Abstract: To enable personalized and safe lane-changing decisions for autonomous vehicles in multi-vehicle interactive environments, a game theory-based multi-cost lane-changing decision model was proposed. The total decision cost comprised three sub-costs, namely safety, comfort and traffic efficiency, and a driving style recognition mechanism was introduced to support personalized lane-changing games among multiple vehicles. Principal component analysis was employed to extract features from drivers’ lane-changing data, K-means clustering was used to classify driving styles, and the convolutional neural network was adopted for style recognition. The interaction and decision-making process of lane changing between vehicles was modeled as a strategic game, through which the lane-changing strategy minimizing the total decision cost was derived via game-theoretic analysis. Typical highway scenarios were selected to conduct two-stage game simulations for four driving style combinations, to accomplish overtaking maneuvers. The research results indicate that when the driving styles of the ego vehicle and the obstacle vehicle tend to be consistent, the “cooperative” effect between the two parties in the game is enhanced. Specifically, compared with combination 2 #, combination 1 # reduces the total cost of the first stage game by 8.80% and that of the second stage game by 4.31%, respectively; compared with combination 4 #, combination 3 # reduces the total cost of the first stage game by 7.01% and that of the second stage game by 2.57%, respectively. When the ego vehicle’s speed is substantially higher than that of the obstacle vehicle, the game cost decreases significantly, with the second-stage game cost respectively reduced by 50.50%, 52.80%, 41.33% and 44.00% for the four combinations, compared to the first-stage game cost. At this time, the obstacle vehicle is more likely to accept lane changing, significantly reducing conflicts and risks during the lane changing process, and improving traffic efficiency. The proposed game-theoretic lane-changing decision model forms a targeted technical pathway for driving style recognition, multi-style strategic game modeling, and personalized trajectory planning, which can support optimal lane-changing decisions for autonomous vehicles in complex traffic environments.

Key words: traffic engineering; lane change decision-making; personalized driving style; Stackelberg game; trajectory planning; NGSIM dataset

摘要： 为实现自动驾驶车辆在多车交互环境下的个性化安全换道决策，提出了一种基于博弈论的多成本换道决策模型。决策总成本由安全、舒适及通行效率等3项子成本构成，通过引入驾驶风格识别机制，以支撑多车间的个性化换道博弈；采用主成分分析法提取驾驶者换道数据特征，通过K-means聚类实现驾驶风格分类，并利用卷积神经网络完成风格识别；将车辆间换道互动与决策过程建模为策略博弈，通过博弈分析求解决策总成本最小的换道策略；选取典型高速公路场景，针对4种驾驶风格组合开展两阶段博弈仿真，完成超车动作。研究结果表明：当主车与障碍车辆驾驶风格趋于一致时，博弈双方的“合作”效应增强； 1#组合较2#组合、 3#组合较4#组合，第一阶段博弈总成本分别降低8.80%、 7.01%，第二阶段博弈总成本分别降低4.31%、 2.57%。当主车车速显著高于障碍车车速时，博弈成本明显下降，这4种组合在第二阶段较第一阶段博弈成本分别降低50.50%、 52.80%、 41.33%、 44.00%，此时障碍车更易接受换道，换道过程中的冲突与风险显著降低，通行效率得以提升。所提出的博弈换道决策模型在驾驶风格识别、多风格策略博弈建模及个性化轨迹规划方面形成了具有针对性的技术路径，可为复杂交通环境下自动驾驶车辆的优化换道决策提供支撑。

关键词: 交通工程；换道决策；个性化驾驶；Stackelberg博弈；轨迹规划；NGSIM数据集