Abstract
The modern-day influx of vehicular traffic along with rapid expansion of roadways has made the selection of the best driver based on driving best practices an imperative, thus optimizing cost and ensuring safe arrival at the destination. A key factor in this is the analysis of driver behavior based on driver activities by monitoring adherence to the features comprising the established driving principles. In general, indiscriminate use of features to predict driver performance can increase process complexity due to inclusion of redundant features. An effective knowledge-based approach with a reduced set of features can help attune the driver behavior and improve driving patterns. Hence, a Deep Mutual Invariance Feature Classification (DMIFC) model has been proposed in this study for predicting driver performance to recommend the best driver. To achieve this, first, the driver behavior is broken down into various features corresponding to a simulated driving dataset and subjected to preprocessing to reduce the noise and form a redundant dataset. Thereafter, a Mutual Invariance Scale Feature Selection (MISFS) filter is used to select the relational features by calculating the spectral variance weight between mutual features. The observed mutual features are promoted to create a dominant pattern to estimate the Max feature-pattern generation using Driver Activity Intense Rate (DAIR). The features are then selected for classification based on the DAIR weightage. Additionally, the Interclass-ReLU (Rectified Linear Unit) is used to generate activation functions to produce logical neurons. The logical neurons are further optimized with Multiperceptron Radial Basis Function Networks (MP-RBFNs) to enable better classification of driver features for best prediction results. The proposed system was found to improve the driver pattern prediction accuracy and enable optimal recommendations of driving principles to the driver.
1. Introduction
Enhancing the driving ethics effective analysis of the driver comportment is very essential. The effective analysis of driver behavior, his conduct, psychology, talent, and attitude, is the key to improving adherence to driving principles for both drivers and passengers for ultimately improving the management of resources and service quality. Cost, safety, and compliance are all parameters that can be positively impacted by investigating and monitoring driver behavior and performance based on driving related features utilized by the driver. Such scrutiny of driver’s behavior relying upon analysis of several features may be undertaken by a machine learning approach. The various attributes of driving usually vary from driver to driver and by accessing such variances certain driving patterns can be identified and promoted. The real-time identification of risky behavior or policy violations based on a driver’s current driving pattern is near impossible due to the variability of the activities at the time of handling the vehicle. Utilizing in-vehicle driving support for analyzing driver features with respect to performance best practices by using deep learning methods can yield best results in pattern prediction. In this study of driver behavioral analysis, the data regarding the selected features was collected during the performance of driving simulations handled by the drivers that simulated real-life situations. The features that were considered for observation were related to various driving activities like harsh braking, fast driving, slow driving, and driver speed while turning the vehicle, as well as other driver skills to gauge driver performance. The features that were observed to be the most frequent were part of a pattern, which were then attributed to the drivers to predict the best performance recommendations.
Consequently, the best adherence to the recommended pattern of features while driving was one of the most important qualities of a good driver. Figure 1 illustrates the process of analyzing driver activities and establishing access patterns. The data analysis model did not include the nonrelated features accessed for the big data analysis for the prediction; only those features based on the performance, knowledge, skills, and the perceptive abilities of the driver were analyzed. Therefore, it was a feature-based analysis focusing on the performance of the actual drivers selected for the operation. Based on the driver performances and functions in the vehicle system, the best features were recognized and implemented using the Filter Case Feature Selection Methods (FCFSM). The conventional vehicle data provided were also analyzed by various Driver Activity Intense Rates (DAIR), to form patterns of behavior analysis, mainly from the driver and the vehicle data to form safety invariance features. The DAIR was utilized to carry out intense weight estimations accessed by the mutual features. Most of the mutual features that had relational patterns demonstrated by the drivers were used to choose the right patterns associated with the best driving. Thus, the best patterns were based on the real-time driving activities observed for all the identified features.
