Situation-Aware Causal Inference-Driven Vehicle Lane-Changing Decision-Making

Featured Application

Enhancing vehicle lane-changing safety and efficiency within intelligent transportation systems. This model significantly reduces the collision risk by 63% and lane-changing time by 31.9% using a dynamic safe distance, attention mechanisms, and counterfactual reasoning. Optimized with particle swarm methods, it improves traffic flow and safety simultaneously.

Abstract

For the decision-making challenge of ensuring vehicle lane-changing safety, this study proposes a context-dependent causal inference-based model for safe lane changes. Emphasizing multi-vehicle interactions within dynamic traffic scenarios, we construct a three-layer decision-making framework that relies on real-time data collection of speed, acceleration, and spacing information from both the target vehicle and adjacent-lane vehicles. The framework consists of (1) a context-aware layer that extracts standardized dynamic features; (2) an attention mechanism layer that dynamically assigns weights to critical risk factors; and (3) a counterfactual causal reasoning layer where lane-changing risks are quantified through virtual interventions, with multi-objective safety strategies optimized via particle swarm algorithms. The simulation results indicate significant enhancements in high-density traffic conditions. When compared to traditional safety distance models and built-in models from simulation software (SUMO v1.18.0), the proposed model achieves reductions in average conflict counts by 63.0% (from 12.7 to 4.7 instances) and by 37.3% (from 7.5 to 4.7 instances), respectively. Additionally, lane-changing durations are reduced by 10.9% (from 5.5 to 4.9 s) and by 31.9% (from 7.2 to 4.9 s), while fluctuations in risk values decrease by 53.3% (from 0.75 to 0.35) and by 36.4% (from 0.55 to 0.35), respectively. The experimental validation confirms that the integration of dynamic safety distance computation with causal reasoning significantly enhances decision-making robustness in complex scenarios through coordinated risk quantification and multi-objective optimization

1. Introduction

In the contemporary traffic environment, the complexity of driving behavior is escalating daily, particularly during lane changes. Drivers are required to simultaneously monitor traffic conditions in front, behind, and alongside their vehicles while making timely decisions based on multi-dimensional information inputs. When executing a lane change, the onboard computer system must accurately assess the surrounding traffic environment and make informed decisions based on both environmental states and its own motion dynamics. This includes determining whether to initiate a lane change and how to execute it effectively. However, current autonomous lane-changing capabilities of vehicles remain relatively limited, especially regarding enhancements in decision-making processes and execution strategies for lane changes. This area necessitates further research for improvement.

Over the past two years, motor vehicle accidents resulting from lane changes have accounted for 16.9% of total motor vehicle accidents [1]. This statistic underscores an urgent need for effective lane change assistance systems. Nevertheless, most existing lane change strategies and trajectory models rely on static planning approaches that lack timely responsiveness; as a result, they struggle to adapt to real-world driving environments. Such conservative strategies expose drivers to heightened safety risks during actual operations. Therefore, developing technology that can effectively assist drivers in making safe and optimal lane change decisions is of paramount importance. Research into safe and optimal lane change (LC) technology is crucial not only for advancing traffic flow theory but also for enhancing autonomous driving capabilities—impacting safety, throughput efficiency, and congestion management significantly. While notable progress has been made in this field, achieving real-time robust decision-making within highly dynamic environments remains a formidable challenge.

1.1. Decision-Making Models

Decision-making models serve as the fundamental logic for initiating and executing lane changes (LCs). Peng [2] (2020) pioneered a safety-and-ergonomics trajectory decision mechanism, providing a novel redefinition of safety considerations. However, the real-time performance of this approach in complex scenarios, along with its computational demands, necessitates further validation. Building upon safety concerns while emphasizing interaction, Yao [3] (2021) introduced a leading-vehicle-intention-based decision-making model that utilizes lateral offset and constraints. This model has demonstrated effectiveness in facilitating efficient vehicle-associated LC maneuvers. Nevertheless, Yao’s approach falls short in addressing more intricate lateral interactions and multi-vehicle dynamics, while also imposing significant requirements on sensor processing speed. In a broader context, Ma [4] (2023) presented a comprehensive systematic review of LC research, synthesizing the evolution and scope of the field.

1.2. Behavior Modeling and Prediction

Complementing decision-making processes, research in Behavior Modeling and Prediction emphasizes the understanding and anticipation of driver and vehicle actions to facilitate safer decisions. Zhang [5] (2019) utilized Long Short-Term Memory (LSTM) networks trained on Human–Robot Collaboration (HRC) data to simultaneously model car following (CF) and lane change (LC) behaviors. Shifting the focus towards risk assessment, Wu [6] (2020) developed the TSRE model specifically for real-time identification of LC risks. For predictive tasks, Li [7] (2020) employed Support Vector Regression (SVR) to forecast lane change initiation (LCI), utilizing microscopic traffic variables. Additionally, Das A [8] (2020) proposed a trajectory-level LC detection model that leveraged data from Naturalistic Driving Studies (NDSs) and the Roadway Information Database (RID). More recently, Guo J [9] (2022), by integrating game theory with advanced neural networks, introduced a cellular-based game-theoretic LC model. Furthermore, Guo [10] (2022) and Gao [11] (2024) established connected vehicle (CV)-based data-driven models for LC detection and prediction using attention mechanisms, which demonstrated enhanced accuracy and robustness compared to previous methodologies.

