Ensuring Driving and Road Safety of Autonomous Vehicles Using a Control Optimiser Interaction Framework Through Smart “Thing” Information Sensing and Actuation

Abstract

Road safety through point-to-point interaction autonomous vehicles (AVs) assimilate different communication technologies for reliable and persistent information sharing. Vehicle interaction resilience and consistency require novel sharing knowledge for retaining driving and pedestrian safety. This article proposes a control optimiser interaction framework (COIF) for organising information transmission between the AV and interacting “Thing”. The framework relies on the neuro-batch learning algorithm to improve the consistency measure’s adaptability with the interacting “Things”. In the information-sharing process, the maximum extraction and utilisation are computed to track the AV with precise environmental knowledge. The interactions are batched with the type of traffic information obtained, such as population, accidents, objects, hindrances, etc. Throughout travel, the vehicle’s learning rate and the surrounding environment’s familiarity with it are classified. The learning neurons are connected to the information actuated and sensed by the AV to identify any unsafe vehicle activity in unknown or unidentified scenarios. Based on the risk and driving parameters, the safe and unsafe activity of the vehicles is categorised with a precise learning rate. Therefore, minor changes in vehicular decisions are monitored, and driving control is optimised accordingly to retain 7.93% of navigation assistance through a 9.76% high learning rate for different intervals.

1. Introduction

Autonomous vehicle control involves safe driving and navigation capability by incorporating advanced sensors, communication systems, and control mechanisms [1]. The vehicles are informed continuously in real-time by the environment, such as road conditions, traffic flow, object presence, speed, direction, and braking action, to decide on the appropriate control actions [2]. The primary components of autonomous vehicle control are stability, collision avoidance, and following traffic flow rules. Environmental sensing, communication with infrastructure, and object detection systems enable safe navigation in urban and highway environments [3,4]. The navigation systems within autonomous vehicles make necessary adaptations to changes through real-time route recalculations based on new data received [5]. The control systems should efficiently manage different driving scenarios, from congested traffic to unforeseeable conditions on the road. The systems must ensure responses are consistent and dependable to realise the safe operation of the autonomous vehicle under variant conditions [6].

The control of safe driving in an autonomous vehicle faces various problems and challenges in an unpredictable or complex environment. Sensors and controls are associated with the vehicle to interpret and respond appropriately to real-world data such as pedestrians, roadway obstacles, or rapid traffic flow changes [7,8]. Sensor data, weather conditions, and road scenarios include uncertainty that may cause control failures and make decisions hard for autonomous systems [9]. Smooth coordination is supposed to be present when integrating various data sources such as GPS, cameras, and lidar. The mismatch in interpretation from all data sources may create navigation errors [10]. Moreover, ensuring the reliability and safety of the decision-making process of a vehicle against incomplete or inaccurate data is another challenge [11]. System failures, hacking risks, and technical malfunctions presuppose other obstacles to keeping control in critical situations [12].

The solution of intelligent learning algorithms in enhancing autonomous vehicle navigation comes up with adaptive and dynamic solutions while interpreting complex driving environments [13]. Machine learning algorithms analyse vast data from multiple sensors to understand patterns for correct decision-making in runtime. These enable vehicles to learn from experience to continuously improve navigation and control performance using reinforcement learning methods [14,15]. The understanding of complex driving scenarios—for instance, pedestrian crossings or adverse weather conditions—becomes more profound with autonomous systems supported by neural networks and deep learning models [16,17]. The learning systems further enhance the detection and classification of objects in the vehicle’s environment for safer and more accurate navigation. The decision-making process will be refined by intelligent algorithms, wherein the vehicles can learn from their past driving experience and refine their strategy with time. Artificial intelligence applied at each stage can enable an autonomous vehicle to adapt to different driving conditions with very limited human intervention [15,18,19].

The COIF was created to make AVs safer on the road by simplifying their point-to-point connection with real-world objects and other “Things”. Unlike older AV frameworks that use preset environmental parameters, COIF uses a neuro-batch learning algorithm to adjust to real-time traffic and road conditions. This technology helps the AV make better decisions and controls by classifying vehicle actions as safe or risky. It achieves this despite constant data from population density, possible obstacles, and traffic incidents.

COIF assures that AVs can interact with fixed and moving objects in complicated contexts via adaptive learning and better knowledge-sharing mechanisms. This technology increases safety, driving control, and the AV’s capacity to anticipate and avoid problems by more precisely monitoring and reacting to environmental signals. The project uses information-sharing technology and adaptive learning algorithms to assure real-world autonomous driving safety and consistency.

The key highlights of the article are:

• A brief discussion of the previous works related to AV safe driving and navigation support through different learning and optimisation methods
• The introduction, discussion, and theoretical explanations of the proposed framework with illustrations and mathematical models
• The discussion includes data-based analysis, functional representations, and partial outputs.
• The metric-centric discussions through comparisons using related metrics and congruent methods with percentage improvement

The rest of the paper is followed by Section 2, which discusses the latest literature review of the proposed idea. Section 3 describes the proposed control optimiser interaction framework in detail. The results and discussion are given in Section 4 and Section 5, respectively, while the paper’s conclusion is drawn in Section 6.

2. Related Works

Kang et al. [20] developed a data-driven, control-policy-based system for automated driving safety analysis. It extracts the control policies based on historical driving data and simulates them to detect hazard scenarios. It achieves 89% coverage of hazard scenarios validated with Baidu Apollo driving data. The method exhibits great competence in the detection of hazards and safety analysis. Validations on extensive data sets prove the robustness of the system. Significant advances have been made in the system to enhance vehicle safety. Han et al. [21] introduced a framework of multi-agent reinforcement learning for connected autonomous vehicles. The framework enhances decision-making and processes by sharing information among agents and uses a truncated Q-function to protect itself. Simulations within CARLA environments demonstrate its ability to maintain appropriate distance apart and prevent unsafe actions. The approach enhances traffic safety and efficiency profile for autonomous vehicle networks.

Wang et al. [22] developed a methodology for lane change decision-making in autonomous vehicles. The fuzzy inference was implemented through the decision-making process as factors such as vehicle velocity, acceleration, and delay are all considered to control safe and comfortable lane changes. Simulation across several scenarios indicated the method’s effectiveness in increasing safety and comfort when changing lanes. The technique translates to a safe and comfortable lane change. Sun et al. [23] proposed an adaptive robust formation control of connected autonomous vehicle swarms. The method presents collision avoidance and time-varying uncertainties through a constraint-following scheme with diffeomorphism transformation. The method presents a compact formation of vehicle swarms with an effective management approach to the risk associated with the collision. Simulation runs showed its validity, which presented reliable performance in complex environments.

Li et al. [24] introduced a global sorting-local gaming framework for managing complex multivehicle interactions in the scope of autonomous driving scenarios. Through game theory, the global sorting-local gaming method has effectively quantified interaction disturbance and resolved multiplayer interaction. Simulations and human-in-the-loop experiments have indicated improvements in safety and traffic efficiency compared to the traditional models. Overall, traffic management can be improved because it deals with multivehicle scenarios. Jond et al. [25] proposed a differential game-theory-based optimal control method for autonomous vehicle convoys. The method transforms the dynamics of individual vehicles to a relative dynamics-based optimal control problem, which ensures an openness Nash equilibrium for distributed convoy control. Simulations illustrate the successful application of the method in stability and safety during highway driving. The method supports stable and safe operations in highway environments. Stability is enhanced in a convoy using this method.