The best feature patterns were identified based on various factors of driver activities that contributed to the best performance using the deep neural network (DNN). It is a pattern of ML procedures identical to AI neural linkages which targets imitating the brain info processing. For example, to improve safety, drivers needed to be trained to become familiar with the more frequent features associated with safety to develop the resource. Figure 2 shows the driver behavioral relational feature analysis. This process has become a common way to complement the basic training especially for those drivers whose patterns are not aligned well with the recommended driving principles. It must be underlined, however, that alignment with the driving principles does not provide sufficient reason to believe that driver will always choose the best pattern as identified by the classification algorithm. Classification features derived without optimizing the redundant features present a big problem in establishing behavioral analysis. In this case, support for the individual driver needs has been personalized for a specific driver behavior performance as predicted by the pattern classified by the Recurrent Neural Network (RNN) through deep learning methods. Additionally, the feature selection and classification of the process based on the driver activities were undertaken using Rectified Linear Unit (ReLU) which was performed on the activation unit. The classification was performed on the trained features with weightage factors from different tasks that were considered such as steering and speed control, shifting their function to fix the marginal weightage to the classifier. The marginal features were evaluated to influence the rate of the pattern average which was trained with the neuron class weightage. The deep neural network was utilized in the classification of features based on the class by reference using the Rectified Linear Units (ReLUs) to determine the activation function for optimizing the training features with the logical neurons which were ruled out to provide the optimal conditions. The driver behavior analysis of feature and classification was further improved by using Multiperceptron-RBFNN for driver behavioral analysis to predict performance. A lot of research had been done in this field, which is to be discussed in the next section of the paper, but our proposed system of Deep Mutual Invariance Feature Classification (DMIFC) model is so efficient for predicting driver performance to recommend the best driver. To accomplish this, first, the driver behavior is shattered down into various features corresponding to a simulated driving dataset and subjected to preprocessing to reduce the noise and form a redundant dataset. Thus, bridging the gap between the past trends, this paper is exploring driver behavior and performance by means of deep learning techniques for envisaging the patterns of the driver activities which holds a lot of possibilities; moreover, performance records have been composed to benefit feature selection and the classification of driver conduct and advance exploration has been undertaken to predict the best driver. The paper has been organized as follows: part 1 is the introduction of the topic followed by part 2 which discusses the related work, part 3 discusses the Deep Mutual Invariance Feature Classification (DMIFC) Model, part 4 gives details about results and discussions, and part 5 gives the details of conclusion obtained.
2. Related Work
Enforcing driving behavior is compliance package that ensures that people are safe and aligned with driving regulations. This refers to the safety and compliance measures, for instance, respectable speed limits, predicted driver control, and traffic rules with respect to unsigned intersections. Safety also has attracted more and more attention [1]. In recent times, machine learning models have not been capable of handling feature analysis in driver behavior pattern prediction well. Since the active control of the vehicle handling is high dimensional, for example, with the lane departure warnings and advanced driver assistance systems, such features usually create high dimensions in complex data analysis models [2]. Additionally, with increased use of road risk predictions and driver behavior-driven safety, a mixed approach is required for conventional vehicles that require autonomous vehicle as well as manual operation. The driver behavior quantity model with identification and classification of driver behavior extracts a quantitative variable from the quantitative model to designate the priorities better [3, 4]. Similarly, classifications using driver information which collected the statistical properties of the accelerometer signal are often used as inputs to the classifier [5]. The driver specific information, like accuracy of active and normal driving style classifications and so forth, is usually derived using the collected data [6]. Function checks, using the exact classification of the state of specific drivers as well as the eye movement of the driver, are used to derive states of downiness and are a prerequisite for preventing traffic accidents due to driver sleepiness [7]. Based on preceding approaches of classification, it was observed that, regardless of the behavioral constructions mined from a single eye, like the shape and texture of the driver’s viewpoint, the most utilized stationary threshold for the positioning error and visual impairment of the eye, classification is susceptible to the influence of the value [8]. Using the Laplace support vector machines and semisupervision of the eye and head, the driver condition is classified as a prudent and cognitive distraction [9]. Driver distraction detection has been evaluated using machine learning. It has been observed that driver distraction firmly influences risk of crash. Such a prediction, reflecting that