1.3. Safety and Trajectory Planning

Ultimately, the objectives of lane change (LC) research frequently focus on enhancing safety and trajectory planning to mitigate accidents and alleviate congestion. Within safety frameworks, Yang [12] (2018) proposed a Dynamic Lane-Trajectory Planning (DLTP) method that incorporates constraints for rollover and collision avoidance. Similarly emphasizing safety thresholds, Wang [13] (2020) developed a decision-making framework that integrates safety thresholds for approaching vehicles. To address uncertainties in safety factors, Luo [14] (2021) introduced fuzzy logic models. A prevalent approach for assessing safety involves distance metrics alongside trajectory planning; for example, Li [15] proposed potential field-based safe distance models. In direct response to collision avoidance challenges, He [16] employed simulation techniques to establish lateral acceleration profiles and separation criteria. Concurrently, various trajectory planning techniques—such as polynomial curves, Bézier curves, and Model Predictive Control (MPC)—are dedicated to generating feasible paths that are free from collisions. Liu [17] synthesized multiple path-planning paradigms specifically tailored for automated LC systems, underscoring the continuous advancements in this domain. Demonstrating a cutting-edge development in the field, Wu [6] (2020) introduced collaborative evolutionary trajectory planning aimed at minimizing overall risk through cooperative maneuvers.

In summary, despite the positive progress achieved in existing research, there are still significant deficiencies that need to be systematically addressed:

• Real-Time Decision-Making Capabilities: Current models struggle to meet the millisecond-level decision-making requirements of highly dynamic and strongly interactive scenarios. The rapid and complex nature of real-world driving conditions demands real-time decision-making capabilities that current models have not fully achieved.
• Multi-Dimensional Information Integration: There is a lack of effective mechanisms for efficiently integrating multi-dimensional information, particularly video streams from electronic rearview mirrors, to achieve in-depth environmental understanding and causal inference. The complexity of integrating diverse data sources and extracting meaningful insights remains a significant hurdle.
• Optimal Solution Under Multi-Objective Games: Finding the optimal solution rapidly under complex multi-objective games is still a formidable challenge. The interaction between multiple objectives and the dynamic nature of the driving environment makes it difficult to find the best course of action in a timely manner.

1.4. Electronic Rearview Mirror Technology

Electronic rearview mirrors (Blind Spot Detection—BSD systems) enhance lane change safety by monitoring vehicle surroundings using cameras and sensors. Li [7] (2020) proposed an “Uncertainty Shadow Safety” method for quantitative multi-directional risk assessment and warning. Modern BSD aims to significantly reduce accidents, warn drivers of hazards, and potentially initiate corrective actions. Statistics indicate that road blind zones contribute to ~30% of ~500,000 annual traffic accidents in China. Consequently, Wu [18] (2023) developed an active collision avoidance method for right-turn blind zones using an improved YOLOv4-tiny algorithm for VRU detection and warning based on distance/speed. Kim [19] (2023) focused on signal processing, tracking algorithms, and AI-based design for high-precision BSD radar systems. Muzammel [20] (2022) integrated advanced CNN feature descriptors with Faster R-CNN for heavy vehicle blind spot collision detection.

Electronic rearview mirror technology has been gradually becoming a standard feature in vehicles, leveraging its advantages of expanding the field of vision, eliminating blind spots, and enhancing perception in harsh environments. In regions such as Japan, the European Union, and the United Kingdom, legislation has been introduced to promote its adoption [21,22]. This development presents new opportunities for addressing the aforementioned challenges, especially in making efficient use of high-dimensional perception information.

Against that backdrop, this paper utilizes video monitoring data and vehicle status data from vehicle electronic rearview mirrors. By constructing a vehicle lane-changing safety model, it conducts a thorough analysis of the vehicle lane-changing mechanism under the condition of electronic rearview mirrors. Moreover, by introducing attention mechanisms, counterfactual analysis [23], and structural causal models, this paper constructs a vehicle lane-changing decision-making model based on context-dependent causal inference. Particle Swarm Optimization (PSO) [24,25,26] is employed to optimize the solution of the lane-changing strategy. This model features a three-layer architecture encompassing dynamic perception of context features, causal graph reasoning, and multi-objective game decision-making. Integrating the causal mechanism into the lane-changing decision-making loop enables efficient and safe vehicle lane-changing decisions.

