A method for recognizing distracted driving behavior based on an improved DCH neural network

Abstract

The application discloses a method for recognizing distraction driving behavior based on an improved DCH neural network, which comprises the following steps: constructing a hyper-dimensional calculation model; firstly, constructing an improved DCH neural network; replacing an ADAM optimizer in the DCH model with an improved artificial traveling mouse optimizer; then, using the improved DCH neural network to perform hyper-dimensional vector coding; in a hash code space, the Hamming distance of hash codes of semantically similar images is close, and the Hamming distance of semantically dissimilar images is far; finally, classifying image information of the same category to form a category hyper-dimensional vector; constructing an AdaptHD retraining model; using the model to calculate a classification accuracy rate, and calculating the similarity between all training samples and the category hyper-dimensional vector again; training the model, testing a data set by using the trained model; and outputting a test result; and the application improves the efficiency of the algorithm within an acceptable range of detection precision, reduces the time for training the model, and no longer depends on high-performance edge computing devices.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method for recognizing distracted driving behavior based on an improved DCH neural network. Background Technology

[0002] Currently, most methods for identifying distracted driving behavior fall into two main categories: those based on vehicle dynamics and those based on driver state. The former identifies driver behavior by capturing vehicle state information, such as speed, steering angle, and acceleration. It determines driving state based on vehicle speed, acceleration, and wheel angular velocity; these vehicle dynamics parameters reflect the vehicle's state, not the driver's behavior. A distracted driver might maintain straight-line driving for several seconds, resulting in no significant change in vehicle data. The system must wait a sufficiently long time until the distracted behavior actually affects vehicle control before abnormal data patterns appear. Therefore, the real-time performance of this method cannot be guaranteed, and it fails to provide preventative or warning effects.

[0003] The latter method assesses the driver's driving status based on physiological parameters, typically using eye-tracking data or a combination of computer vision and deep learning to determine posture. While wearable devices can collect physiological parameters like heart rate and brainwaves, these methods have limitations. Wearable devices may interfere with driving behavior. Eye-tracking data, on the other hand, cannot obstruct the driver's eyes, such as by prohibiting sunglasses.

[0004] The methods described above are other existing approaches to identifying distracted driving behavior, and these methods can all detect dangerous driving behaviors to some extent. However, these methods all apply neural networks to detect targets. As the difficulty of the task increases, the training scale of neural network models becomes increasingly massive. For example, different modalities of data, differences in brightness between day and night (RGB modal data is unusable at night), data from different angles, and complex in-vehicle environments all make the task of classifying distracted driving behavior increasingly difficult, thus increasing the number of model parameters and computational load. Currently, the number of training parameters for these neural networks is generally over one million. The huge computational load leads to long training times and increasingly demanding hardware requirements. The computation of massive amounts of unaudited parameter tuning and floating-point data is difficult to perform on IoT, embedded, and other devices. Summary of the Invention

[0005] To address the aforementioned technical problems, this technical solution provides a distracted driving behavior recognition method based on an improved DCH neural network. The improved DCH neural network utilizes an enhanced artificial lemming optimizer, resulting in a model with higher robustness and faster convergence speed; effectively solving the technical issues.

[0006] This invention is achieved through the following technical solution:

[0007] A method for recognizing distracted driving behavior based on an improved DCH neural network includes the following steps:

[0008] S1: Construct the dataset and preprocess it;

[0009] S2: Constructing a hyperdimensional computational model:

[0010] S2.1: Constructing an improved DCH neural network: Replacing the ADAM optimizer in the DCH model with an improved artificial lemming optimizer;

[0011] When improving the artificial lemming algorithm, the random number rand is optimized using the Circle chaotic mapping; and when lemmings migrate long distances to explore, the learning mechanism of the eagle-fish algorithm is added to the population.

[0012] S2.2: Use an improved DCH neural network for hyperdimensional vector encoding;

[0013] S2.3: Classify image information of the same category to form a category hyperdimensional vector;

[0014] S3: Construct the AdaptHD retraining model: Use this model to calculate the accuracy of hyperdimensional vector classification and calculate the similarity between the hyperdimensional vectors of the training samples and the hyperdimensional vectors of the categories.