distraction can be dangerous and a cause of concern towards traffic safety, will remind the driver of ensuring a safer lane-keeping aid system [10]. Current methods have usually tried to detect any emergencies, from environmental sensors to approximation of the feature analysis related to the driver’s behavior only [11]. A multivariate statistical model that has used driving behavior as a behavioral indicator among younger drivers has studied the relationship between self-reported and impartial events surrounding driving problems to be nonobserved as far as classification methods are concerned [12]. Driving speed and fuel have been treated as important feature patterns of driving style using support vector machine (SVM) to reflect the nature of driving styles. Identifying the behavior of the driver style, to reduce the number of support vector machines and using clustering techniques to reduce the variability in the recognition of driving styles have both been used as two distinct types of data-driven approaches [13]. Driver behavior on the road can also be considered as the dynamic driving style. Sometimes drivers may be calm but may become aggressive under certain influences [14]. Information about the dynamic driving style of the driver has been considered important. The Double Articulation Analyzer Machine Learning (DAA-ML), a temporal prediction method, uses maximum observed feature selection to make predictions, but this can lead to the creation of nonrelevant patterns when making recommendations. This method determines driving situations from the driver activities to create features that are used with the time series data of the incoming driving behavior, to predict the outcome that may occur in the driving behavior [15]. The driver behavior pattern has also been considered with the help of machine learning algorithms. It requires a new digital image x as input and can be expressed as a function for generating an encoded output vector in the same manner as the target vector y (x) [16]. The development of such systems utilizing machine learning to reduce the consumption of transport energy, for instance, can also be applied to behavioral evaluation techniques. Detecting and analyzing driver behavior using deep learning performance systems has unique benefits [17]. In this paper, the humanoid driver has divergent styles of driving style skills and using the ordering and feature variety of their own driving conduct to drive the features to be used in machine learning technology [18]. The application of medical signals and images to these methods can enable the authorities involved to support the right decisions in cases of arbitration. Application of Internet of Medical Things (IoMT) to medical data can identify and help contain instances of aggressive driving behavior to reduce traffic accidents and crashes [19, 20]. Deep learning methodologies have been utilized for extraction of such latent feature from the data collected. Aggregation of extraction of autonomous driver feature from the duration of the vehicle operation has been used to create a velocity distribution to enable better understanding of the drive by a very human process [21, 22]. It has been observed that, based on the recognition of the hidden features extraction from the driver, the recognition model, using complex human feature extractions, has found it difficult to report good performance [23, 24]. Classifications around the state of the skill/drive, though presenting a basic problem in establishing the supporting viewpoint of the driver, can nevertheless be applied to create custom recommendations personalized for a particular driver. Classifications around driver sleepiness have also been applied using deep learning [25]. In such instances, machine learning has been used on data around performance and classifications and for the analysis of driving performance. Similarly, light footprint driving behavior classification with the Random Forest classifier with well-trained extracted features has been shown to help select the most relevant subset of the function, thereby obtaining a much higher accuracy rate, as can be seen from the independent test data [26]. Experimental extraction of blinking and head movement of the driver in a driving simulator has also been considered. Large-scale datasets have also been utilized to develop a feature selection methodology with K-nearest neighbor algorithms based on classification of the driver state and evaluation. In this classification system of driving behavior, feature extraction has, as in many other cases, played an important role [27, 28]. Various feature extraction methods have therefore been used to drive the behavior classification, using various datasets, enabling direct comparison of feature extraction methods for behavior analysis [29, 30]. This review not only provides an in-depth understanding of the scope of the problem of the state of the shortage but also underlines the importance of the input of the performance index used in deep learning algorithms, which can offer the most advanced systems [31]. Solving the issues around driver behavior analysis using machine learning algorithms seems to be the most relevant for feature selection and extraction as well as the classifications used for predicting the driver behavior performance [32]. This paper meticulously inspects the dataset and implements data training using variables to facilitate prediction in vehicular driving [3]. Thus, bridging the gap between the past trends, the proposed model in this paper analyzes driver behavior and performance using deep learning methods for predicting the patterns of the driver behavior that holds a lot of promise and has been utilized in this study where performance data has been collected to aid feature selection and then the classification of driver behavior and further analysis have been undertaken to predict the best driver.