The innovative content of this article primarily encompasses the following two points:

• A lane-changing safety and decision-making model is proposed, which integrates multimodal perception information along with context-related features. This research introduces electronic rearview mirror technology to acquire high-dimensional environmental data and employs an attention mechanism to enhance the extraction of critical scene features. Through dynamic perception and behavioral correlation analysis, this model can more accurately identify real-time risk factors in complex traffic environments, thereby improving decision-making accuracy while optimizing the selection of lane-changing strategies.
• The introduction of causal relationship modeling into the domain of lane-changing research represents a significant innovation. By constructing a structured causal graph and implementing a counterfactual analysis framework, this study elucidates the causal relationships present in multi-objective games across various driving scenarios. This approach not only facilitates a clearer understanding of the dynamic influence relationships among different factors during lane changes but also provides theoretical support and technical foundations for real-time safety assessments and optimal strategy optimization based on causal mechanisms.

2. Description of the Scenario

Lane-changing behavior is a complex operation in driving, requiring drivers to consider multiple factors when executing lane changes. This behavior not only involves assessing the current status of the vehicle (such as speed and distance [27]) but also necessitates real-time evaluation of dynamic information regarding surrounding vehicles (including their speed and position in adjacent lanes). Furthermore, lane-changing actions can significantly impact other drivers in both the original and target lanes, thereby increasing the risk of traffic conflicts or accidents. Consequently, ensuring safe lane changes demands advanced technical means for real-time perception and analysis of the operational status of multiple vehicles.

As an advanced driving assistance device, the electronic rearview mirror enhances collected visual data through image-processing algorithms and transmits this information to the vehicle’s display screen, providing drivers with clear and stable visual feedback. By integrating real-time data on target vehicles along with environmental information, the electronic rearview mirror accurately perceives the movement statuses of surrounding vehicles. This capability offers a scientific basis for decision-making regarding lane changes for both drivers and autonomous driving systems.

In this context, the present study aims to utilize data from electronic rearview mirrors to develop a causal relationship model for analyzing and forecasting lane-changing behavior in terms of safety. The goal is to provide decision support for automated vehicle auxiliary lane changing. In real-world driving scenarios, the interactions between the target vehicle and other vehicles—particularly those in adjacent lanes, such as leading and trailing vehicles—are both dynamic and complex. To design a comprehensive safe driving model, it is essential to consider the relative motion states of the target vehicle with respect to both the leading and following vehicles, as well as factors related to road conditions and operational constraints.

(1)

Define input parameters

As shown in Figure 1, Target vehicle(indicated in black color): speed, acceleration, reaction time, maximum acceleration, maximum braking acceleration; $v$ $a$ $t_{react}$ $a_{\max }$ $b_{\max }$

Preceding vehicle in adjacent lane: distance between preceding vehicles, velocity, acceleration; $d_f$ $v_f$ $a_f$

Next vehicle in adjacent lane: rear distance, velocity, acceleration; $d_r$ $v_r$ $a_r$

Road limit: maximum speed limit. $v_{\max }$

(2)

Calculate the basic safe distance

The basic safe distance is defined as the distance the driver needs to travel until they detect the problem and safely change lanes to protect the vehicle. This distance is related to the operation state of the vehicle in front of the adjacent lane, and its mathematical description is as follows.

$$d_{base}=v\cdot t_{react}+{\displaystyle \frac{v^2-v_f^2}{2\max \left(b_{\max }+a_f,\epsilon \right)}}$$

(1)

where $\epsilon$ is an infinitesimal parameter, and in this paper, is ${10}^{-9}$. This prevents division by zero in edge cases, a standard practice in numerical stability [28].

(3)

Calculate the dynamic safe distance

Considering that in the process of lane changing, there is a dynamic interaction between the safe distance of the target vehicle and the speed, and another between acceleration and deceleration states of the front and back vehicles in the adjacent lane, in order to further quantify the dynamic safe distance, the dynamic adjustment coefficient is introduced as follows:

$$\begin{array}{l}{k}_f=1+{\displaystyle \frac{\max \left(v-v_f,0\right)}{v}}+{\displaystyle \frac{\max \left(a-a_f,0\right)}{b_{\max }}}\\ k_r=1+{\displaystyle \frac{\max \left(v_r-v,0\right)}{v}}+{\displaystyle \frac{\max \left(a_r-a,0\right)}{b_{\max }}}\end{array}$$

(2)

The dynamic adjustment coefficients $k_f,k_r$ quantify risk escalation during lane changes. $k_f>1$ indicates a higher risk when the subject vehicle approaches a slower preceding vehicle ($v>v_f$) or decelerates less abruptly ($a>a_f$). The normalization by $v$ and $b_{\max }$ ensures scale-invariant risk assessment, aligning with vehicle dynamics principles [29].

Therefore, it can be determined that the dynamic safe distance between the front and rear vehicles in adjacent lanes is as follows.

$$\begin{array}{l}{d}_{safe,f}=d_{base}\cdot k_f\\ d_{safe,r}=d_{base}\cdot k_r\end{array}$$

(3)

3. Safe Lane-Changing Model Based on Context-Dependent Causal Inference

This model aims to realize a safe and efficient lane-changing decision by integrating context awareness, attention mechanism, and causal reasoning technology. The overall framework can be divided into the following three main parts: Context awareness layer: collects real-time dynamic features related to lane changing. Attention mechanism: dynamically adjusts the attention degree of each feature to highlight the factors that have the greatest impact on the lane-changing decision. Counterfactual causal inference layer: enhances the ability of risk prediction and evaluates the causal impact in different situations.