Wang et al. [26] presented a hybrid approach combining reinforcement learning with model predictive control to apply to cars following manoeuvres of autonomous vehicles. An algorithm is used to dynamically adjust the parameters from model predictive control according to the calculated levels of risk, which applies reinforcement learning for the classification of risks in traffic scenes. It further points to improved safety and performance compared to pure model predictive control. It further provides adaptability through the real-time tuning of parameters. Nie et al. [27] proposed an autonomous highway driving strategy with a time-to-collision and reinforcement learning-based safety check system. Unsafe actions are only substituted by safer ones if the TTC is less than some certain value; therefore, safety is increased. Simulations conducted in ramp merging tasks demonstrate elevated arrival rates but decreased collisions; the approach has good prospects for dense traffic conditions. The approach enhances collision avoidance by its feasibility in making immediate action adjustments, and validation confirms improvements in safety and efficiency.

Deng et al. [28] proposed a deep reinforcement learning approach to decision-making in uncertain highway driving scenarios. The method densifies training data to focus the process on only the safety-critical events, accelerating the evaluation speed by as much as 105 times. It enhances the building of autonomous vehicles through enhanced efficiency in safety testing. The extensive testing and strategy demonstrate the method’s fast evaluation capabilities and greatly boost safety validation processes. Feng et al. [29] offered a dense reinforcement learning technique for the safety validation of self-driving vehicles. It adapts Markov decision processes to emphasise safety-critical events more, accelerating evaluation. Considering crucial safety events enhances the efficiency of testing as a whole. The results of validation validate its efficiency in accelerating evaluation tests. It is one of the efficient accelerative techniques for safety validation.

Lee et al. [30] offered a deep hybrid learning network for autonomous driving control in adverse weather conditions. The method utilised a variational autoencoder to learn environmental features and an inception-bidirectional, long, short-term memory network to predict the driving based on these features. Highly impressive performance improvements were reported, with a mean square error of 0.0464 and an inference time of 6 milliseconds. It is an enhancement technique of control while driving, which gains response when dealing with adverse weather conditions. He et al. [31] proposed a decision-making approach for safe autonomous driving with observation robust reinforcement learning. The adversarial agent simulates attacks to the observations and utilises a robust actor-critic model to optimise policies in bounded variations. Lane-change tasks showing resilience against adversarial attacks ensure safety and good performance in uncertain conditions. The approach ensures decision-making safety in complex driving environments by addressing perception uncertainties, and the method improves safety in complex driving environments.

Ben Elallid et al. [32] proposed a reinforcement learning-based control method for an autonomous vehicle under weather and daylight conditions. Methodology driver-specific behaviour is incorporated into a model predictive control scheme using a proportional integral differential feedback channel. Hardware-in-the-loop experiments have verified its suitability for personalised control and collision prevention. It addresses varied driving conditions and styles. Enhanced tracking control and safety were offered, contributing to better driver adaptation and comfort. Shi et al. [33] proposed an integrated deep reinforcement learning framework for intelligent autonomous driving. The approach integrates macrolevel routing with real-world, microlevel behaviours, lane changes, and car-following-the approach outperforms human drivers. Its performance significantly improves traffic efficiency and safety performance, especially in networks with high penetration rates of autonomous vehicles. The approach improves traffic efficiency and safety and supports advanced autonomous driving systems.

Li et al. [34] introduced a method for controlling an autonomous vehicle with an orientation oriented towards personalised driving behaviour. The method utilises Q-ABSAS optimisation within a framework of model predictive control, thus enabling the method to add driver-dependent characteristics. The method is flexible with various driving behaviours, so an improvement in tracking control, safety, and comfort occurs. The approach improves the driving experience through individual differences in adaptation and provides customised solutions in different driving conditions. Mihaly et al. [35] developed a supervised reinforcement learning scheme for applying trajectory-tracking control in an autonomous vehicle. The technology utilises reinforcement learning to improve stability and the precision of the applied control action on trajectory tracking. Simulation outputs depict a high degree of stability in the retention of vehicle stability throughout the trajectories. Different simulations have proven the performance of the approach to stable trajectory handling. The proposed method promotes the stabilisation of the handling trajectory.

Gao et al. [36] proposed an online safety verification approach for autonomous driving based on dynamic reachability analysis. It predicts the possible future moves of neighbouring vehicles and incorporates traffic rules to analyse trajectory overlaps with legal zones. Experimental results demonstrate the framework’s effectiveness in determining and mitigating safety threats. The approach ensures collision-free driving conditions, predicts the future movements of vehicles, and provides a sound solution to safety verification in autonomous vehicles. Wang et al. [37] designed a control learning design that learns to ensure the safety of traffic with both self-driving and human-driven vehicles. The technique blended a traditional method and a learning model to predict how human-driven vehicles act in mixed groups. The combined model reduced the root mean square error to 35.64% in speed prediction from the HV compared to when the first principles were used alone. GP-MPC improves safety by better dealing with uncertainties and performing well in challenging scenarios like emergency braking. A mathematical safety buffer for cars that collide is developed, this temporal separation is linearised, and a combination mixed-integer linear program is constructed [38]. Every possible scenario, including regular, emergency, and recovery scenarios, is detailed and tested via simulations. The tradeoff between fast response times (larger safety buffer) and low reaction times (more false positives) is shown via sensitivity analysis on different reaction times. The study’s main goal is to determine how much space should be left between cars to prevent a crash in the event of a vehicle failure.

In contrast to static or rule-based solutions, the suggested COIF can adapt in real time to changing settings because of the use of a neuro-batch learning algorithm. COIF’s exclusive emphasis on point-to-point interaction guarantees that the AV extracts data directly from its immediate environment, improving safety by providing accurate, situationally driven reactions. This contrasts with other approaches, such as ORRL, HDSE-DRL, and CRL-MPC, that frequently function in sterile settings and are not flexible enough for use in the actual world. As a result of its superior fit for complicated and unpredictable environments, COIF’s method guarantees constant AV performance and enhanced safety.

3. Proposed Control Optimiser Interaction Framework

3.1. Data Collection and Description

The data used in this article is a consolidated inheritance from [39,40]. The first source provides a “self-driving cars” dataset comprising images (of lanes) and acceleration parameters (brake, steering, speed, and throttle). This dataset observes 8036 entries for the above parameters under straight and left lane changing settings. The second dataset provides independent accelerometer, GPS, timestamp, distance covered, and magnetometer readings from in-vehicle and in-road sensors. Using the images from the first dataset, the navigation scenario is illustrated in Figure 1.

With an emphasis on improving navigational and driving safety via sophisticated environmental learning, the proposed study presents a new COIF for autonomous cars. In contrast to conventional methods, this architecture allows for consistent and uninterrupted communication between the AV and its environment’s “Things” using a neuro-batch learning algorithm. To maximise driving choices, this technique guarantees accurate, real-time information transmission. A highly adaptable and safety-oriented AV system is made possible by the framework’s batch optimisation methodology for risk assessment and a unique risk indicator to change navigation in response to risky situations quickly. There has been a giant leap forward in autonomous vehicle technology with the integration of real-time learning, adaptive response systems, and risk management. The first image marks the sensors and visual input in the AV driving scenario. The information acquired for processing towards navigation and safe driving is illustrated in Figure 1. The second image showcases the change of AV direction towards the lefthand side with the possibilities of $\theta$ (angle). The third image shows the “after lane change” position of the AV for which the $\theta$ variation $[180° to\left(45°,60°,90°\right) to 180°]$ is illustrated. Based on this, $1695\times 4$ entries, accelerometer, speed, GPS, etc. sensors are observed. The maximum distance accounted for is 1000 m, and the testing time is between 0 and 55 min from which the navigation data is extracted.

Autonomous vehicles employ a variety of sensors to alter their trajectory, as seen in the image. These include a camera, speed, acceleration, and global positioning system (GPS). In the initial lane view, with a steering angle of $\mathsf{\theta }\mathrm{o}\mathrm{r}180°$, the sensors watch the automobile for its position and speed. This setup guarantees that the vehicle knows its intended path and orientation before changing lanes.