3. Deep Mutual Invariance Feature Classification (DMIFC) Model
In this proposed system, a deep neural network based feature selection and classification has been implemented to predict the best driver pattern. All the driving simulations have been collected into the driver feature data log using the best data forums for information mining and analysis. The Deep Mutual Invariance Feature Classification method used only considers the mutual relationship between the most integrated features from the driver pattern to be included in the feature selection process under the filtering process. Based on the variance dependencies, the mutual relationships are analyzed to estimate the intention rate of Max feature selection. The following steps have been followed to make better driver behavioral pattern prediction:(i)Analyze the feature dependencies based on the mutual relationships between the formal features among the invariance scaling factors.(ii)Estimate the Driver Activity Intense Rate from the observed feature depending on Max feature attainment with low margin driving factor utilization.(iii)Create a logical rule based on Interclass ReLU activation function to train the neuron for feature classification using Multiperceptron RBFNN.(iv)Predict the Max feature pattern of driver dependencies to suggest the best feature patterns for recommendations of drivers and driving principles.
The proposed system contains the following steps during the implementation of the driver behavior analysis and the architecture is shown in Figure 3 explaining the working principles of the driver pattern prediction and recommendation.
The feature selection comes under the process of driver utilization during which driving weights are assigned depending on the time factor, for instance, by which time, the behavior of the driver changes. Multiple features accessed at the best level are part of the driver pattern.
3.1. Data Initialization
Firstly, the dataset pertaining to collective drivers is preprocessed with plenty of elements which are supposed to possess attributes. Such inducements impacts appraise redundant data with uncovered data filled or nonfilled in raw dataset and verify the values occurrence and ingenuity. Recent real-world data has been collected from driving transports logs to be processed with a denoising filter algorithm for this work.
|
Processing stage has been explained in the above algorithm for reduction of dimensions established upon an attribute by values filtering. Entire data carries multiple attributes as features implied as the single record of info. The raw data conversion is initialized for cleansing to remove noise and by filling attribute values to the variables by preprocessing. For instance, frequent data entries found with identical numerical values are removed. Unfilled dataset values filled with constant values are included to be processed for prediction.
3.2. Mutual Invariance Scale Feature Selection (MISFS)
In this stage, the preprocessed R-Fds dataset is taken for further processing for the MISFS features selection. This step finds the mutual relationships by dividing the scaling weightage based on observed features. It is a quantity to measure between two arbitrary parameters by which we can compute the data of one weight through the other. Thus, using MISFS step, we can obtain mutual relations between various parameters that are used as weights for the proposed system. The variance relationships are generated among the feature subcategory by surfing the related weights. The feature weights are scanned by the surfing feature for selection of the improving values between the class end margins. On every choice, driver feature accessed weights are grouped into subcategories depending on inner and outer set subgroups. By this, redundant deterministic values are enabled to make small dimension data points.
|
The above algorithm defines the mutual relationship based on the driver accessed level, whether the feature weights are marginalized to select the continual access of weights during feature accesses by connecting the weights as patterns depending on the Max frequent accessed features.
3.3. Max Feature Pattern Generation
In this stage, the features are observed based on the maximum driver feature utilized depending on frequent pattern mining relationships between the features. Frequent mining generates the graph points accessed by feature weight correlation. Each feature creates a domain space that holds the max weightage depending on driver utilization; similarly, all the domains are a class for holding relevant features. The features are then marginalized by access patterns to create access patterns. This is estimated to sum the redundant pattern from the mutual relation score identified with the Max terms feature observed from the driver behavior.
|
The above algorithm finds the maximum frequent access features from the driver’s usage pattern. The frequent graph mining estimates the correlation between the mutual relations of accessed driver features. Further best patterns are categorized based on the measurement of Driver Activity Intense Rate.
3.4. Driver Activity Intense Rate
In this stage, DAIR estimates the even similarity among the driver pattern’s most accessing rate which is the marginalized CRs pattern class. This finds the most appropriate pattern from a frequent class in which the generated patterns are categorized into absolute mean weight; the mean rate implies to the attributes available in every event class. In accordance with measuring event support, a specific event is shortlisted and presented as result.
|
The pattern prediction process is explained above in the pseudocode, and the methodology calculates the event support for several event classes produced by the best driver behavioral pattern. Ultimately, a single event is shortlisted as a possible event for result production.