3.1. Emotion Perception Layer

The objective of this layer is to capture real-time dynamic features that are directly associated with lane changing. In light of the aforementioned introduction, and in order to eliminate dimensional discrepancies while enhancing the model’s convergence efficiency, the input state vector for this layer is defined as follows, in accordance with standardization requirements:

$$S=\left[{\displaystyle \frac{v}{v_{\max }}},{\displaystyle \frac{a}{a_{\max }}},t_{react},b_{\max },{\displaystyle \frac{v_f-v}{v_{\max }}},{\displaystyle \frac{v_r-v}{v_{\max }}},{\displaystyle \frac{a_f-a_r}{a_{\max }}}\right]$$

(4)

3.2. Attention Mechanism Layer

The goal of this layer is to dynamically adjust the attention level of each feature [28].

(1)

Query vector, key vector, and value vector are obtained by linear transformation

When combining Equation (4) and applying a different linear transformation to each element of the same input sequence, the query vector, key vector, and value vector for this article are determined as follows:

$$\begin{array}{l}Q=W_Q\cdot S^T=\left[\begin{array}{ccc}{\omega }_{11}^{\left(Q\right)}& \dots & {\omega }_{17}^{\left(Q\right)}\\ ⋮& \ddots & ⋮\\ {\omega }_{71}^{\left(Q\right)}& \dots & {\omega }_{77}^{\left(Q\right)}\end{array}\right]\cdot S^T\\ K=W_K\cdot S^T=\left[\begin{array}{ccc}{\omega }_{11}^{\left(K\right)}& \dots & {\omega }_{17}^{\left(K\right)}\\ ⋮& \ddots & ⋮\\ {\omega }_{71}^{\left(K\right)}& \dots & {\omega }_{77}^{\left(K\right)}\end{array}\right]\cdot S^T\\ V=W_V\cdot S^T=\left[\begin{array}{ccc}{\omega }_{11}^{\left(V\right)}& \dots & {\omega }_{17}^{\left(V\right)}\\ ⋮& \ddots & ⋮\\ {\omega }_{71}^{\left(V\right)}& \dots & {\omega }_{77}^{\left(V\right)}\end{array}\right]\cdot S^T\end{array}$$

(5)

where $W_Q,W_K,W_V$ is the learnable parameter matrix, $\omega$ is the weight, and the dimension is 7.

(2)

Calculate the attention score matrix

$$a_{ij}=soft\max \left({\displaystyle \frac{{Q_i{K}_j}^T}{\sqrt{d}}}\right)$$

(6)

where $d=7$ is the feature dimension.

(3)

Solve the context vector $c_t$

Through the dynamic weighted fusion of input features, the most critical feature combination for the current decision is extracted. Its role can be compared to the cognitive process of human drivers who actively focus on key risk points in complex road conditions.

$$c_t={\displaystyle \sum _{i=j}^7{a}_{ij}}{V}_j$$

(7)

where $c_t\in {ℝ}^7$ contains the weighted key features ($\Delta d_f,\Delta d_r,\Delta v,\Delta a,\Delta v_f,\Delta v_r,\delta$).

3.3. Counterfactual Causal Inference Layer

(1)

Construct a structural causal model

As shown in Figure 2, the causal relationship [29,30] between variables is formally expressed through the causal diagram, which provides an interpretable logical framework for risk assessment. The critical path is introduced as follows:

①

Main causal path of target vehicle:

$$\left(v,a,t_{react},b_{\max }\right)\to D_{dynamic}\to R$$

(8)

②

Leading vehicle influence path:

$$\left(v_f,a_f\right)\to k_f\to D_{dynamic}$$

(9)

③

The following vehicle affects the path:

$$\left(v_r,a_r\right)\to k_r\to D_{dynamic}$$

(10)

④

Counterfactual intervention path:

$$do\left(\Delta a\right)\to a^{\prime }\to v^{\prime }\to D_{dynamic}^{\prime }$$

(11)

where $D_{dynamic}^{\prime },D_{dynamic}$ are the old and new dynamic safety distances, $R$ is the quantitative risk, and $\Delta a,a^{\prime },v^{\prime }$ are the acceleration adjustment amount, the adjusted acceleration, and the adjusted speed of the target vehicle counterfactual intervention.

(2)

Counterfactual intervention construction

This part simulates the safety state under different decisions through virtual intervention. In this paper, the target vehicle acceleration is selected as the intervention object.

①

Calculation of original safety distance:

$$D_{dynamic}=\max \left(d_{safe,f},d_{safe,r}\right)+L$$

(12)

where $L$ is the length of the car body.

②

Counterfactual intervention operation:

This part simulates the safety state under different driving strategies by virtually modifying the acceleration of the target vehicle. $do\left(a^{\prime }=a+\Delta a\right)$, meaning the updated speed of the target vehicle can be obtained as $v^{\prime }=v+a^{\prime }\cdot \Delta t$, where $\Delta t$ is the lane-changing time.