During the direction change phase, the steering angle is adjusted in a series of steps, beginning at $\left(-45°,-60°\right)$ degrees and finishing at $90°$ degrees. The vehicle can adjust to environmental changes effectively since the camera sensor records this direction shift. In the final frame, After the direction change, the vehicle returns to a straight path at an angle between zero and ninety degrees after changing directions $\mathsf{\theta }\mathrm{o}\mathrm{r}180°$. This series demonstrates the vehicle’s directional manoeuvring capabilities, including its ability to maintain a constant speed and orientation.

3.2. Framework Description

The COI framework aims to provide AV users with a safe driving and navigation experience. This framework considers vehicle, environmental, and computational intelligence to fulfil its purpose. This framework application for the AV scenario is depicted in Figure 2. In the above depiction, the “Things” refer to the access points, traffic lights, cloud devices, base stations, etc. that are interfaced using the wireless communication media to the AVs. These “Things” provide users with infotainment, navigation, location services, and messaging features. The AVs define their paths through a persistent interaction schedule and learn the environment to ensure safe driving. Environmental learning refers to the visual and communication sensed information the AV utilises for speed and direction control, movement, lane hanging, etc. The learning by the AV is facilitated by the actuating and communicating sensors/devices attached to the vehicle. The fundamental information requirement is the starting and ending location, lane, safe distance, and vehicle acceleration. The following/trailing vehicles communicate with the AV to broadcast their status (speed, direction, etc.), which further serves as the learning input for the vehicle (Figure 2). Communication between the autonomous car and the interacting “Things” allows it to extract and use local data, including location, lane, visual, etc., for reliable data sharing.

Road safety management is improved using consistent vehicle interaction that obtains the knowledge of novel sharing between the AV and interacting “Thing” to retain driving and vehicle safety. In the case of different locations or lanes or visuals in front, vehicle interaction and consistency change accordingly. Estimating the precise and accurate minimum safety distance between the vehicles and interacting “Things” is important. Hence, the minimum safety distance ${Safe}_D$ is computed as

$${Safe}_D={\displaystyle \frac{{\left({RV}_s\right)}^2}{2E\left(rd\right)}}{OV}_s+{LV}_s+c$$

(1)

where, ${RV}_s$ is the relative vehicle speed, $E\left(rd\right)$ signifies the expected deceleration of the relative vehicle, ${OV}_s$ signifies the overtaking vehicle speed, ${LV}_s$ signifies the leading vehicle speed and $c$ denotes the consistency of vehicle interaction. The point-to-point interaction is defined as a distance to the AV and interacting “Things” that exceeds the minimum safety distance, while the following extraction and utilisation process is characterised by a location, lane, and visual that is less than ${Safe}_D$.

To guarantee road safety, the shown equation determines the minimal safe distance (safe) that must be maintained between an autonomous vehicle (AV) and an interacting “Thing” (such as another AV or an object). This component considers the following: the relative vehicle’s speed ${RV}_s$, the relative vehicle’s predicted deceleration $E\left(rd\right)$ the relative vehicle’s speed when passing ${OV}_s$, the relative vehicle’s speed when leading ${LV}_s$ and a stability factor $\left(c\right)$ for the interaction between vehicles. Autonomous vehicles can maintain a safe following distance in all situations by figuring this out and considering things like acceleration and deceleration rates. An important safety buffer during AV encounters in many traffic conditions, this dynamic measure may respond to location, lane, or visual signal changes. The equation shows that the consistency factor in vehicle encounters is represented by c. It ensures that the AV’s interactions with other “Things” in its surroundings, such as other cars or barriers, are stable and dependable. This factor modifies the minimum safety distance to account for variations in interaction quality that occur in real-time. The model ensures that the AV has a flexible safety margin by integrating c, which makes it more resistant to changes on the road.

In this scenario, the information transmission refers to the AVs and interacting “Thing” undertaking a location, lane, and visual change trick.

$${Int}_{RV}=\left\{\begin{array}{c}{max}_s, interacting\\ max\left\{{LV}_s,{max}_s\right\}, extracting\\ {max}_s, utilizing\end{array}\right.$$

(2)

where, ${Int}_{RV}$ represents the interaction of relevant vehicles, ${max}_s$ is the maximum speed, and ${LV}_s$ represents the leading vehicle speed.

The above equation describes the interaction, ${Int}_{RV}$ that occurs during data transmission between AVs and an external “Thing” (such as other AVs or traffic-related factors). The three possible interactions—interacting, extracting, and utilising—depend on the top speed and leading vehicle speed. The AVs modify their pace and behaviour depending on their environment in each situation, which constitutes a unique kind of information exchange. When interacting, an AV takes its maximum speed into account; when extracting, it compares the speed of the leading vehicle ${LV}_s$ with its maximum speed and chooses the greater value. This method enables the AV to make real-time, behavioural adjustments that improve its interaction with its surroundings while minimising risk.

3.3. Information-Sharing Process

This proposed framework relies on consistent vehicle interaction to address decision issues suitable for adaptability using the interacting “Things”. This monitors the complete location, lane, visual extraction, and utilisation process. In this process, a typical consideration of extracted data, such as location, lane, and visual, is utilised to analyse the information on environmental knowledge. Based on the instance, the front vehicle (FV), interacting “Thing” (IT), and rear vehicle (RV) in the current location/lane/visual is required for appropriate information transmission, while the traffic, objects, the population in the current location/lane/visual is monitored on the roadside.

$$\left\{\begin{array}{c}{X}_i\left(T\right)=X_i^0+W_i{\mu }_i\left(T\right)+{\exists }_i\left(T\right)\\ Y_i\left(T\right)=X_i^0+W_i{\mu }_i\left(T\right)\end{array}\right.$$

(3)

In Equation (3), the variable $X_i\left(T\right)$ represents the initial extracted factors from a certain location, lane, and visual and ${\mu }_i\left(T\right)$ is the environmental knowledge data observed from different time instances. Where, ${\exists }_i\left(T\right)$ signifies the uncertainty in location, lane, and visual resulting in noise occurrence. The $Y_i\left(T\right)$ are the utilised factors and $W_i$ are the weights. Based on the above Equation (3), the uncertainty case is discussed as a lemma that impacts the vehicle’s behaviour.

Lemma 1. 

${\exists }_i\left(T\right)$ impacts $c$ for which $In{t}_{RV}$ is expected to be optimal.

Discussion 1. 

The first lemma of ${\exists }_i\left(T\right)$ assessment over $c$ and $In{t}_{RV}$ impacts the ${\mu }_i\left(T\right)$ of the AV at any distance or time. Therefore, the impact is described using the derivatives defined below in

Table 1. Impacts of vehicle’s behaviour Lemma 1.

