Figures (8)  Tables (6)

    • Figure 1. 

      Annual trend in publication volume.

    • Figure 2. 

      Keyword visualization.

    • Figure 3. 

      Discretionary lane change.

    • Figure 4. 

      Mandatory lane change.

    • Figure 5. 

      Schematic diagram of the phases of lane change.

    • Figure 6. 

      Classification diagram of lane change decision model.

    • Figure 7. 

      LSTM-GNN network architecture. 'sv1–sv8' respectively represent different surrounding vehicles.

    • Figure 8. 

      Lane change decision system.

    • Method used Advantages Year Ref.
      Long short-term memory Analyzes the factors affecting driving safety and the scope of potential risks 2020 [23]
      Multimodal trajectory prediction Explains human driving behavior by adjusting subjective risk parameters; suitable for mixed traffic scenarios. 2022 [24]
      Fuzzy logic theory Reveals the spatiotemporal dynamics of risk states and quantifies the influencing factors of risk. 2023 [25]
      Probabilistic model Fully considers position uncertainty and distance-based safety indicators. 2022 [26]
      Traffic conflict index Breaks through traditional single-threshold methods and enables dynamic risk grading. 2020 [28]
      Traffic conflict index Dynamically expands the TTC model based on high-precision microscopic trajectory data. 2022 [29]

      Table 1. 

      Lane change risk assessment methods.

    • Input variables Algorithm Issues Advantages Year Ref.
      Vehicle state and driver behavior features Rule-based two-stage decision model Driver behavior is not classified Quantitatively considered the response of following vehicles in the target lane 1986 [30]
      Traffic flow data and vehicle dynamic information MITSIM Does not consider nearby vehicles during lane changes Incorporated probabilistic logic, reflecting driver risk preferences 1996 [31]
      Position, speed, and acceleration of the vehicle and surrounding vehicles MOBIL Ignored vehicle dynamics during lane changes Applied benefit–safety dual-condition mechanism, closer to human driving 2007 [32]
      Road information, vehicle information, and other obstacle data HSM + RCS Insufficient model generalization and simplified safety evaluation Used real-road data and RCS to quantify lane change risk 2018 [34]
      Dynamic traffic data of vehicles Fuzzy inference + FSM Limited applicability of the algorithm, strong rule dependence Effectively handled uncertain information 2023 [35]
      NGSIM database Cellular automata model Overly simplified model Quantifying the differences in human driving styles 2023 [36]
      US-101 NGSIM database Rule-based + machine learning Motion prediction model remained relatively simple Balanced interpretability with adaptability to complex scenarios 2022 [37]

      Table 2. 

      Rule-based lane change decision method.

    • Input variables Algorithm Issues Advantages Year Ref.
      US-101 and I-80 NGSIM database SVM Relies on manual feature extraction, limited dynamic scene adaptability Incorporates physical, interaction, and road structure features 2020 [38]
      US-101 NGSIM database ABC-SVM Relies on ego-vehicle data, neglects surrounding traffic environment Avoids grid search inefficiency and local optima 2021 [39]
      US-101 NGSIM database Bayesian Network Lacks a real-time updating mechanism; limited adaptability Transforms driver decision uncertainty into quantifiable probabilistic outputs 2020 [40]
      High D dataset LSTM + Bayesian Network Neglects personalized factors such as driver style and vehicle type Dynamically updates decision thresholds, adapts to diverse traffic environments 2021 [41]
      Vehicle state, driver operations, and environmental data Improved IOHMM Lacks validation under real-world driving environments Addresses IOHMM limitations in sequential memory and continuous outputs 2021 [44]
      High D GMM-HMM Does not consider applicability in mixed traffic flows Identifies spatiotemporal interactions and high-risk patterns in lane changes 2023 [45]

      Table 3. 

      Lane change decision-making method based on traditional learning.