Then, the $d_{base}^{\prime }$ updated basic safety distance, the $k_f^{\prime },k_r^{\prime }$ dynamic adjustment coefficient of the front and rear vehicles, and the dynamic safety distance of the front and rear vehicles $d_{safe,f}^{\prime },d_{safe,r}^{\prime }$ can be obtained according to Equations (1)–(3).

Therefore, according to Formula (12), the safe distance after intervention can be obtained as follows.

$$D_{dynamic}^{\prime }=\max \left(d_{safe,f}^{\prime },d_{safe,r}^{\prime }\right)+L$$

(13)

③

Quantitative assessment of lane change risk:

Considering the distance, speed, acceleration, and other factors, combined with normalization processing, the quantitative assessment model of lane-changing risk was finally deduced and constructed as follows.

$$R=\exp \left(-{\displaystyle \frac{\Delta d_f}{L}}\right)\cdot \exp \left(-{\displaystyle \frac{\Delta d_r}{L}}\right)\cdot \exp \left(-{\displaystyle \frac{\Delta {v_f}^2+\Delta {v_r}^2}{\delta \cdot v}}\right)$$

(14)

where the definitions $\Delta d_f=\left|d_f-d_{safe,f}\right|,\Delta d_r=\left|d_r-d_{safe,r}\right|$ represent the difference between the front and rear vehicles and the dynamic safety distance between the front and rear vehicles, respectively; $\Delta v_f=\left|v-v_f\right|,\Delta v_r=\left|v-v_r\right|$ are defined as the speed difference between the target vehicle and the vehicles ahead and behind in the adjacent lane; and $\delta$ is the weight coefficient, which can be obtained according to historical data and a simulation test.

4. Optimization of Lane-Changing Strategy Based on Particle Swarm Optimization

4.1. Objective Function Construction

In addition to the safety factors, driving efficiency, control cost, and other factors should be further considered in the vehicle lane-changing process. Therefore, in order to optimize the vehicle lane-changing process, the lane-changing decision is modeled as a multi-objective optimization problem, so the objective function can be defined as

$$J={\lambda }_{1}R+{\lambda }_2\Delta v^{adj}+{\lambda }_3\left|\Delta a^{adj}\right|+{\lambda }_4{\displaystyle \int \left|j\right|dt}$$

(15)

where ${\lambda }_1\sim {\lambda }_4$ is the weight coefficient of each factor, $\Delta v^{adj}$ is the target vehicle speed adjustment, $\Delta a^{adj}$ is the target vehicle acceleration adjustment, and $j$ is the target vehicle jerk degree (describes the physical quantity of the vehicle acceleration change rate, measures the degree of acceleration change in unit time, and is the derivative of acceleration with respect to time).

Formula (15) shows that the target of vehicle lane change mainly has four contents, which are the comprehensive risk, speed adjustment, acceleration adjustment, and jerk degree.

In addition to considering the above factors, the vehicle should also meet the safety constraints of vehicle driving in the process of lane changing, as follows:

①

Safety distance constraint

$$\begin{array}{l}{d}_f\ge d_{safe,f}\\ d_r\ge d_{safe,r}\end{array}$$

(16)

②

Speed constraint

$$v+\Delta a\cdot t_{react}\le v_{\max }$$

(17)

4.2. Particle Swarm Optimization Algorithm Optimization

Combined with the rapid decision-making and safety-first optimization needs, this paper chooses Particle Swarm Optimization (PSO) as the solution algorithm, where the specific content is as follows:

①

Particle encoding and initialization: Each particle $x_i=\left[\Delta a^{adj},\Delta v^{adj}\right]$ is defined to represent a set of candidate solutions, the particle velocity is initialized to random values, and the feasible region is defined as $N$.

②

Fitness function design: The objective function is shown in Formula (15), and to achieve the optimization purpose $J_{\min }$, the penalty amount $\rho$ is introduced into it, and the penalty coefficient is

$$J_{penalty}=J+\rho \left(\max \left(0,d_{safe,f}-d_f\right)+\max \left(0,d_{safe,r}-d_r\right)\right)$$

(18)

③

Particle update rules: This part mainly includes two parts, which are the velocity update and position update.

$$\left\{\begin{array}{l}{v}_i^{k+1}=\xi v_i^k+c_1{r}_1\left(p_{best,i}-x_i^k\right)+c_2{r}_2\left(g_{best,i}-x_i^k\right)\\ x_i^{k+1}=x_i^k+v_i^{k+1}\end{array}\right.$$

(19)

where $\xi$ is the inertia coefficient, $c_1,c_2$ are the individual and group learning factors, and $r_1,r_2\in \left(0,1\right)$ is a random number.