$\forall In{t}_{RV}={max}_{s}in$ ${\exists }_i\left(T\right)=X_i\left(T\right)-X_i^o\pm W_i{\mu }_i\left(T\right)$ where $T=\left\{\mathrm{1,2},\dots ,{\displaystyle \frac{{max}_s}{Saf{e}_D}}\right\}$
Based on $X_i\left(T\right)$		Based on ${\exists }_i\left(T\right)$
Define $X_i\left(T\right)={\sum }_{i=1}^T{X}_i^o,W_i{\mu }_i\left(T\right)=0$ ${\exists }_i\left(T\right)=0$ unless $W_i{\mu }_i\left(T\right)=0$ $In{t}_{RV}={max}_s$ and $In{t}_{RV}={min}_s$ $\forall {max}_s$, if $E\left(rd\right)=0$, then ${\mu }_i\left(T\right)=Y_i\left(T\right)=\left\{{max}_s,{min}_s,O{V}_s,L{V}_s\right\}$ Impact 1: ${\mu }_i\left(T\right)=X_i\left(T\right)-Y_i\left(T\right)$ //low where $In{t}_{RV}={\sum }_{i=1}^T{W}_i\times \left[Y_i\left(t\right)-c\right]$ $\forall In{t}_{RV}={min}_s,c=W_i\times {\displaystyle \frac{Y_i\left(T\right)}{X_i\left(T\right)}}$ Impact 2: ${\mu }_i\left(T\right)=X_i\left(T\right)-{\exists }_i\left(T\right)$//high ${\exists }_i\left(T\right)=X_i-{\mu }_i\left(T\right)\forall {\displaystyle \sum _{i=0}^T}c=In{t}_{RV}$ where $In{t}_{RV}={\sum }_{i=1}^T{W}_i\times \left[Y_i\left(t\right)-c\right]$ $\forall In{t}_{RV}={min}_s,c=W_i\times {\displaystyle \frac{Y_i\left(T\right)}{X_i\left(T\right)}}$ Impact 2: ${\mu }_i\left(T\right)=X_i\left(T\right)-{\exists }_i\left(T\right)$//high ${\exists }_i\left(T\right)=X_i-{\mu }_i\left(T\right)\forall {\displaystyle \sum _{i=0}^T}c=In{t}_{RV}$		Define ${\mu }_i=X_i\left(T\right)-{\exists }_i\left(T\right)$ For the $Saf{e}_D={\displaystyle \frac{R{V}_s}{2E\left(rd\right)}}$, the ${\mu }_i$ impacts as follows $E\left(rd\right)=R{V}_o-{\displaystyle \frac{{max}_s}{T}}$ such that $X_i\left(T\right)=Y_i\left(T\right)+W_i\times {\exists }_i\left(T\right)$ and Impact 1: $c=Saf{e}_D-Y_i\left(T\right)-X_i\left(T\right)$ //low Impact 2: $c=0\forall {\exists }_i>Y_i\left(T\right)$

.

In this process, the extracted data addresses the unique characteristics of AVs and interacting “Things”, including environmental knowledge data. The environmental knowledge data are defined in Equation (4),

$$\left\{\begin{array}{c}{\mu }_{NL,Nl,NV}=\left[X_i^0+{∆}_i{X}_i\left(T\right)+{\exists }_i\left(T\right)\right]\\ {\mu }_{CL,Cl,CV}=\left[X_i^0+{∆}_i{X}_i\left(T\right)\right]\end{array}\right.$$

(4)

where, the variables $NL$ and $CL$ represent the nearby and current locations; $Nl$ and $Cl$ represent the nearby and current lanes;$NV$ and $CV$ represent the nearby and current visual. If $\alpha$ and $\beta$ are the position and interaction between the nearby vehicles and “Things”; $\gamma$ signifies the type of AV and interacting “Things”. In this scenario, when the AVs interact with other “Things”, the output is $1;$ otherwise, it is $0$. In this process, the maximum extraction and utilisation are estimated. Using the estimation output, the autonomous vehicle with appropriate environmental knowledge is tracked to improve the consistency and adaptability measures through the neuro-batch learning algorithm. The interactions are batched with the environmental knowledge data, with the measures of $-1$, $0,$ and $1$ corresponding to the batches and driving time. Each factor in environmental knowledge is categorised to avoid unsafe navigation responses.

3.4. Batch Optimisation Algorithm

The neuro-batch learning algorithm is based on a random utilisation scenario, which assumes that the random integer. $q_i$ is alone and follows the learning rate throughout its driving time. The learning, as mentioned in the above algorithm, implies that the probability of data extraction and utilisation corresponds to the precise tracking of AV.

$$P\left(X_i,Y_i\right)={\displaystyle \frac{exp\left[X_i\left(T\right)\right]}{{\sum }_{N\in i}exp\left[X_N\left(T\right)\right]}}$$

(5)

To enhance driving and road safety, the interactions are adaptively batched over time to determine the utilisation of location, lane, visual, and extraction probability from the current instance to the future. If $N$ represents the type of traffic information observed from the roadside environment in different time intervals. Based on the objects, population, and traffic conditions, the environmental knowledge of the AV and its learning rate are categorised until its driving time through the learning process. This algorithm increases the environmental knowledge data for performing appropriate batching. The target location is chosen based on the actuated and sensed information from the vehicle’s environmental knowledge. This enhances driving and road safety by appropriately selecting the utility of each location and lane in the future.

$$B\left({\mu }_i\right)=\arg \begin{array}{c}max\\ i\end{array}\left({\int }_T^{T+1}P\left(X_i,Y_i\right).TdT\right)$$

(6)

The batch optimisation model for risk assessment is illustrated in Figure 3.

The model of the batch optimisation is illustrated in Figure 3 for $T·dT$ detection using $X_i$ inputs. The $P\left(X_i,Y_i\right)$ is split into two combinations: $\left(X_i=1,P_i=0\right)$ and $\left(X_i=1,P_i=1\right)$. In the other two conditions (i.e.,) $X_i=0$ the input is not fetched, and therefore, previous decisions are retained. Therefore, for the two conditions, the batch combinations are: $\left(c,W_i\right)$ and $\left(In{t}_{RV},W_i\right)$. Using these combinations, $\left[X_i\left(T\right),X_i\left(T-1\right),X_i\left(T-2\right),\dots ,X_i\left(1\right)\right]$ an $\left[X_N\left(1\right),X_N\left(2\right),\dots ,X_N\left(T\right)\right]$ integration, are verified to ensure ${\mu }_i=max$ and ${\exists }_i=0$. This verification is pursued with $W_i$ to ensure $T·dT$ outputs a unity value that holds $P\left(X_i,Y_i\right)=1$. Besides, ${\mu }_{NL},Nl,NV$ is the $T$ derivative for ${∆}_i{X}_i\left(T\right)+{\exists }_i\left(T\right)$ such that $Y_i\left(T\right)=c+W_i\left(T\right)=c\forall W_i=0$. The second batch optimisation requires ${\exists }_i\left(T\right)$ under $\left[{∆}_i\left(T\right)-X_i\left(T\right)\right]$ is validated to identify any $W_i=0$ in $T$. The ${\mu }_i\left(T\right)$ is iterated until ${\mu }_i\left(T\right)=max$ and ${\exists }_i=0$ achieving high precision of $P\left(X_i,Y_i\right)$.

The neuro-batch learning algorithm is well suited to the proposed method, which enables the AV to adapt instantaneously to its environment. The system can effectively assess and classify traffic data, including vehicle speeds, obstacles, and pedestrian movements, using batch processing. It is safer to utilise the AV when things are continually changing because of its flexibility, allowing it to adapt to changing conditions precisely. In contrast to traditional methods, neuro-batch learning continuously improves the AV’s decision-making with new information, allowing it to manage the multiple interconnections required by the COIF.

3.5. Risk Assessment

Once the intention is to categorise, safe and unsafe activity of the vehicle has been identified using the connected learning neurons to the information for unidentified scenarios. The risk assessment technique is used to verify the risk occurrence in the vehicle’s decision over unavailable or inconsistent information to regularise its existing neuro batches for ensuring safe navigation responses. The framework for autonomous vehicles identifies the multiple risk factors from the different types of traffic information required in the current location, lane, and visual. However, this module considers only single-vehicle interaction and does not consider consistency measures adaptability. This article proposes a new risk indicator for easily cut navigation during unsafe driving.

By optimising risk assessment in batches and using the neuro-batch learning algorithm for dynamic communication between AVs and interacting “Things,” this technique aims to manage complicated, real-world situations better. Installing a new risk indication for early detection of hazardous driving is an extra precaution for preventative safety that allows swift navigational alterations. This technique aims to enhance road safety by making AVs more dependable and adaptive in unpredictable situations.