[0015] S4: Train the model and test it on the dataset; output the test results.

[0016] Furthermore, the optimization of the random number rand using the Circle chaotic mapping specifically involves employing a spiral winding mechanism to simulate the random foraging behavior of lemmings, the expression of which is:

[0017] ;

[0018] in, This represents the position of the i-th search target at iteration (t+1) during the foraging process. This represents the current optimal solution; F is used as a flag to change the search direction, avoiding local optima and providing the search agent with a higher chance to rigorously scan the problem domain; that is:

[0019] ;

[0020] in, To round down;

[0021] The spiral shape representing random searching during foraging is expressed as:

[0022] ;

[0023] ;

[0024] The radius represents the foraging range, which is the Euclidean distance between the current position and the optimal solution;

[0025] The circle chaotic mapping is used to apply the rand value in the helical winding mechanism, and its expression is:

[0026] ;

[0027] in, For random values ​​after using the Circle chaotic mapping, is the original random value, and mod is the modulo function.

[0028] This method utilizes the Circle chaotic mapping to make the mapping values ​​more average; all settings should conform to the range of rand values ​​[0,1].

[0029] Furthermore, during the long-distance migration and exploration of lemmings, the direction and distance of their migration are affected by ecological and environmental factors. To simulate this behavior, the following equation is proposed:

[0030] ;

[0031] In the above formula, Let represent the position of the i-th search target at iteration (t+1) during long-distance migration exploration. BM describes the characteristics of Brownian motion, utilizing dynamic and uniform step sizes to explore potential regions in the search space; R is a space of size . A vector whose elements are random numbers uniformly distributed in the interval [-1, 1];

[0032] An eaglefish algorithm learning mechanism was added within the population, and its expression is:

[0033] ;

[0034] in, for The optimized value, in The bias was increased; Let w be the optimal individual position element within the population, and w be the learning coefficient. This represents the position of the i-th search target at iteration t.

[0035] Furthermore, the improved DCH neural network includes an input layer, three sets of convolutional downsampling layers, a fully connected layer, a fully connected hash layer, a Cauchy cross-entropy loss function, and an output layer. After the image is input from the input layer, it undergoes three convolutional downsampling processes, then a fully connected layer is used to construct a multi-dimensional one-dimensional vector, and finally the fully connected hash layer and the Cauchy cross-entropy loss function output the required multi-dimensional vector.

[0036] Furthermore, the hyperdimensional vector encoding described in step 2.2 is performed as follows: using an improved DCH neural network, a Cauchy cross-entropy loss function is used for classification testing, and normalized Hamming distance is used to determine the classification: in the hash code space, similar images will have similar hash codes; semantically similar images have a close Hamming distance in their hash codes, while semantically dissimilar images have a far Hamming distance.

[0037] Furthermore, the similarity calculation in step S3 requires re-encoding and standardizing the images in the training samples to obtain new image hyperdimensional vectors; then, the similarity between this hyperdimensional vector and the category hyperdimensional vector is calculated; the similarity calculation method includes:

[0038] When the hyperdimensional vector is encoded using a binary data type, the normalized Hamming distance is used to measure similarity.

[0039] When the hyperdimensional vector is a non-binary hyperdimensional vector, the cosine distance is used to calculate the similarity between the two vectors.

[0040] Beneficial effects

[0041] The present invention proposes a distracted driving behavior recognition method based on an improved DCH neural network, which has the following advantages compared with existing technologies:

[0042] (1) This invention improves the DCH neural network by replacing the ADAM optimizer in the DCH model with an improved artificial lemming optimizer. Furthermore, during the improvement of the artificial lemming algorithm, the random number rand is optimized using the Circle chaotic mapping. Additionally, during the long-distance migration exploration of the lemmings, an eaglefish algorithm learning mechanism is added to the population, resulting in higher robustness and faster convergence speed for the artificial lemming algorithm. This method improves the efficiency of the algorithm within an acceptable range of detection accuracy, significantly reduces the training time of the model, and no longer relies on high-performance edge computing devices. Moreover, compared to traditional encoding methods, in the hash code space, semantically similar images have a very close Hamming distance in their hash codes, while dissimilar images have a very large Hamming distance. This encoding method reduces the amount of data, removes redundant information, and retains core feature data, thereby improving the accuracy of recognition.

