten the downsampled feature map and convert it into a one-dimensional feature vector to meet the input requirements of the subsequent classifier module.

[0019] Step 6: Classification and Detection: The flattened feature vector is fed into the classifier module. This module, which includes a batch normalization layer, a fully connected layer, a dropout layer, and an ELU activation function, is responsible for real-time detection and classification of distracted driving behaviors, outputting the final detection results.

[0020] Step 7. Result output: Compare the output of the classifier module with the predefined category label, calculate the cross entropy loss, and perform backpropagation and optimize the model parameters based on the loss value to continuously improve the detection accuracy and system stability.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] Efficient feature extraction: By introducing depthwise separable convolution, the computational complexity and parameter size of the model are significantly reduced, and the efficiency of training and inference is improved, enabling the distracted driving detection system to operate efficiently in resource-constrained environments.

[0023] Improved accuracy: The multi-spectral attention mechanism enhances the model's ability to focus on key features, enabling it to more accurately capture subtle characteristics of driver distraction and significantly improve detection accuracy.

[0024] Strong real-time performance: Optimized feature extraction and attention mechanism processing reduce computational latency, enabling real-time monitoring of distracted driving detection and meeting the high real-time performance requirements of intelligent driving systems.

[0025] Model lightweighting: The application of depthwise separable convolution reduces the number of model parameters and computational complexity, lowers the demand for hardware resources, and facilitates deployment and expansion in practical applications.

[0026] Strong generalization capability: The use of diverse data augmentation techniques, including brightness adjustment, random cropping, and horizontal flipping, improves the model's generalization capability under different lighting, viewing angles, and driving environments, ensuring the detection system's stable performance in complex and changing real-world scenarios.

[0027] Easy to train and optimize: The introduction of cosine annealing learning rate scheduling strategy and data parallel technology accelerates the model training process and improves training efficiency. The optimization strategy prevents model overfitting and further improves detection performance.

[0028] By systematically combining depthwise separable convolution and multi-spectral attention mechanism, this paper designs an efficient, accurate and easy-to-deploy distracted driving detection method, which overcomes many shortcomings of the existing technology and has broad application prospects and significant technical advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the system architecture of distracted driving detection method based on depthwise separable convolution and multi-spectral attention

[0030] Figure 2 Schematic diagram of the workflow of the distracted driving detection method

[0031] Figure 3 Schematic diagram of the design framework for the depth-wise separable convolution module

[0032] Figure 4 Schematic diagram of the workflow of the multi-spectral attention mechanism

[0033] Figure 5 Schematic diagram comparing the traditional convolutional neural network (CNN) and the network architecture of the present invention DETAILED DESCRIPTION

[0034] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings.

[0035] A distracted driving detection method based on deep separable convolution and multi-spectral attention, the design block diagram is as follows Figure 1 As shown, its workflow is as follows Figure 2 As shown, it includes the following steps:

[0036] Step 1: Image Preprocessing: When a driver image is input, it first passes through the image preprocessing module, resizing the image to 256×256 pixels, then randomly cropping it to 224×224 pixels with padding, randomly flipping it horizontally, and finally normalizing it. These preprocessing steps are designed to enhance the model's generalization and robustness, ensuring that the model can adapt to different lighting, viewing angles, and driving environments.

[0037] Step 2, Feature Extraction: The preprocessed image is passed to the feature extraction module based on depthwise separable convolution. This module consists of multiple depthwise separable convolution layers, such as Figure 3 As shown in the figure, each depthwise separable convolution layer includes depthwise convolution and pointwise convolution. Depthwise convolution is used to perform spatial convolution operations on the input feature map, while pointwise convolution is used to integrate cross-channel information. This structure significantly reduces the computational complexity and number of parameters of the model while maintaining the effectiveness of feature extraction.

[0038] Step 3: Multi-spectral attention enhancement: The extracted feature map is passed to the multi-spectral attention layer. The specific implementation is as follows Figure 4 As shown in Figure 2, the multi-spectral attention mechanism uses multiple spectral submodules to independently process features in different spectral ranges and dynamically adjusts the weights of each spectral feature during the final feature fusion process. This step aims to enhance the model's ability to focus on key features and improve the accuracy of distracted driving detection.

[0039] Step 4: Feature downsampling: The feature map processed by the attention mechanism is passed to the maximum pooling layer for downsampling, such as Figure 1 As shown in Figure 2, the pooling kernel size is set to 28×28, with a stride of 1, to reduce the feature dimension while retaining important spatial information. This step effectively reduces the computational complexity of subsequent processing.

[0040] Step 5: Feature Flattening: The downsampled feature map is flattened and converted into a one-dimensional feature vector to fit the input requirements of the subsequent classifier module. The flattening process includes resizing the feature map to (batch_size, -1) to ensure that the shape of the feature vector is suitable for the fully connected layer.

[0041] Step 6, Classification and Detection: The flattened feature vector is input to the classifier module. This module includes batch normalization layer, fully connected layer, Dropout layer and ELU activation function. The specific structure is as follows Figure 1 As shown in the figure, the batch normalization layer accelerates training and stabilizes the model. The dropout layer prevents overfitting, and the ELU activation function enhances the model's nonlinear expression capabilities. The classifier module is responsible for real-time detection and classification of distracted driving behaviors, outputting the final detection results.

[0042] Step 7. Result output: The output of the classifier module is compared with the predefined category label, the cross entropy loss is calculated, and backpropagation and model parameter optimization are performed based on the loss value to continuously improve the detection accuracy and system stability.