The convergence behavior and parameter sensitivity of the PSO algorithm are analyzed in Figure 3. For high-density scenarios, the optimization consistently converges within 47 iterations on average, demonstrating computational efficiency suitable for real-time decision-making. Sensitivity analysis (Figure 3b) confirms that safety weight λ₁ dominates the collision avoidance performance, causing an 18.3% variation in conflict count when perturbed by ±20%, while the comfort weight λ₄ shows minimal impact (<5%). This parameter robustness ensures a stable performance across varying traffic conditions.

5. Verify the Analysis Experimentally

5.1. Experimental Scene Construction

Experiments used SUMO v1.18.0 with a simulation timestep of 0.1 s. Vehicle types: 85% passenger cars (length = 4.5 m), 15% trucks (length = 7.2 m). Random seeds 0–9 were used for 10 independent runs, with the results averaged (±std). Route files were generated via randomTrips.py with flow rates 800/1600/2400 veh/h for low/medium/high density.

To verify the effectiveness of the model presented in this paper, a road segment simulation scenario was constructed using the Sumo platform to replicate a one-way, three-lane driving environment. The specific parameters were as follows:

Road Environment Data: The length of the straight lane is 1 km, with a lane width of 3.5 m and a speed limit set at 70 km/h. The target vehicle was initially positioned at the start of the middle lane, with an initial speed randomly assigned within the range of 35–60 km/h. Leading and following vehicles were randomly distributed in adjacent lanes (left and right), maintaining an initial distance ranging from 20 to 50 m, with speeds also falling within the range of 35–60 km/h (following a normal distribution).

Vehicle State Data: The maximum acceleration for the target vehicle is specified as 2.5 m/s², while its maximum braking deceleration is −4.5 m/s²; the reaction time varies between 0.8 and 1.2 s (randomly assigned). In adjacent lanes, both front and rear vehicles exhibit an acceleration range of ±1.5 m/s², with an initial spacing error margin set at ±5%.

Vehicle Perception Data: Through secondary software development, real-time data regarding speed, acceleration, and spacing information for following vehicles in adjacent lanes relative to the target vehicle were collected instantaneously. Additionally, real-time metrics concerning speed, acceleration, and spacing for leading vehicles in these adjacent lanes were obtained to simulate CAN collection data pertinent to the target vehicle (such as that from driving recorders or laser radars). Furthermore, real-time speed and acceleration information pertaining directly to the target vehicle was acquired similarly for CAN simulation purposes. Data acquisition occurred at a frequency of 10 Hz; Gaussian-distributed interference data with a mean value of zero and standard deviation set at five percent was employed to simulate noise effects.

The Gaussian noise perturbation (±5%) applied to perceived positions and velocities reflects empirical sensor error characteristics in automotive systems. This noise level corresponds to industry-reported millimeter-wave radar velocity errors (3–5%) in BSD systems under typical operating conditions [19]. Each simulation experiment involved 50 vehicles (including the target vehicle), with vehicle quantities dynamically adjusted according to traffic density levels defined in

Table 1. Traffic density parameters.

Density Level	Vehicles per km	Inter-Vehicle Gap	Speed Difference
Low	<15	>50 m	<10 km/h
Medium	15–30	30–50 m	10–20 km/h
High	>30	<30 m	>20 km/h

.

5.2. Simulation Experiment Setup

For each traffic density level (Low/Medium/High), 10 independent simulation runs were conducted using random seeds (0–9) to ensure statistical consistency. Results are reported as mean ± standard deviation.

According to the complexity of traffic scenarios, three distinct density levels—high, medium, and low—have been established. These varying density scenes are characterized by the distance between vehicles and the speed differentials, as detailed below:

⮚

Low-density scenarios: The distance between vehicles in adjacent lanes exceeds 50 m, with a speed difference of less than 10 km/h between leading and trailing vehicles.

⮚

Medium-density scenario: The distance between vehicles in adjacent lanes ranges from 30 to 50 m, while the speed difference between front and rear vehicles is between 10 and 20 km/h.

⮚

High-density scenario: The distance between vehicles in adjacent lanes is less than 30 m, accompanied by a speed differential exceeding 20 km/h.

Within each density scenario, three experimental groups were established: Experimental Group 1 utilized a basic safety distance model (Equation (1)), Experimental Group 2 employed the lane-changing model from simulation software (LC2013 model), and Experimental Group 3 implemented a context-dependent causal lane-changing model.

During experimentation, the lane-changing intention of the target vehicle was triggered randomly at intervals of every thirty seconds. The feature dimension was set at d = 7 with an individual vehicle length of L = 4.5 m. A weight coefficient δ = 0.5 was determined through simulation calibration and historical data fitting. Following multi-objective optimization principles, safety was prioritized while balancing efficiency and control costs; comfort considerations were subsequently integrated into this framework. Factor weight coefficients were assigned as follows: λ₁ = 0.5, λ₂ = 0.2, λ₃ = 0.2, λ₄ = 0.1 based on PSO parameter sensitivity analysis results. The parameters for the PSO algorithm are configured as follows: number of particles N = 50; number of iterations T = 100; inertia coefficient ξ = 0.6; learning factors c₁ = c₂ = 1.5; penalty coefficient ρ = 1000.