[0043] (2) Through the above improvements, this invention solves the problem of strong binding between two neurons in a neural network in deep learning, and eliminates the need for a large number of neurons to process data, thus giving the model strong nonlinear capabilities. Instead, it uses hyperdimensional vectors with dimensions in the thousands to represent different types of data. These data are mostly stored in binary and integer form, reducing the amount of data computation and eliminating the need for a large amount of training time, allowing the model to run on edge computing devices with lower computing power. It no longer relies on devices with powerful computing power to train the model, nor does it rely on high-performance edge computing devices. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0045] Figure 2 This is a schematic diagram of the framework of the hyperdimensional computation model in an embodiment of the present invention.

[0046] Figure 3 This is an architecture diagram of the artificial lemming optimizer in an embodiment of the present invention.

[0047] Figure 4 This is a diagram of the improved DCH neural network architecture in an embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of the training process of the hyperdimensional computation model in an embodiment of the present invention.

[0049] Figure 6 This is a schematic diagram of the identification process of the hyperdimensional computing model in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. The described embodiments are merely some embodiments of the present invention, and not all embodiments. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention should fall within the protection scope of the present invention.

[0051] Example 1:

[0052] like Figure 1 As shown, a method for recognizing distracted driving behavior based on an improved DCH neural network includes the following steps:

[0053] S1: Construct the dataset and preprocess it; the specific steps are as follows:

[0054] The system collects images of distracted driving behavior from different angles, including a side-mounted camera on the passenger side to capture images of the driver's overall posture, and a camera directly in front of the driver to capture images. Furthermore, due to the different brightness levels during the day and night, images need to be categorized using different modalities: RGB images for daytime and near-infrared images for nighttime.

[0055] like Figure 2 As shown, the categories of distracted driving behaviors include: safe driving, sleeping, yawning, using a mobile phone, making phone calls, drinking water, operating GPS, operating the radio, communicating with passengers, and fixing makeup. Furthermore, to further expand the dataset, operations such as rotation, zooming in, and zooming out were used to enrich the dataset.

[0056] The acquired images are preprocessed to standardize their scale. Since the input images may vary in size, preprocessing is necessary. This involves increasing the image scale using nearest neighbor interpolation or decreasing it using convolution operations, ultimately resizing the images to 640*640*3 pixels.

[0057] Finally, the dataset was split, with the ratio of training set to test set being 7:3.

[0058] S2: Constructing a hyperdimensional computational model:

[0059] S2.1: Constructing an improved DCH neural network: Replacing the ADAM optimizer in the DCH model with an improved artificial lemming optimizer; when improving the artificial lemming algorithm, optimizing the random number rand using a Circle chaotic mapping; and adding the eaglefish algorithm learning mechanism to the population during long-distance lemming migration exploration; specifically:

[0060] like Figure 3As shown, the artificial lemming algorithm is a population-based algorithm that mathematically models four different behaviors of lemmings in nature: long-distance migration, burrowing, foraging, and predator avoidance. These behavioral characteristics are highly compatible with the dynamic balance requirement of "exploration-exploitation" in optimization problems. Furthermore, the artificial lemming algorithm introduces an energy diminishing mechanism, which can dynamically adjust the balance between exploration and exploitation, thereby improving the algorithm's resistance to local optima and its robustness.

[0061] The Artificial Lemming Algorithm initializes the positions of all search agents before entering the iterative process. The set of all initial candidate solutions is a matrix consisting of N rows and Dim columns between a given upper and lower bound, where N is the population size and Dim is the number of dimensions. The best position in each iteration is considered the optimal solution or the best solution obtained so far. The decision variables are calculated for each dimension using the following equation:

[0062] ;

[0063] Calculate the decision variables for each dimension using the following equation. :

[0064] ;

[0065] Where rand is a random value in the range of 0-1. Let j be the lower bound of the j-th dimension. Let be the upper bound of the j-th dimension; Z is the decision variable, where i and j are different dimensions.

