【摘要】 驾驶员的情绪变化会直接影响其驾驶行为．目前对于驾驶员行为预测的研究依然采用统计分析的方式，无法反映情绪状态对驾驶员行为的影响，导致对驾驶员动力需求的预判存在信息偏差．为此，针对实际驾驶过程中的驾驶员情绪状态判断，提出了满足实车应用的驾驶员情绪识别算法，同时设计试验研究了不同情绪间驾驶行为的特征差异．首先，针对车载计算单元相对不足的问题，结合人脸动作单元和面部表情编码系统，简化了高兴和愤怒两种表情的面部关键点，进而降低特征维度．然后，使用支持向量机分别建立了两种情绪的识别模型，通过网格搜索算法优化了模型参数，提升离线识别准确率．最后，结合情绪诱导设计了实车道路试验，基于通用计算平台测试了模型的在线识别效果，同时对比了驾驶员在高兴和愤怒时的驾驶行为以及车辆运行差异．结果表明：利用简化特征建立的高兴和愤怒识别模型，数据集离线测试的准确率分别达到了84.45%和84.43%；通用计算平台的测试结果显示该算法满足车载实时应用的需求，与情绪诱导结果对应，能够在线识别驾驶员的情绪状态；与高兴情绪相比，驾驶员在愤怒情绪下的驾驶行为更加激进，加速踏板的操作从长期维持不变转为周期性的剧烈踩踏，踏板频谱幅值最大提高了165.1%，整车油耗最大增加了14.2%.更多还原

【Abstract】 Drivers’ emotional changes considerably impact their driving behavior. However,current research on driver behavior prediction mainly relies on statistical analysis,failing to capture the influence of emotional states on driving behavior. This limitation introduces information bias in predicting driver power demand. To address this issue,a driver emotion recognition algorithm was developed for real-vehicle applications to assess drivers’ emotional states during actual driving. Experiments were conducted to investigate variations in driving behavior characteristics across different emotional states. To overcome the challenge of limited onboard computing resources,the facial key points for happy and angry expressions were simplified by integrating the facial action coding system and face action units,effectively reducing the feature dimensions. Recognition models for the two emotions were separately developed using a support vector machine,with model parameters optimized via a grid search algorithm to enhance offline recognition accuracy. Subsequently,a real-vehicle road test incorporating emotion elicitation was conducted to evaluate the model online recognition performance using a general-purpose computing platform. In addition,the driving behaviors of drivers in happy and angry states were analyzed,and differences in vehicle operation under these emotional conditions were compared. The results indicate that the offline testing accuracy of the simplified feature dataset for recognizing happy and angry emotions reach 84.45% and 84.43%,respectively. Tests conducted on the generalpurpose computing platform demonstrate that the algorithm meets the requirements for in-vehicle real-time applications,successfully recognizing the driver’s emotional state online and aligning with the outcomes of emotion elicitation. Furthermore,compared with happy state,drivers in angry state exhibit more aggressive driving behaviors. In particular,accelerator pedal operation changes from long-term steadiness to periodic vigorous pressing,with the pedal spectrum amplitude increasing by up to 165.1% and overall vehicle fuel consumption increasing by up to 14.2%.更多还原