For high-density scenarios, the optimization consistently converges within 47 iterations on average, demonstrating computational efficiency suitable for real-time decision-making. Sensitivity analysis (Figure 3b) confirms that safety weight λ₁ dominates collision avoidance performance, causing an 18.3% variation in conflict count when perturbed by ±20%, while the comfort weight λ₄ shows minimal impact (<5%). This parameter robustness ensures a stable performance across varying traffic conditions.

5.3. Analysis of Experimental Results

As illustrated in Figure 4, under varying traffic density scenarios, the number of collisions recorded in Experiment 3 (contextual causal model) was significantly lower than those observed in Experiment 1 (basic safety distance model) and Experiment 2 (built-in model of simulation software). In the low-density scenario, the average number of collisions in Experiment 3 was 0.3 per ten trials, representing a reduction of 62.5% compared to Experiment 2 (0.8 times) and an impressive decrease of 85% relative to Experiment 1 (2.0 times).

In the high-density scenario, the conflict count for Experiment 3 stood at 4.7, which is a notable decline of 37.3% from that recorded in Experiment 2 (7.5 times) and a substantial reduction of 63.0% when compared to Experiment 1 (12.7 times). In the medium-density scenario, the conflict occurrences in Experiment 3 were reduced by approximately 53.6% and by as much as 74.5% when contrasted with Experiments 2 and 1, respectively.

Through context-aware analysis and counterfactual causal reasoning, Experiment 3 demonstrates a significant enhancement in conflict avoidance capabilities within complex environments—particularly evident in scenarios characterized by dense traffic flow.

Discussion on Conflict Reduction: The substantial reduction in conflict counts, particularly evident in high-density scenarios (Figure 4), can be attributed to the synergistic effect of the model’s core components. The context-aware layer provides standardized, real-time inputs reflecting the dynamic interplay between vehicles. The attention mechanism dynamically identifies and weights the most critical risk factors (e.g., large relative speed differences $\Delta v_f,\Delta v_r$ or critically small gaps $\Delta d_f,\Delta d_r$), a capability absent from the static rules of Exp1 and the less adaptive weighting potentially present in Exp2. Crucially, the counterfactual causal inference layer allows the model to simulate ‘what-if’ scenarios (e.g., “What if I accelerate slightly more/less?”) before executing a lane change. This enables proactive risk quantification (R in Equation (15)) based on predicted future states (D’dynamic in Equation (14)), rather than relying solely on instantaneous or rule-based thresholds like Exp1 and Exp2. This predictive risk assessment is paramount in preventing collisions in rapidly evolving, high-density traffic where the reaction time is critical.

As illustrated in Figure 5, Experiment 3 demonstrates a superior lane-changing efficiency. Its time consumption consistently remains lower than that of Experiment 2 and is significantly better than that of Experiment 1 in high-density scenarios. In low-density conditions, the average time for Experiment 3 was recorded at 2.9 s, which represents a 9.4% increase compared to Experiment 1 (3.2 s) and a 3.3% increase relative to Experiment 2 (3.0 s).

In high-density situations, Experiment 3 required only 4.9 s, marking a reduction of 10.9% from the time taken by Experiment 1 (5.5 s) and an impressive decrease of 31.9% compared to Experiment 2 (7.2 s). For medium-density scenarios, the duration for Experiment 3 was measured at just 3.8 s—24% less than the corresponding time for Experiment 2 (5.0 s).

Experiment 3 effectively balanced safety and efficiency through dynamic safe distance calculations combined with a Particle Swarm Optimization strategy, thereby mitigating the surge in time consumption typically associated with conservative strategies employed by traditional models in high-density environments.

Discussion on Efficiency Gains: The significant reduction in lane-changing time, especially the 31.9% improvement over Exp2 in high-density scenarios (Figure 4), underscores the effectiveness of the multi-objective PSO optimization. While Exp1 prioritizes safety conservatively (leading to longer, slower maneuvers) and Exp2 might struggle to find efficient paths under complex constraints, our model explicitly balances safety (minimizing R), efficiency (minimizing time, linked to $\Delta v_{adj}$), and comfort (penalizing large $\Delta a_{adj}$ and $\int ${\displaystyle \int \left|j\right|dt}$) within the optimization objective J (Equation (16)). The PSO efficiently navigates this trade-off space, finding strategies that satisfy the dynamic safety constraints (Equations (17) and (18)) while minimizing overall maneuver duration. This demonstrates the model’s ability to achieve safe and efficient lane changes, not just safe ones.

As illustrated in Figure 6, the risk line chart derived from nine experimental groups (with ten trials for each group) reveals that Experiment 1 exhibits the most significant fluctuation in risk values across all scenarios, indicating a lack of stability. Conversely, Experiment 3 demonstrates the smoothest risk curve, with a risk range confined to only 0.2–0.5 in high-density scenarios, thereby confirming the model’s robustness under complex conditions. In high-density situations, the mean risk associated with Experiment 3 (0.35) is markedly lower than those of Experiment 1 (0.75) and Experiment 2 (0.55), underscoring the advantages offered by its dynamic risk assessment mechanism.