    • Input variablesAlgorithmIssuesAdvantagesYearRef.
      NGSIM databaseDBN-LSTMLacks dynamic modeling of driving stylesSimulate the entire lane changing process and its impact on traffic flow2019[46]
      NGSIM databaseLSTM-GNNIgnores key behavioral features; simplified interaction modelingFully models multi-vehicle interactions, significantly improving decision accuracy2023[47]
      Dynamic motion imagesCNN-based dynamic motion image representationLimited generalization to real-world scenarios and high computational costCaptures surrounding vehicles' positions and motion comprehensively2023[48]
      Comma2k19 and Udacity datasetsSpatiotemporal attention-based deep learningComplex model, long training timeSpatiotemporal attention highlights important frames and key regions2022[49]
      Vehicle state space and action spaceDSCNN- TransformerNot suitable for complex urban roadsMaintains temporal modeling capability while reducing computational load2023[50]

      Table 4. 

      Lane change decision method based on deep learning.

    • Input variablesAlgorithmIssuesAdvantagesYearRef.
      Relative positions and speeds of vehiclesRainbow DQNUnable to jointly optimize longitudinal speed and lane changing timingIntroduces a safety feedback reward mechanism2020[52]
      Longitudinal/lateral distances, yaw angle, relative distancesDRL + Risk assessment functionIgnores driving styles; model applicability is limited to specificAchieves optimal driving strategy with minimum expected risk2022[26]
      US-101 NGSIM databaseD3QN+DDPGFails to consider trajectory continuity and heterogeneous traffic flowSafely and efficiently handles lane changing and car-following behaviors2022[53]
      High D datasetDRLLimited application scenarios; reward function lacks consideration of comfortIncorporates a collision-avoidance strategy to ensure longitudinal safety2023[55]
      Relative distance, relative speed, and lane-relative positionDQN + IRLHigh computational complexity and depends on expert data from driving simulatorsIntegrates behavior, planning, and control modules for joint training and execution2025[56]
      Ego vehicle's longitudinal acceleration, yaw rate, surrounding vehicles' speeds and distancesOARLLimited by discrete actions; cannot handle continuous steering controlMaintains high performance and safety under observation disturbances2023[57]
      Vehicle position, speed, acceleration, and traffic light informationAH-TD3Lacks timeliness and poor continuity in modeling interactive behaviorMixed action representation integrates discrete and continuous actions2024[58]

      Table 5. 

      Lane change decision method based on reinforcement learning.

    • Input variablesAlgorithmIssuesAdvantagesYearRef.
      Traffic conflict indicatorsTwo-player non-zero-sum non-cooperative gameOversimplified assumptions underestimate real-road complexityDemonstrates feasibility of applying game theory to traffic interaction modeling1999[59]
      Vehicle state informationStackelberg Game with Incomplete InformationRelies on sensor data, subject to perception errorsCapable of handling uncertainty and enabling dynamic decision updates2016[60]
      Vehicle state and environmental informationNon-cooperative mixed-strategy gameInsufficient consideration of driving style diversityIncorporates a dynamic risk model into the lane changing game2022[61]
      Vehicle state information and driving style parametersCoalition game modelHigh computational complexity; relies on stable V2X communicationIntroduces perceived risk field theory to quantify uncertainties in mixed traffic2023[62]
      High D datasetIncomplete information gameOveremphasis on human driving uncertainty while ignoring vehicle dynamics constraintsUses Risk–Response (R-R) diagram to interpretively quantify social driving preferences2024[63]
      NGSIM datasetHierarchical game theory modelAssumes fully rational vehicles, while real-world mixed traffic is not fully rationalDynamically selects game model based on RDS2024[64]
      Vehicle status information and traffic trend prediction dataPrediction-enhanced game theory modelOversimplified scenarios, only two-vehicle games are considered while real lane changes involve multi-vehicle interactionUtilizes macroscopic traffic flow information for proactive decision-making2024[65]
      US-101 NGSIM databaseImproved Stackelberg game theoryIgnores multi-lane continuous lane changing demand; insufficient safety mechanismsStrong adaptability in multi-lane scenarios with online real-time classification of driving styles2025[66]

      Table 6. 

      Decision-making method for lane change based on game theory.