[0066] When lemmings migrate long distances to explore, the direction and distance of their migration are affected by ecological factors. To simulate this behavior, the following equation is proposed:

[0067] ;

[0068] In the above formula, This represents the position of the i-th search target at iteration (t+1) during the long-distance migration exploration process; Represents the optimal individual location element within the population; R is a variable of size . A vector whose elements are random numbers uniformly distributed in the interval [-1, 1];

[0069] F is used as a flag to change the search direction, that is:

[0070] ;

[0071] This can help avoid local optima and provide a higher chance for the search agent to rigorously scan the problem domain. To round down;

[0072] Brownian motion (BM) describes the characteristics of Brownian motion, utilizing dynamic and uniform step sizes to explore potential regions in the search space; the step size of standard Brownian motion is obtained through the probability density function of a normal distribution with variance of 1 and mean of 0, i.e.:

[0073] .

[0074] To enhance the convergence speed of the artificial lemming algorithm, an eagle-fish algorithm learning mechanism was added within the population, the expression of which is:

[0075] ;

[0076] in, Let w be the optimal individual position element within the population, and w be the learning coefficient. This represents the position of the i-th search target at iteration t. This indicates the optimized position.

[0077] When lemmings dig burrows in their habitat, they randomly dig new burrows based on the location of their current burrow and the location of random individuals within the population; the expression for this is:

[0078] ;

[0079] in, This represents the position of the i-th search target at iteration (t+1) during the hole-digging process; L is a random number related to the current iteration number. Let L represent a search individual randomly selected from the population, where b is a random integer index value between 1 and N. Used to characterize the interactions among individual lemmings as they dig new burrows. The value of L is calculated as follows:

[0080] .

[0081] When foraging, lemmings move widely and freely within their burrows, wandering randomly within their foraging areas to ingest as much food as possible. A spiral winding mechanism is used to simulate this random foraging behavior of lemmings, expressed as follows:

[0082] ;

[0083] in, This represents the position of the i-th search target at iteration (t+1) during the foraging process; The spiral shape representing random searching during foraging is expressed as:

[0084] ;

[0085] ;

[0086] The radius represents the foraging range, which is the Euclidean distance between the current position and the optimal solution.

[0087] When lemmings evade predators, they utilize their exceptional running ability to escape back to their burrows and also employ deceptive maneuvers to evade capture. The lemmings' avoidance and protective behaviors when facing danger are a key focus of modeling; their expression is:

[0088] ;

[0089] in, Let represent the position of the i-th search target at iteration (t+1) during the process of avoiding predators; G is the lemming's escape coefficient, representing their escape ability, which decreases as the number of iterations increases. Tmax represents the maximum number of iterations. yes Flight function used to simulate the deceptive maneuvers of lemmings during their escape;

[0090] ;

[0091] in, , represents the levy flight momentum strategy, u and ν are random values ​​in the interval 0-1, and β is a constant equal to 1.5. Let u be the standard deviation. This is a gamma function.

[0092] In the artificial lemming algorithm, the four search strategies are closely related to the lemming's own energy level. In the initial stage, the lemming often participates in the exploration process to locate promising areas, while in the later stages of the search, it tends to implement local exploration to obtain the global optimum.

[0093] To maintain a robust balance between exploration and development, an energy factor was designed. This energy factor gradually decreases during the iteration process. When lemmings have sufficient energy, they selectively migrate or burrow; otherwise, they forage and evade predators. The calculation formula is as follows:

[0094] ;

[0095] As the number of iterations increases, Gradually decrease it to 0, with fluctuations; set the threshold to 1 so that the lemmings have almost the same opportunities to explore and exploit during the iteration process.

[0096] when If migration or digging is not possible, then foraging or escaping is not possible. Once the maximum number of iterations is reached, the globally optimal solution is output.