Discussion on Risk Stability: The markedly lower and more stable risk profile of Exp3 (Figure 6), particularly in high-density scenarios, reflects the model’s robustness to environmental dynamism. The combination of dynamic feature weighting (attention) and causal reasoning provides a more consistent and interpretable assessment of risk (R) compared to rule-based (Exp1) or potentially heuristic (Exp2) approaches, which might exhibit more erratic risk estimations as sensor inputs fluctuate or scenarios become complex. The ε term in Equation (1) and the normalization in feature vector S (Equation (4)) also contribute to numerical stability.

Justification of Experimental Design: The chosen traffic density levels (Low, Medium, High—Table 1) were specifically designed to stress test the model under conditions ranging from benign (Low) to highly challenging (High), mirroring real-world traffic variability. The selection of Exp1 (Traditional Safety Distance) as a baseline provides a fundamental physics-based benchmark widely understood in the field. Exp2 (LC2013 model in SUMO) represents a state-of-the-art, widely used model within a leading transportation simulation platform, serving as a practical benchmark for performance in simulated environments. The parameters for the PSO algorithm (N = 50, T = 100, ξ = 0.6, c1 = c2 = 1.5) and risk model weights (λ1–λ4, δ = 0.5) were determined through extensive simulation calibration runs aimed at balancing convergence speed, solution quality, and adherence to safety-critical constraints. The Gaussian noise level (±5%) applied to perception data (Section 5.1) reflects typical sensor uncertainties reported in the automotive literature [19] and ensures the model’s evaluation under realistic noisy conditions. These design choices collectively ensure a comprehensive and rigorous evaluation of the proposed model’s capabilities in the target operational domain (complex, interactive lane-changing).

To address traffic condition variability, the results were stratified by density level (Table 1). As shown in Figure 4, Figure 5 and Figure 6, the proposed model (Exp3) consistently outperformed baselines across all density scenarios, with most significant improvements in high-density conditions (e.g., 63.0% conflict reduction vs. Exp1). The statistical significance of differences between models was validated through one-way ANOVA with Tukey’s post hoc test (α = 0.05). All reported improvements (e.g., conflict counts, time reduction) were statistically significant (p < 0.01).

The aforementioned experimental data validate the efficacy of the context-dependent causal model in facilitating lane-changing decision-making within intricate traffic environments. This provides both theoretical support and practical engineering references for autonomous driving lane-changing decisions.

6. Conclusions

The context-dependent causal model developed in this study, utilizing data from electronic rearview mirrors, demonstrates substantial improvements in both the safety and efficiency of vehicle lane-changing behavior, as rigorously validated through comparative simulation experiments. The differences between models were statistically significant (p < 0.01 by ANOVA), confirming the robustness of results.

Safety Enhancement (Conflict Reduction):

(1)

Compared to the traditional safety distance model (Experiment 1), the proposed model (Experiment 3) achieved a remarkable 63.0% reduction in the average number of conflicts (from 12.7 to 4.7 instances) under high-density traffic conditions.

(2)

Compared to the built-in model of the simulation software (Experiment 2), a significant 37.3% reduction was observed (from 7.5 to 4.7 instances).

Efficiency Gains (Lane-Changing Time Reduction):

(1)

The proposed model (Exp3) reduced the average lane-changing duration by 10.9% compared to the traditional model (Exp1) (from 5.5 s to 4.9 s in high-density scenarios).

(2)

More strikingly, it achieved a 31.9% reduction compared to the simulation software’s model (Exp2) (from 7.2 s to 4.9 s), highlighting its superior ability to maintain efficiency even in congestion.

Decision Stability (Risk Fluctuation Mitigation):

The fluctuation range of risk values exhibited by the proposed model (Exp3) was 53.3% lower than that of the traditional model (Exp1) (from 0.75 to 0.35) and 36.4% lower than that of the simulation software’s model (Exp2) (from 0.55 to 0.35), demonstrating significantly improved robustness and stability in decision-making under dynamic conditions.

These quantitative results clearly demonstrate the superiority of the proposed context-aware causal inference model over both a fundamental physics-based approach (Exp1) and a representative state-of-the-art model implemented in industry-standard simulation software (Exp2). The integration of dynamic safe distance calculations, the attention mechanism for critical feature extraction, and counterfactual causal reasoning effectively addressed the challenges of real-time risk assessment and multi-objective optimization in complex, interactive traffic scenarios. Furthermore, the Particle Swarm Optimization (PSO) algorithm successfully harmonized safety constraints with driving efficiency and comfort, ensuring a stable and smooth lane-changing process. The findings substantiate the synergistic benefits of combining causal reasoning with multi-objective optimization in autonomous driving decision-making frameworks and provide a solid theoretical foundation and practical reference for enhancing safety control within intelligent transportation systems. Future research will extend these methodologies into more intricate scenarios involving multi-vehicle collaboration as well as extreme weather conditions aimed at bolstering the generalization capabilities of this model.