[0043] Traditional distracted driving detection methods such as Figure 5 As shown in (a), it mainly includes the traditional convolutional neural network (CNN) architecture, which has the following disadvantages:

[0044] Low feature extraction efficiency: Traditional CNN has high computational complexity when processing high-dimensional data, making it difficult to meet the needs of real-time detection.

[0045] Limited recognition accuracy: Due to the limitations of feature extraction methods, it is difficult to fully capture the subtle features of driver distraction behavior, resulting in low recognition accuracy.

[0046] Large model parameters: While existing deep learning models improve their accuracy, they are often accompanied by a sharp increase in model parameters, resulting in high consumption of computing resources during training and inference.

[0047] Insufficient generalization ability: Existing methods perform unstably under different lighting, viewing angles, and driving environments, have poor generalization capabilities, and are difficult to adapt to complex and changing practical application scenarios.

[0048] In order to overcome the above problems, this paper introduces depth-wise separable convolution and multi-spectral attention mechanism, such as Figure 5 As shown in (b), an efficient and accurate distracted driving detection method is designed. The specific implementation steps are as follows:

[0049] Step 1: Depthwise Separable Convolution Feature Extraction: The preprocessed image is fed into the depthwise separable convolution module, which uses depthwise convolution for spatial feature extraction and then performs channel integration through point-by-point convolution. This approach effectively reduces the amount of computation and the number of parameters while maintaining the effectiveness of feature extraction.

[0050] Step 2: Multi-spectral Attention Mechanism: The extracted features are passed to the multi-spectral attention layer, which consists of multiple spectral submodules that independently process features in different spectral ranges. Each spectral submodule adaptively adjusts the feature weight based on the spectral information of the current feature, ensuring the model's sensitivity and recognition of key features.

[0051] Step 3: Feature Downsampling and Flattening: The feature map processed by the attention mechanism is downsampled through a max pooling layer, using a 28×28 pooling kernel and a stride of 1 to reduce the feature dimension. The downsampled feature map is then flattened into a one-dimensional feature vector to fit the input of the fully connected layer.

[0052] Step 4: Classifier Module Design: The flattened feature vector is input into the classifier module, which includes a batch normalization layer, a fully connected layer, a dropout layer, and an ELU activation function. The batch normalization layer accelerates training and stabilizes the model, the dropout layer prevents overfitting, and the ELU activation function enhances the model's nonlinear expression capabilities. The classifier module is responsible for real-time detection and classification of distracted driving behaviors and outputs the final detection results.

[0053] Step 5. Result output and optimization: Compare the output of the classifier module with the predefined category label, calculate the cross entropy loss, and perform backpropagation and optimize the model parameters based on the loss value to improve the detection accuracy and system stability.

Claims

1. A distracted driving detection method based on depthwise separable convolution and multi-spectral attention, characterized by: The following steps are involved: Step 1: In the distracted driving detection system, a feature extraction module based on depthwise separable convolution is implemented to receive the input driver image and extract efficient feature representation; Step 2: Define a multi-spectral attention mechanism and use the multi-spectral attention layer to adaptively focus on the extracted features to enhance the model's ability to capture key features. Step 3: After feature extraction and attention mechanism processing, a pooling module containing a maximum pooling layer is placed to downsample the feature map to reduce the feature dimension; Step 4: Pass the pooled feature map to the classifier module, which includes a batch normalization layer, a fully connected layer, a dropout layer, and an activation function to detect and classify distracted driving behaviors in real time. Step 5: Compare the output of the classifier module with the predefined category label and output the distracted driving detection result.

2. The distracted driving detection method based on depthwise separable convolution and multi-spectral attention according to claim 1, characterized in that The following technical steps are included: The depthwise separable convolution module consists of multiple depthwise separable convolution layers, which is used to significantly reduce the computational complexity and number of parameters of the model while ensuring the feature extraction effect; Multi-spectral attention mechanism, which improves the model's sensitivity and recognition of key features of distracted driving by adaptively weighting different spectral information; An optimized data augmentation strategy performs diverse processing on input images during training, including brightness adjustment, random cropping, and horizontal flipping, to improve the generalization ability of the model; The classifier module uses batch normalization, fully connected layers, Dropout, and ELU activation functions to build an efficient classification network and accurately classify distracted driving behaviors.

3. The distracted driving detection method based on depthwise separable convolution and multi-spectral attention according to claim 1, characterized in that The multi-spectral attention mechanism further includes multiple spectral sub-modules, which independently process features in different spectral ranges and dynamically adjust the weights of each spectral feature in the final feature fusion process.

4. The distracted driving detection method based on depthwise separable convolution and multi-spectral attention according to claim 2, characterized in that The following steps are involved: Step 1: Input the preprocessed driver image into the depthwise separable convolution module and use the depthwise separable convolution layer for efficient feature extraction; Step 2: Pass the extracted features to the multi-spectral attention layer and dynamically adjust the feature weights based on different spectral information; Step 3: Downsample the feature map after attention processing through the maximum pooling layer to reduce the feature dimension; Step 4: Flatten the pooled feature map and input it into the classifier module. The classifier module performs batch normalization, full connection, dropout, and ELU activation calculation in sequence, and finally outputs the classification result of distracted driving behavior; Step 5: Compare the classification results with the true labels, calculate the cross entropy loss, perform backpropagation and optimize the model parameters based on the loss value.

Images