The risk indicators assessment is validated to categorise safe and unsafe driving. Subsequently, the learning neurons are connected to compute the likelihood of safe navigation at each level of risk. Ultimately, the identifying risk level with the interacting “Thing” is detected as the unsafe navigation risk level. The first assessment is to categorise the environmental knowledge of the vehicle $\delta$ throughout its driving time using neural learning. The neurons are connected to the actuated and sensed information, which can be employed to identify the unsafe activity where the extracted feature is close to zero, and interaction undergoes some variation.

$$\delta =\left\{\begin{array}{c}{\displaystyle \frac{{\mathbb{V}}^f-{\mathbb{V}}^l}{{\mathbb{L}}^f-{\mathbb{L}}^l}}, {\mathbb{V}}^f>{\mathbb{V}}^l \\ 0, {\mathbb{V}}^f\le {\mathbb{V}}^l\end{array}\right.$$

(7)

In Equation (7), the variables $L^f$ and $L^l$ represent the longitudinal positions of the vehicles is observed in both leading and following of the certain AV, while ${\mathbb{V}}^f$ and ${\mathbb{V}}^l$ are the corresponding velocities of the vehicles. The second assessment is the computation of the safety distance index $\left(sdi\right)$ of the vehicle. The environmental knowledge gained of the AV using the $Y_i\left(T\right)$ is discussed using Lemma 2. Table 2 explains the environmental knowledge gain Lemma 2.

Lemma 2. 

${\mu }_i\left(T\right)$ Gain using $Y_i\left(T\right)$

Discussion 2. 

Table 2. Environmental knowledge gain Lemma 2.

	Batches $\begin{array}{c}B\left({\mu }_i\right)={\int }_{T=1}P\left(\mathrm{1,0}\right)·TdT\\ :\\ B\left({\mu }_i\right)={\int }_{T=1}P\left(\mathrm{1,1}\right)·TdT\end{array}$
$\delta =1\forall {\exists }_i=0$$\left(\mathrm{or}\right) {\exists }_i\ne 0$ The batch derivatives are: $P\left(\mathrm{1,0}\right)={\displaystyle \frac{X_i\left(T\right)}{X_n\left(T\right)}},{\mu }_i=max$ $P\left(\mathrm{1,1}\right)={\displaystyle \frac{X_i\left(T\right)}{X_n\left(T\right)}}-W_i\times {\mu }_i$ ${\exists }_i=1-\left[Y_i\left(T\right)\cap X_i\left(T\right)\right]$ ${\mu }_i\left(T\right)=\left\{\begin{array}{c}P\left(\mathrm{1,0}\right)-{\exists }_i\forall \delta =0\\ P\left(\mathrm{1,1}\right)-{\displaystyle \frac{W_i}{{\exists }_i}}\forall \delta =1\end{array}\right.$ Therefore, $B\left({\mu }_i\right)=\left\{\begin{array}{c}{\mu }_i\left(T\right)·P\left(\mathrm{1,0}\right)dT\\ {\mu }_i\left(T\right)-{\exists }_i·P\left(\mathrm{1,1}\right)dT\end{array}\right.$		$\delta =1\forall {\exists }_i\notin 0\&1$$, \mathrm{but} 0<{\exists }_i<1$ The batch derivatives are: $Y_i\left(T\right)=X_i^o+W_i{\mu }_i\left(T\right)$ $\mathrm{Subs} \mathrm{for} {\mu }_i\left(T\right)$ from LHS, $Y_i\left(T\right)=\left\{\begin{array}{c}{X}_i^o+W_i·P\left(\mathrm{1,0}\right)-{\exists }_i\\ X_i^o+W_i·P\left(\mathrm{1,1}\right)-{\displaystyle \frac{W_i}{{\exists }_i}}\end{array}\right.$ and thus, ${\mu }_i\left(T\right)·P\left(\mathrm{1,0}\right)=X_i^o+W_i·P\left(\mathrm{1,0}\right)$ ${\mu }_i\left(T\right)-{\exists }_i·P\left(\mathrm{1,1}\right)=X_i^o+W_{i}P\left(\mathrm{1,1}\right)-{\displaystyle \frac{W_i}{{\exists }_i}}$ $\mathrm{The} \mathrm{Crain} \mathrm{is} {\mu }_i\left(T\right)$$\mathrm{for} {\exists }_i$ and ${\mu }_i\left(T\right)$$\mathrm{in} \mathrm{any} T·dT=1$

Table 2. Environmental knowledge gain Lemma 2.

	Batches $\begin{array}{c}B\left({\mu }_i\right)={\int }_{T=1}P\left(\mathrm{1,0}\right)·TdT\\ :\\ B\left({\mu }_i\right)={\int }_{T=1}P\left(\mathrm{1,1}\right)·TdT\end{array}$
$\delta =1\forall {\exists }_i=0$$\left(\mathrm{or}\right) {\exists }_i\ne 0$ The batch derivatives are: $P\left(\mathrm{1,0}\right)={\displaystyle \frac{X_i\left(T\right)}{X_n\left(T\right)}},{\mu }_i=max$ $P\left(\mathrm{1,1}\right)={\displaystyle \frac{X_i\left(T\right)}{X_n\left(T\right)}}-W_i\times {\mu }_i$ ${\exists }_i=1-\left[Y_i\left(T\right)\cap X_i\left(T\right)\right]$ ${\mu }_i\left(T\right)=\left\{\begin{array}{c}P\left(\mathrm{1,0}\right)-{\exists }_i\forall \delta =0\\ P\left(\mathrm{1,1}\right)-{\displaystyle \frac{W_i}{{\exists }_i}}\forall \delta =1\end{array}\right.$ Therefore, $B\left({\mu }_i\right)=\left\{\begin{array}{c}{\mu }_i\left(T\right)·P\left(\mathrm{1,0}\right)dT\\ {\mu }_i\left(T\right)-{\exists }_i·P\left(\mathrm{1,1}\right)dT\end{array}\right.$		$\delta =1\forall {\exists }_i\notin 0\&1$$, \mathrm{but} 0<{\exists }_i<1$ The batch derivatives are: $Y_i\left(T\right)=X_i^o+W_i{\mu }_i\left(T\right)$ $\mathrm{Subs} \mathrm{for} {\mu }_i\left(T\right)$ from LHS, $Y_i\left(T\right)=\left\{\begin{array}{c}{X}_i^o+W_i·P\left(\mathrm{1,0}\right)-{\exists }_i\\ X_i^o+W_i·P\left(\mathrm{1,1}\right)-{\displaystyle \frac{W_i}{{\exists }_i}}\end{array}\right.$ and thus, ${\mu }_i\left(T\right)·P\left(\mathrm{1,0}\right)=X_i^o+W_i·P\left(\mathrm{1,0}\right)$ ${\mu }_i\left(T\right)-{\exists }_i·P\left(\mathrm{1,1}\right)=X_i^o+W_{i}P\left(\mathrm{1,1}\right)-{\displaystyle \frac{W_i}{{\exists }_i}}$ $\mathrm{The} \mathrm{Crain} \mathrm{is} {\mu }_i\left(T\right)$$\mathrm{for} {\exists }_i$ and ${\mu }_i\left(T\right)$$\mathrm{in} \mathrm{any} T·dT=1$

This validation ensures that AVs are maintained at the precise distance between interacting “Things”.

$$sdi={\displaystyle \frac{{Safe}_D}{\left|{\mathbb{L}}^f-{\mathbb{L}}^l\right|}}$$

(8)

where, ${Safe}_D$ signifies the minimum safety distance as defined in Equation (1). The third assessment is the computation of the total risk $\left(\omega \right)$ based on the traffic information obtained. The risk factor is related to the kinetic energy of the traffic unit, as shown in Equation (9).

$$\omega =T\left(\mathbb{M}+{\displaystyle \frac{1}{2}}\mathbb{M}{\mathbb{V}}^2+{\displaystyle \frac{2N}{{\mathbb{V}}_{lim}^2}}\right)$$

(9)