[0097] Traditional metaheuristic algorithms often face severe challenges in handling such problems due to premature convergence, insufficient local search capabilities, or low global exploration efficiency, especially in dynamic environments and multimodal problems. To enhance the convergence speed and robustness of the artificial lemming algorithm, this embodiment employs a Circle chaotic mapping for the rand value in the burrowing, spiral winding, and energy regulation mechanisms, with the expression:

[0098] ;

[0099] in, For random values ​​after using the Circle chaotic mapping, is the original random value, and mod is the modulo function.

[0100] This method utilizes the Circle chaotic mapping to make the mapping values ​​more average; all settings should conform to the range of rand values ​​[0,1].

[0101] The improved DCH neural network, such as Figure 4 As shown, it includes an input layer, three sets of convolutional downsampling layers, a fully connected layer, a fully connected hash layer, a Cauchy cross-entropy loss function, and an output layer.

[0102] S2.2: Use an improved DCH neural network for hyperdimensional vector encoding;

[0103] An improved DCH neural network is used to encode the input image data into a superdimensional vector. After the image is input from the input layer, it is downsampled by three convolutions, and then a one-dimensional vector with multiple dimensions is constructed through a fully connected layer. Finally, the required superdimensional vector is output through a fully connected hash layer and the Cauchy cross-entropy loss function.

[0104] This network utilizes the Cauchy distribution to construct a more accurate similarity-preserving loss function. It employs an improved DCH neural network and uses the Cauchy cross-entropy loss function for classification testing, determining the images based on normalized Hamming distance: in the hash code space, similar images will have similar hash codes; semantically similar images have closer Hamming distances, while semantically dissimilar images have farther Hamming distances. By using normalized Hamming distance for determination, this encoding method is more suitable for hyperdimensional computation.

[0105] S2.3: Classify image information of the same category to form a category hyperdimensional vector;

[0106] The encoded hyperdimensional vector is standardized to obtain... * A superdimensional vector, called the training superdimensional vector, is denoted as . (i=1, 2… * ).Right now Each category contains a hypervector with multiple categories. There are several hyperdimensional vectors. Then, training hyperdimensional vectors belonging to the same class are combined using addition operations in hyperdimensional computation to form a set, denoted as . After standardization, it becomes A hyperdimensional vector, called the category hyperdimensional vector, denoted as . .

[0107] This category of hyperdimensional vectors, along with their labels, are then stored in a storage space called associative memory.

[0108] S3: Construct the AdaptHD retraining model: Use this model to calculate the accuracy of hyperdimensional vector classification and calculate the similarity between the hyperdimensional vectors of the training samples and the hyperdimensional vectors of the categories.

[0109] As the amount of information stored increases, errors in hyperdimensional calculations can occur due to the influence of other erroneous features. Therefore, it is necessary to add an AdaptHD retraining model; the specific process is as follows.

[0110] ;

[0111] ;

[0112] This method is based on iterative AdaptHD, which changes the average training error rate after multiple iterations. The value of , when the average error rate is higher The larger the value, the smaller the value. AdaptHD retraining is equivalent to the process of calculating classification accuracy in other algorithms. It recalculates the similarity between all training samples and the class hypervectors; it removes misclassified training samples from the incorrect categories and adds them back to the correct categories, strengthening the model's correct understanding of the samples and removing misclassifications.

[0113] S4: Train the model and test it on the dataset; output the test results; the specific operation method is as follows:

[0114] Once the model is fully built, training begins. The dataset is divided into training and test sets in a 7:3 ratio. The training set is then randomly divided into mini-batches, which are fed into the built model for training. Figure 5 As shown.

[0115] After training, the model's accuracy is tested using a test set, such as... Figure 6 As shown; during testing, the minimum batch size was set to 16.

[0116] The accuracy on the test set exceeds 85%, which is within an acceptable range, indicating that the model has been successfully trained. During testing, the number of iterations and the learning coefficient can be adjusted to find the best model.