3.5. Interclass ReLU (Rectified Linear Unit) Activation Function
In this stage, logical neurons are created through Interclass ReLU adaptive patterns and are trained with Boolean representation. This creates an activation function to train the logical neurons by ruling out access weight. The weights are marginalized into the hidden activated neurons and create links depending on the mutual pattern relation score. All over the neurons, the link with logical combination feature access depends on the weight accessed by the feature limit.
The feature margin weights are input to the Rectified Linear Unit as input is D denotes driver features.
For driver and the vehicle layer from for ReLU input , the Rectified Linear Unit function is ReLU = max (0, Zn). Consider i as the feature weights and j as pattern features; driver feature = ReLU − y initialized i = ReLU − y at K features at pattern matched at maximum weight. The linear units have two derivates depending on the order of the rule prediction principle.
If we use behavior of driver as measure, then 1 will become something else, but 0 will stay the same, and so for a driver ,
The behaviors ReLU with the driver instance rate = 0 on the left side.
Driver function attends that generates i-th features from output function in n-th layer, D(n) = max(0,n) D(k) = max(0,n) is the ReLU unit features, and is linear input into (i + 1) (i + 1) intent the marginal value of the driver behavior analysis the i-th features and (k + 1). The first-order activation function is as follows:
The first order initializes the driver features taken and logical points fixed with conditional operation to check the normalized weightage. Similarly, the second order follows the second-order derivative to fix the scaling conditions to check logical facts of pattern weights to fix the rules adaptive to the neurons.
If the input is a positive value, then that value is returned otherwise 0. This activation function sparse integration to create a logical function supports the RBFNN. The steps given below show the activation function.
|
The back-propagation algorithm is a perspective to train the neurons into the convolution network-based deep neural network. Based on nonnegative driver patterns, the weights are updated into the ReLU function to optimize the driver behavior analysis to make the logical neurons. Further logical functions activate the logical conditions to verify the class by reference to feed the values to the nest layer which is constructed through RBFNN.
3.6. Multiperceptron Radial Basis Function Networks (MP-RBFNs)
Multiperceptron network intent with RBFNN in the hidden layer makes a communal link on radial basis function. In this neural network, the activation function creates the logical neurons to fix the rules assuming that, with the distance from each position, it will form an inverted feature selection through each position. Radial basis function network is a mathematical method used in neural networks. It uses instigation tasks as radial basis function to enhance the ordering accuracy to analyze the best pattern class by reference. The following principles are used to make the input features for classification by construction of MP-RBFNs to compute the feature distance-vector space. Let us construct the network as m-dimensional input vector and n-dimensional output vector . Let the features f: x ⟶ y. pattern margins construct over RBFNN using ReLU activation function.where x = , is the center, and is the radius width of the hidden node.
The output is represented as
The features are defined by the targeted weightage in the centroid radii point based cluster to optimize the feature selection, to reduce the training nature of nodes, and to avoid overtraining when the radius has been reduced indefinitely. The position of the center of the hidden layer is adjusted with the activated neuron as variable radial width can reduce the region overlap neurons. The hidden links verify the feature distraction to transform the conditions using the inverse variable at radial function.
W = and T = Let us consider W as pattern feature weightage into the H hidden layer with link transmission T. Similarly, the process links P which activates the Logical neuron function to construct hidden layers = .
For K-number of hidden nodes,
T = –target output
The sample belongs to the class. The sampling features are marginalized with targeted conditions in the hidden layers.
The neurons training state gets the normalization progress to intraclass logistic activation construction level to achieve optimal weight rate for all individualized features achieved in the feed-forward layer. The following steps shows the MP-RBFNN.
|
This algorithm is postlearning on a feed-forward network to classify the patterns. The optimized RBFNN activates functions based on the ReLU function and categorizes the weightage values tested with driver behavioral input, the neurons are then trained with DIAR formalized with hidden weights, and the optimized output is formalized for the best driver pattern from features to produce recommended and nonrecommended class.