In Equation (9), $\mathbb{M}$ indicates the mass value, $\mathbb{V}$ indicates the velocity, $N$ is the type of traffic information required coefficient and ${\mathbb{V}}_{lim}^2$ indicates the speed limit of the selected road lane. The risk indicator based on lane characteristics is completed.

$$D_{ji}={\displaystyle \frac{1}{{\left|{\epsilon }_{v_{ji}}\right|}^x}}\times {\displaystyle \frac{{\epsilon }_{v_{ji}}}{\left|{\epsilon }_{v_{ji}}\right|}}$$

(10)

From the instance, the variable ${\epsilon }_{v_{ji}}$ represents the coordinate vector from $j$ to $i$ time interval and $x$ means the descending gradient coefficient of the vehicle. The risk assessment is related to the unsafe navigation of the movement $\vartheta$. The learning categorisation coefficients can be expressed as

$$D_{\vartheta ,ji}=exp\left[\left|R_{\mathbb{V}}\right|\times cos\theta \right]$$

(11)

In Equation (11), $R_{\mathbb{V}}$ means the relative velocity of $i$ and $j$ vectors and $\theta$ represents the angle between the AV and interacting “Things”. From the considered dataset values, the impact of batch optimisation over the metrics discussed in Equations (8)–(11) is studied in Figure 4.

The impact of $Saf{e}_D$ and $P\left(X_i,Y_i\right)$ is induced by the batches in any $T$ provided the $\omega$ lends are precisely identified. The derivatives in Lemma 2 are validated to maximise $B\left({\mu }_i\right)$ under $TdT$. In this assessment, the $Y_i\left(T\right)$ and it is corresponding $W_i$ adjustments are pursued to ensure $-{\exists }_i$ is suppressed. Through different $\left\{1-\left[Y_i\left(T\right)\cap X_i\left(T\right)\right]\right\}<{\mu }_i\left(T\right)$ instances are extracted to ensure ${\exists }_i$ is suppressed. In the batch optimisation process, the $Y_i\left(T\right)$ differentiations are pursued to ensure $X_i\left(T\right)=Y_i\left(T\right)$. If this case is true then $T·dT$ generates a unity output. The chances of unity pose $D_{ji}$ and $D_{v_{ji}}$ through ${\epsilon }_{v_{ji}}$ differences for ${\mathbb{V}}_{lim}^2$. Hence, $\omega$ from $-{\exists }_i$ and $\left[sdi>\left|{\mathbb{L}}^f-{\mathbb{L}}^l\right|\right]$ validations, $W_i$ modifications are pursued to ensure $n_o$ unsecured/information omission takes place. Besides, the physical parameter $\theta$ of the vehicle for $L{V}_s$ or $O{V}_s$ following or $E\left(rd\right)$ are identified based on the rate of $x\forall$ any $\omega$. Thus, $Y_i\left(T\right)={\displaystyle \frac{-W_i}{{\exists }_i}}$ is the $\omega$ deciding factor across $P\left(\mathrm{1,0}\right)$ and $P\left(\mathrm{1,1}\right)$ such that $\left(W_i\ast {\mu }_i\right)$ is the balancing factor for $B\left({\mu }_i\right)$. This case is inferred in the Lemmas 2’s ${\mu }_i\left(T\right)·P\left(\mathrm{1,0}\right)$ and ${\mu }_i\left(T\right)-{\exists }_i·P\left(\mathrm{1,1}\right)$ derivatives for $B\left({\mu }_i\right)$ (Refer to Figure 4 Representations). The traffic regulations and lane marks are used to limit vehicle navigation, and the environmental knowledge of the vehicle will filter all risk factors on the roadside.

The coefficient of the filtering process can be expressed as

$$F_i=\left\{\begin{array}{c}N\times \mathit{cos}\left(\left|p_v\right|\times {\displaystyle \frac{\pi }{6\tau }}\right), \left|p_v\right|>{\displaystyle \frac{\tau }{2}} \\ 1, \left|p_v\right|\le {\displaystyle \frac{\tau }{2}}\end{array}\right.$$

(12)

In Equation (12), $p_v$ is the position vector, $\tau$ is the width of the lane and $N$ is the type of traffic information coefficient from the centre line. The proposed framework precisely identifies the risk factors associated with vehicle velocity, mass, and consistency measures adaptability are taken to address the impact of unsafe driving navigation response. The final risk assessment index ${Risk}_a$ can be expressed as

$${Risk}_a=\left|{\omega }_{front-leadingvehicle}+{\omega }_{rear-leadingvehicle}\right|$$

(13)

The above three assessments are used to assess the accurate level of the risks as $X=\left\{\delta ,sdi,\omega ,{Risk}_a\right\}$ and gives notification to the vehicle based on the three risk levels as expressed as $\mathsf{\Psi }=\left\{safe\left(\mathbb{S}\right),unsafe\left(\mathbb{U}\right)\right\}$, unsafe in the sense of providing dangerous or alert signals. Assume that the consistency of the vehicle interaction $\mathbb{x}$ and the risk level $\mathsf{\Psi }$ satisfies the following Equation:

$$p_v\left(\left.X=\mathbb{x}\right|\mathsf{\Psi }={\mathsf{\psi }}_i\right)\propto \left\{\begin{array}{c}exp\left(-{\displaystyle \frac{{\left(\mathbb{x}-G\right)}^2}{2\sigma }}\right)+{\displaystyle \frac{1-rg}{2}}, if \mathbb{x}>G \\ exp\left(-{\displaystyle \frac{{\left(\mathbb{x}-H\right)}^2}{2\sigma }}\right)+{\displaystyle \frac{1-rg}{2}}, if \mathbb{x}>H\\ rg, otherwise\end{array}\right.$$

(14)

In Equation (14), $G$ and $H$ are upper and bottom threshold conditions in decision-making. Where the variable $rg$ denotes the regularisation factor, and $\sigma$ is the uncertainty of observational information. The neurons’ connectivity can be estimated by incorporating the different risk weight assignments. In Lemma 3, the $p_v$ derivatives for $Ris{k}_a$ classification (safe and unsafe) are discussed. Table 3 describes the risk classification of Lemma 3.

Lemma 3. 

The role of $p_v$ in AV safety deciding is inevitable based on $Ris{k}_a$

Discussion 3. 

Table 3. Risk classification of Lemma 3.

	$\mathit{P}(\mathit{X}=\mathbb{X})\mathit{\alpha }{\displaystyle \frac{{\left(\mathbb{X}-\mathit{G}\right)}^2}{2\mathit{\sigma }}}$ $\mathit{P}(\mathit{X}=\mathbb{X})\mathit{\alpha }{\displaystyle \frac{{\left(\mathit{X}-\mathit{H}\right)}^2}{2\mathit{\sigma }}}$
$\mathbb{X}>\mathit{G}$		$\mathbb{X}>\mathit{H}$
$\mathrm{Consider} D_{v_{ji}}$$\mathrm{as} exp\left|R_V\times Co{s}_{\theta }\right|$ $\mathrm{for} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =0°$ (no change) $D_{v_{ji}}=R_V;F_i=N;$ $\mathrm{Equating} \mathrm{with} p_v$ $R_V={\displaystyle \frac{-{\left(x-G\right)}^2}{2\sigma }}$$, rg=0$ and $2\sigma R_V+{\left(x-G\right)}^2=0$ is the safe Driving criteria $\mathrm{For} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =90°$ (maximum) $D_{v_{ji}}=0;F_i=1$ ${\displaystyle \frac{-{\left(x-G\right)}^2}{2\sigma }}=o\Rightarrow {\left(x-G\right)}^2=0$ is the Safe driving criteria		$\mathrm{Consider} F_i=1$$\mathrm{for} Y_i\left(T\right)=-{\exists }_i$ $\mathrm{for} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =0°$ (no change) $P\left(X_i={\mathbb{X}}_i\left|\psi ={\psi }_i\right.\right)={\displaystyle \frac{p_v}{1}}=p_v$ $p_v\left(X=\mathbb{X}\left|\psi ={\psi }_i\right.\right)\alpha {\displaystyle \frac{\left(1-r_g\right)}{2}}\forall 2\sigma ={\left(x-H\right)}^2$ and $p_v\left(X-\mathbb{X}\left|\psi ={\psi }_i\right.\right)\alpha r_g=Ris{k}_a$ $\mathrm{for} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =90°$ (maximum) $D_{ji}={\displaystyle \frac{1}{1^x}}\times {\displaystyle \frac{\left(1-r_g\right)}{1^x}}={\displaystyle \frac{\left(1-r_g\right)}{1^x}}$ $\mathrm{where} x={\displaystyle \frac{{\left(X-H\right)}^2}{2\sigma }}$

Table 3. Risk classification of Lemma 3.

