基于多模态信息联合判断的驾驶员危险行为监测系统

危险驾驶是引发交通事故的主要原因,严重威胁公众的人身和财产安全.车辆在行驶过程中,各种环境因素复杂多变,严重干扰了驾驶员危险行为监测系统的识别精度;现有的驾驶员危险行为监测检测系统主要是通过眨眼和打哈欠的频率来判断驾驶员是否出现危险驾驶行为,忽略了表情、动作以及视线方向等多个重要信息.针对这些问题,提出了一种基于多模态信息联合判断的驾驶员危险行为监测系统,该系统使用分布在驾驶舱3个不同位置的近红外图像作为输入,解决了光照强度变化和视角盲区问题;设计了一种基于多任务学习的深度神经网络,该网络可以同时完成人脸检测、表情识别以及危险动作分类等任务,极大地提升了系统的识别精度,提升了运行效率.实验证明,所提出系统的识别准确率为96.2％,运行速度为81 fps,性能优于目前常用的算法.

Driver's Dangerous Behavior Monitoring System Based on Multi-modal Information

Dangerous driving is the main cause of traffic accidents, which seriously threatens the safety of people and property of the public. When the vehicle is driving, various environmental factors are complex and changeable, which seriously interferes with the recognition accuracy of the driver''s dangerous behavior monitoring system; the existing driver''s dangerous behavior monitoring and detection system mainly judges driving by blinking and yawning frequency Whether the driver has dangerous driving behavior, ignoring multiple important information such as facial expressions, movements, and direction of sight. In response to these problems, this paper proposes a driver''s dangerous behavior monitoring system based on multi-modal information joint judgment. This system uses near-infrared images distributed in three different positions of the cockpit as input, which solves the change of light intensity and blind spots of viewing angle. Problem: A deep neural network based on multi-task learning is designed. The network can complete tasks such as face detection, expression recognition, and dangerous action classification at the same time, which greatly improves the recognition accuracy of the system and improves the operating efficiency. Experiments show that the recognition accuracy of the system proposed in this paper is 96.2%, and the operating speed is 81fps, which is better than the current commonly used algorithms.