[0117] After successful model training, the model is tested on new images: First, the new image is encoded and normalized to obtain its hyperdimensional vector; then, the similarity between this hyperdimensional vector and the category hyperdimensional vector is calculated. Similarity is an important metric for image detection. There are two different calculation methods, as follows:

[0118] Method 1: Use normalized Hamming distance to measure similarity; this method is suitable for measuring the similarity of hyperdimensional vectors encoded with binary data types; the expression for similarity measurement is:

[0119] ;

[0120] The above formula is used to calculate , How many distinct dimensions are present in two hyperdimensional vectors; n represents the number of dimensions. This indicates the calculation of the Hamming distance value;

[0121] ;

[0122] This formula is used to calculate the normalized Hamming distance. express , Hamming distance between two hyperdimensional vectors; defined in hyperdimensional computation: when the normalized Hamming distance between two hyperdimensional vectors is 0.5, the two hyperdimensional vectors are pseudo-orthogonal to each other; the closer the normalized Hamming distance is to 0, the better the similarity between the two hyperdimensional vectors; conversely, the closer the normalized Hamming distance is to 1, the better the orthogonality; the normalized Hamming distance is equal to 0 or 1 only when all corresponding elements in the two hyperdimensional vectors are the same or different.

[0123] Method 2: Calculate the similarity between two vectors using cosine distance; this method is suitable for non-binary hyperdimensional vectors; the expression for its similarity measure is:

[0124] ;

[0125] express , The cosine distance between two hyperdimensional vectors. The closer the cosine distance between hyperdimensional vectors is to 0, the higher the orthogonality between them; the closer it is to 1, the higher the similarity between them; when all corresponding elements are different, the cosine distance is -1.

[0126] After normalizing the Hamming distance, the hypervector of the new image is calculated, and the similarity between the hypervector and the category hypervector is calculated. Then, the images of different categories are judged, and the judgment result of whether the driver in the image is driving distractedly is output.

[0127] To verify the technical effects of this invention and demonstrate the advanced performance of its method, the inventors constructed the same experimental environment and selected ResNet, AlexNet, and the model of this invention to conduct experiments for comparison and verification. The experimental content is as follows:

[0128] The hardware platform configuration used in this experiment is as follows: operating system is Windows 11, equipped with an NVIDIA GeForce RTX 4080 Super with 16GB of RAM, Python version 3.11, PyTorch version 2.1.0, and CUDA version updated to 12.1. ResNet, AlexNet, and the model of this invention were used to predict images of distracted driving behavior.

[0129] Regarding experimental parameters, the maximum number of input batches was 16, the experimental iterations were 100 rounds, and the early stop mechanism was disabled. The collected images of distracted driving behavior were divided into training and test sets in a 7:3 ratio.

[0130] The experimental data is shown in Table 1, which represents the model weights that performed best on the test set during 100 iterations. TOP1 represents the most likely class predicted by the model that matches the true label, and TOP3 represents the top three classes with the highest probabilities that contain the true label. FPS represents the amount of data the model can process per second.

[0131] Table 1

[0132] Model Top1 Top 3 FPS Resnet 75% 81% 712.2 Alexnet 80% 90% 631.7 This invention model 88% 96% 1204.1

[0133] The data in the table above shows that under the TOP1 decision condition, ResNet's accuracy is 75%, and AlexNet's is 80%. The model proposed in this invention achieves an accuracy of 88%, a significant improvement. Under the TOP3 decision condition, ResNet's accuracy is 81%, and AlexNet's is 90%. The model proposed in this invention achieves an accuracy of 96%, also a significant improvement. The experimental data demonstrates that the model proposed in this invention outperforms both AlexNet and ResNet in both TOP1 and TOP3 decision conditions. Furthermore, its classification efficiency far surpasses that of traditional neural network models.

[0134] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered by the present invention.