	$\mathit{P}(\mathit{X}=\mathbb{X})\mathit{\alpha }{\displaystyle \frac{{\left(\mathbb{X}-\mathit{G}\right)}^2}{2\mathit{\sigma }}}$ $\mathit{P}(\mathit{X}=\mathbb{X})\mathit{\alpha }{\displaystyle \frac{{\left(\mathit{X}-\mathit{H}\right)}^2}{2\mathit{\sigma }}}$
$\mathbb{X}>\mathit{G}$		$\mathbb{X}>\mathit{H}$
$\mathrm{Consider} D_{v_{ji}}$$\mathrm{as} exp\left|R_V\times Co{s}_{\theta }\right|$ $\mathrm{for} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =0°$ (no change) $D_{v_{ji}}=R_V;F_i=N;$ $\mathrm{Equating} \mathrm{with} p_v$ $R_V={\displaystyle \frac{-{\left(x-G\right)}^2}{2\sigma }}$$, rg=0$ and $2\sigma R_V+{\left(x-G\right)}^2=0$ is the safe Driving criteria $\mathrm{For} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =90°$ (maximum) $D_{v_{ji}}=0;F_i=1$ ${\displaystyle \frac{-{\left(x-G\right)}^2}{2\sigma }}=o\Rightarrow {\left(x-G\right)}^2=0$ is the Safe driving criteria		$\mathrm{Consider} F_i=1$$\mathrm{for} Y_i\left(T\right)=-{\exists }_i$ $\mathrm{for} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =0°$ (no change) $P\left(X_i={\mathbb{X}}_i\left|\psi ={\psi }_i\right.\right)={\displaystyle \frac{p_v}{1}}=p_v$ $p_v\left(X=\mathbb{X}\left|\psi ={\psi }_i\right.\right)\alpha {\displaystyle \frac{\left(1-r_g\right)}{2}}\forall 2\sigma ={\left(x-H\right)}^2$ and $p_v\left(X-\mathbb{X}\left|\psi ={\psi }_i\right.\right)\alpha r_g=Ris{k}_a$ $\mathrm{for} \mathrm{the} \mathrm{lane} \mathrm{change} \theta =90°$ (maximum) $D_{ji}={\displaystyle \frac{1}{1^x}}\times {\displaystyle \frac{\left(1-r_g\right)}{1^x}}={\displaystyle \frac{\left(1-r_g\right)}{1^x}}$ $\mathrm{where} x={\displaystyle \frac{{\left(X-H\right)}^2}{2\sigma }}$

Based on the instance, the probability of a particular risk level can be determined as

$$P\left(\left.X_i={\mathbb{x}}_i\right|\mathsf{\Psi }={\mathsf{\psi }}_i\right)={\displaystyle \frac{{\sum }_{i=1}^{N_{\mathsf{\Psi }}}{W}_i{p}_v\left(\left.X=\mathbb{x}\right|\mathsf{\Psi }={\mathsf{\psi }}_i\right)}{{\sum }_{i=1}^{N_{\mathsf{\Psi }}}{W}_j}}$$

(15)

When considering the unsafe activity of the vehicle, the risk factors originate from the actuated and sensed information of AV in the vehicle’s decision. Consequently, the probabilities of unsafe navigation responses should be estimated by joining all the respective risk levels by using a new function that is expressed as

$$\left\{\begin{array}{c}P\left({\mathsf{\psi }}_i=\mathbb{S}\right)=1-{\displaystyle \sum _{N=\mathrm{0,1}}}P\left({\mathsf{\psi }}_N=\mathbb{S}|X_i^N={\mathbb{x}}_M^{i,N}\right)\\ P\left({\mathsf{\psi }}_i=\mathbb{U}\right)=1-{\displaystyle \sum _{N=\mathrm{0,1}}}P\left({\mathsf{\psi }}_N=\mathbb{U}|X_i^N={\mathbb{x}}_M^{i,N}\right)\\ P\left({\mathsf{\psi }}_i=\mathbb{S}-\mathbb{U}\right)=1-P\left({\mathsf{\psi }}_N=\mathbb{S}\right)-P\left({\mathsf{\psi }}_N=\mathbb{U}\right)\end{array}\right.$$

(16)

Equation (16) follows the information-sharing process described above that allows the environmental knowledge of the vehicle to be observed and the risk level to be identified through the batch optimisation algorithm and neural learning process. The risk of changing unsafe navigation responses is determined based on the risk level when the learning neurons are activated and connected to classify safe and unsafe driving. During an emergency, an alert signal indicates the vehicle is unsafe. The unsafe AV navigation function outputs above are analysed graphically using Figure 5.

The $F_i$ and $Ris{k}_a$ at $G$ and $H$ for the derivatives $P({\psi }_i=\mathbb{S}-\mathbb{U}$ is represented in Figure 4. The $pv$ variants for $(\mathbb{X}>G)$ and $(\mathbb{x}>H)$ are the key factors that decide the $rg$ and $\psi$ for $AV$ navigation. The possibilities are ${\psi }_N=\mathbb{S},{\psi }_N=\mathbb{U}$ for the individual unity factors for providing precise driving controls. Therefore for $F_i$ the $G$ and $H$ are the inverse for $\psi =\left({\mathbb{X}}_i, {\omega }_i, {\exists }_i, D\right)$ based on the information acquired such that $\left[1-P\left({\psi }_N\right)\right]\forall \mathbb{S}$ (or) $\mathbb{U}$ maximises the chances of safe driving over $Ris{k}_a$. Thus, the cases violating the above are unsafe navigations (H cases) in all of the above $F_i$ differentiated outputs. These points require $\omega$ and $D{v}_{ji}$ modifications for ${\left|\mathbb{X}-G\right|}^2$ and ${\left|\mathbb{X}-H\right|}^2 \forall r_g$.

3.6. Driving Control Design

In this training process, the AV uses the straight road to drive; if there is any abnormality, it takes a new lane using the environmental knowledge data. Therefore, the control optimiser interaction framework must extend the features to make appropriate decisions. The state space of the control optimiser interaction is given as

$$\phi =\left({\mathbb{x}}_i,{\omega }_i,{\exists }_i,offset,laststeer,D\right)$$

(17)

In Equation (17), the vehicle’s position follows the safe navigation after regularisation. The angle between the vehicle and the interacting things is completed to follow the safe lane. In general, multiple factors interrupt the vehicle interaction consistency to ensure driving and pedestrian safety. This article presents a batch optimisation-based consistency measure adaptability prediction model using the neural learning paradigm. The neural learning process identifies the unsafe activity of the vehicle of different risk/abnormality factors towards safe navigation.

4. Results

In the discussion section, the learning rate, environment detection, navigation assistance, and information latency metrics are comparatively analysed between the proposed and existing methods. The existing methods considered are ORRL [31], HDSE-DRL [28], and CRL-MPC [26], which are introduced in the related works section. These metrics are comparatively analysed using distance (100 and 1000 m) and time (up to 55 min). This comparative discussion is used to verify the efficiency of the proposed method over the other methods through the percentage of improvement estimated and described under each metric discussed below. Therefore, in Figure 6, the comparative illustration of the learning rate is depicted.