4. Result and Discussion
A Deep Mutual Invariance Feature Classification (DMIFC) model is employed in this paper where it has been projected for predicting driver performance to commend the finest driver. To achieve this, first, the driver conduct is broken down into various sorts corresponding to a simulated driving dataset and exposed to preprocessing to lessen the noise and form a redundant dataset. The driver interactive pattern expectation result is evaluated with confusion matrix and paralleled with the different levels of algorithms and it has obtained the optimum performance as compared to the existing methods. The behavioral analysis of driver activities is collected as log intents and is carried out by collection from UCI web link repository, and its implementation is carried by the Python running framework. The proposed driver behavior data analysis is implemented using feature selection and spectral classification methodologies for deduction of a set of related categories for every user’s queries depending on the retrieval history of driver behavior search. The proposed method has produced efficient results on context clustering and hybrid data pattern (HDP) and is seen to improve the performance better than the previous method. The parameters are tabulated below.
The parameters listed in Table 1 make up the environment to test the results with the collective dataset. The performance of DMIFC is evaluated through clustering accuracy (cs), precision rate, recall rate, and time complexity.
The driver behavioral pattern prediction result is evaluated with confusion matrix and compared with the different levels of algorithms listed in Figure 4. The proposed system proves best pattern prediction accuracy compared to the other dissimilar methods like Genetic Algorithm-Fuzzy C-Means (GA-FCM).
Table 2 shows that prediction accuracy produced 10 drivers as 96.8%, 20 drivers as 97.9%, and 30 drivers as 98.4%, demonstrating that the planned method has fashioned sophisticated clustering accuracy compared to the principal component analysis (PCA) neural network.
(1) Analysis of Precision Rate. Precision (Pr) is distinct as the quantity of the total numbers of pattern predictions and total numbers of driver behaviors relation, where R is the relevant pattern calculated from the confusion matrix.
The precision rate is seen to produce the performance provenance of the proposed system as well compared to other methods. Figure 5 shows the performance chart produced by the proposed system with comparisons.
Table 3 defines the performance of precision produced from a dissimilar level of 10 drivers as 89.3%, 20 drivers as 85.4%, and 30 drivers as 87.1%, a higher performance ratio that proves the prediction accuracy.
(2) Analysis of Recall. Recall (Rc) is distinct as the proportion of entire amount of retrieved driver behavior and the relevant positive values related to pattern prediction.
The proposed recall rate produced the best performance as well compared to other dissimilar methods shown in Figure 6. The result proves the better prediction rate of the proposed implementation with differential driver user variances.
The best recall rate proves the pattern prediction performance of proposed system as shown in Table 4. The recall rate defines the proposed specificity dependences of true values to produce the best result.
The analysis of time complexity is as follows:
The time taken to complete the feature pattern selection using O(n) execution with different methods is shown in Figure 7. The proposed system is shown to have lower execution time in comparison with the earlier systems and performance too.
Table 5 highlights the time complexity comparison, where the offered perfect clustering is seen to produce 10 drivers as 5.2 (ms), 20 drivers as 6.6 (ms), and 30 drivers as 7.1 (ms), showing that the approach which has been proposed yielded lesser time complexity.
5. Conclusion
To conclude, the driver behavioral prediction based on Deep Mutual Invariance Feature Classification (DMIFC) model with a low-dimensional perspective produced the best results. The best patterns were predicted with the support of preliminary feature selection to make redundant data and the estimated DIR supported the maximum weightage based on pattern utilized by the driver. The adaptive deep neural network based on RBFNN with ReLU activation function predicted the best pattern from the driver behavioral activates. The results prove the high prediction and classification performance up to 98.7% compared to previous systems and produced the best recommendations for driver driving principles. The proposed system proves best pattern prediction accuracy compared to the other dissimilar methods like Genetic Algorithm-Fuzzy C-Means (GA-FCM). In the future, we can use deep neural networks to check the healthy state of the driver while driving a car to attain the best features that can serve to adapt the driver’s comportment and mend driving patterns to the best possibility.
Data Availability
The data shall be made available upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.