Claims

1. A method for recognizing distracted driving behavior based on an improved DCH neural network, characterized in that: Including the following steps: S1: Construct the dataset and preprocess it; S2: Constructing a hyperdimensional computational model: S2.1: Constructing an improved DCH neural network: Replacing the ADAM optimizer in the DCH model with an improved artificial lemming optimizer; When improving the artificial lemming algorithm, the random number rand is optimized using the Circle chaotic mapping; and when lemmings migrate long distances to explore, the learning mechanism of the eagle-fish algorithm is added to the population. S2.2: Use an improved DCH neural network for hyperdimensional vector encoding; S2.3: Classify image information of the same category to form a category hyperdimensional vector; S3: Construct the AdaptHD retraining model: Use this model to calculate the accuracy of hyperdimensional vector classification and calculate the similarity between the hyperdimensional vectors of the training samples and the hyperdimensional vectors of the categories. S4: Train the model and test it on the dataset; output the test results.

2. The distracted driving behavior recognition method based on an improved DCH neural network according to claim 1, characterized in that: The optimization of the random number rand using the Circle chaotic mapping specifically involves employing a spiral winding mechanism to simulate the random foraging behavior of lemmings, the expression of which is: ； in, This represents the position of the i-th search target during the (t+1)th iteration in the foraging process. This represents the current optimal solution; F is used as a flag to change the search direction, avoiding local optima and providing the search agent with a higher chance to rigorously scan the problem domain; that is: ； in, To round down; The spiral shape representing random searching during foraging is expressed as: ； ； The radius represents the foraging range, which is the Euclidean distance between the current position and the optimal solution; The circle chaotic mapping is used to apply the rand value in the helical winding mechanism, and its expression is: ； in, For random values ​​after using the Circle chaotic mapping, The original random value is denoted by 'mod', and 'mod' is the modulo function. This method utilizes the Circle chaotic mapping to make the mapped values ​​more average. All settings should conform to the range of rand values ​​[0,1].

3. A method for recognizing distracted driving behavior based on an improved DCH neural network according to claim 1 or 2, characterized in that: During the long-distance migration and exploration of lemmings, the direction and distance of their migration are affected by ecological and environmental factors. To simulate this behavior, the following equation is proposed: ； In the above formula, where, Let represent the position of the i-th search target at iteration (t+1) during long-distance migration exploration. BM describes the characteristics of Brownian motion, utilizing dynamic and uniform step sizes to explore potential regions in the search space; R is a space of size . A vector whose elements are random numbers uniformly distributed in the interval [-1, 1]; An eaglefish algorithm learning mechanism was added within the population, and its expression is: ； in, for The optimized value, in The bias was increased; Let w be the optimal individual position element within the population, and w be the learning coefficient. This represents the position of the i-th search target at iteration t.

4. The distracted driving behavior recognition method based on an improved DCH neural network according to claim 1, characterized in that: The improved DCH neural network includes an input layer, three sets of convolutional downsampling layers, a fully connected layer, a fully connected hash layer, a Cauchy cross-entropy loss function, and an output layer. After the image is input from the input layer, it undergoes three convolutional downsampling processes. Then, a one-dimensional vector with multiple dimensions is constructed through the fully connected layer. Finally, the required hyperdimensional vector is output through the fully connected hash layer and the Cauchy cross-entropy loss function.

5. The distracted driving behavior recognition method based on an improved DCH neural network according to claim 1, characterized in that: The hyperdimensional vector encoding described in step 2.2 is performed as follows: using an improved DCH neural network, a Cauchy cross-entropy loss function is used for classification testing, and normalized Hamming distance is used to determine the classification: in the hash code space, similar images will have similar hash codes. Images with similar semantics have a short Hamming distance in their hash codes, while images with dissimilar semantics have a long Hamming distance.

6. The distracted driving behavior recognition method based on an improved DCH neural network according to claim 1, characterized in that: The similarity calculation in step S3 requires re-encoding and standardizing the images in the training samples to obtain new image hyperdimensional vectors; then, the similarity between this hyperdimensional vector and the category hyperdimensional vector is calculated; the similarity calculation method includes: When the hyperdimensional vector is encoded using a binary data type, the normalized Hamming distance is used to measure similarity. When the hyperdimensional vector is a non-binary hyperdimensional vector, the cosine distance is used to calculate the similarity between the two vectors.

Images