The environmental data extraction and utilisation process improves the tracking of autonomous vehicles with appropriate environmental knowledge. Through batching of interactions, it is easy to identify the type of traffic information sharing between the AV and interacting “Thing” over different time domains. The precise observation and information extraction from the previous neuro batch and utilisation computation output are used for retaining driving and road safety. The unsafe activity of the AV is addressed by using COIF and batch optimisation algorithms based on the $X_i\left(T\right)$ and $Y_i\left(T\right),$ the condition that satisfies successive observation of the safe navigation, preventing ${\exists }_i\left(T\right)\le Y_i\left(T\right)$ and $Y_i\left(T\right)\le {\exists }_i\left(T\right)\le X_i\left(T\right)$. Therefore, the following information-sharing sequence analysis is performed for a feasible solution and is not represented. Based on (3) and (4), the conditions satisfy a high learning rate in precise categorisation-based neurons connected. In this process, a high safe navigation response in AV and environmental knowledge of the vehicle are based on current observation, and it is comparatively less extraction and complexity. Hence, the distinct AV information for the navigation is reduced and satisfies a high learning rate due to changes in the sharing information series (Figure 6). This proposed framework improves the learning rate by 9.76% and 8.68%, respectively, for the above comparisons. Following the above illustrations, the environment detection ratio is presented in Figure 7.

Based on data extraction and utilisation, the proposed framework achieves high environment detection for safe and unsafe driving categorisation (Figure 7). The threat occurrence is mitigated based on $\arg \begin{array}{c}max\\ i\end{array}\left({\int }_T^{T+1}P\left(X_i,Y_i\right).T dT\right)$ with connected neurons. Information extraction and sharing are pursued to identify unsafe driving due to changes in location and lane, and the vehicle’s visual observations and environmental knowledge are categorised through neural learning. Maximum extraction and utilisation are useful for precise environmental detection to reduce the AV’s unsafe activity. The interactions are batched and addressed the unknown scenarios throughout the driving time based on both observations of ${\mathbb{V}}^f>{\mathbb{V}}^l$ and ${\mathbb{V}}^f\le {\mathbb{V}}^l$ and transmission of organising information through weight assignment. Contrarily, the computed unsafe environmental data observation is used to augment the unidentified scenarios for verifying driving and pedestrian safety, which relies on other factors in the point-to-point interaction. Therefore, the environmental detection is high, and batching also increases. The proposed COI framework improves environment detection by 9.65% and 10.97% for different times and distances. Figure 8 illustrates the navigation assistance rate for the distance and time variants.

The navigation assistance through neuron connectivity ensures consistency and adaptability with the interacting “Thing” for accurate data observation in the information-sharing process (Figure 8). This framework satisfies less learning categorisation by computing the maximum extraction and utilisation. In this sequential process, information actuated and sensed from the AV in different time intervals for preventing unsafe driving based on the condition $X=\left\{\delta ,sdi,\omega ,{Risk}_a\right\}$ and $\mathsf{\Psi }=\left\{safe\left(\mathbb{S}\right),unsafe\left(\mathbb{U}\right)\right\}$ is computed until driving time to retain road safety. The observed vehicle resilience and consistency depend on the position vector and their threshold conditions, wherein the regularisation and uncertainty in the environmental knowledge of the vehicle are identified through neural learning. The unsafe driving behaviour is identified using the observation of information between the AV and interacting “Thing” to compute retaining accuracy based on driving time. The unsafe navigation is addressed with the extracted data and retaining this condition with $p_v\left(\left.X=\mathbb{x}\right|\mathsf{\Psi }={\mathsf{\psi }}_i\right)$ computation. Therefore, the maximum environmental knowledge data extraction is computed to maximise the interactions based on actuated and sensed information from the AV, which satisfies high navigation assistance. The maximum percentage of improvement is 7.93% and 8.1% for the time and distance.

The comparative illustrations of information latency for distance and time are in Figure 9.

The accurate information transmission between the AV and interacting “Thing” through the learning process classifies safe and unsafe navigation responses. The observed environmental information is extracted from the previous neuro batches to ensure highly safe navigation. The available data extraction and environment detection accuracy are considered to identify the risk factor level. If identifying the unsafe activity of the vehicle in any instance, the risk factors originate from the information actuated and sensed for accurate decision-making. Reliable and persistent information sharing is based on the environmental knowledge of the vehicle to be observed, and the risk level can be detected through the batch optimisation algorithm and neural learning process. This detection and batch optimisation accuracy is increased by streamlining its previous neuro batches for filtering safe navigation responses, preventing information latency. The extracted features are extended to the vehicle decision over unavailable information-based environmental data observation, for which the proposed framework satisfies less information latency (Figure 9). The proposed framework suppresses the latency by 10.22% and 10.58% for the time and distance from the above comparative analysis.

5. Discussion

The choices of ORRL, HDSE-DRL, and CRL-MPC as comparison algorithms for the COIF show their utility in increasing autonomous vehicle (AV) safety, flexibility, and decision-making under uncertain conditions.

(i)

COIF’s primary goal is the same as ORRL’s: to ensure AV decisions can withstand unpredictable environments. The overarching objective of ORRL and COIF is to enhance AV performance in novel environments. However, COIF enhances information-sharing consistency and improves control over safety responses by communicating with external “Things” in the environment in real time via neuro-batch learning.

(ii)

HDSE-DRL was selected for its strategic flexibility in uncertain highway driving, aligning with COIF’s emphasis on resilience. The versatility is further enhanced by COIF’s ability to classify and batch various types of data (such as traffic density, objects, and obstacles), which allows for more detailed environmental information and the real-time detection of harmful activities.

(iii)

CRL-MPC is included, which enhances vehicle navigation via predictive control and focuses on managing AV behaviours during car-following manoeuvres. Using CRL-MPC’s predicted performance as a yardstick, one may assess COIF’s real-time adaptability and decision consistency. While CRL-MPC uses reinforcement learning and model predictive control to anticipate how vehicles will behave and ensure safe following distances, COIF uses adaptive neuro-batch learning to enhance real-time responses to various road circumstances.

As a result, COIF can maximise safety by continually analysing hazards, keeping an eye on them, and correcting for small changes in vehicle choices. In contrast to these proven methods, COIF may adaptively enhance vehicle-pedestrian safety through consistent, robust, and classified interaction, and it takes a novel approach to improve AV control.

Indeed, the findings of the comparison study are encouraging, demonstrating that the suggested strategy outperforms current methods. Nevertheless, it would be much more beneficial to provide an exact estimate of the obtained accuracy since this would provide a clear and measurable way to evaluate the development. Such an absolute accuracy number might provide a precise measure of this method’s effect on autonomous vehicle operations’ performance, dependability, and safety. The method’s benefits would be better supported, and its efficacy might be more easily assessed with this information included if it is feasible.

6. Conclusions

This article proposes and discusses the control optimiser interaction framework for autonomous vehicle driving and navigation safety. The AV control optimisation with environmental adaptability through intense learning is designed to improve its performance. Based on the neuro-batch optimisation method, the communication between the roadside intelligent communication “Things” is used to improve the vehicles’ adaptability and environmental knowledge gain. The interaction extracts, computes, and utilises data through batches with different additive or depreciative information exchanged. This included the traffic information, pedestrian density, hindrances, etc., in the driving path of the vehicles. The batches are connected through neurons and are recurrently trained for any new information acquired throughout the interaction. Based on the risk and driving parameters, the safe and unsafe activity of the vehicles is categorised with a precise learning rate. Therefore, minor changes in vehicular decisions are monitored, and driving control is optimised accordingly to retain 7.93% of navigation assistance through a 9.76% high learning rate for different intervals.