Unsafe Driver Behavior Detection Using Novel Dictionary Algorithm

城市和高速公路上发生事故的原因有很多。导致交通事故的一些因素包括司机和行人的情绪状态、疲劳和注意力不集中。这个研究项目的短期目标包括通过分析司机的眼睛、眉毛和嘴唇来识别他们的面部表情。安装在车内的摄像头网络被用来拍摄静止图像。图像处理技术用于提取情感检测所需的面部物体和特征。这项研究计划的长期目标是扩大车内摄像头系统，在十字路口安装智能摄像头网络，以提醒司机和行人由于疲劳或注意力不集中而产生的潜在危险。这项研究的新颖之处在于在图像处理中使用了一种增强的字典方法，其中引入了新的原子。情绪/疲劳识别是当图像通过经过训练的算法时，基于输入图像数据库对各种表情的识别。情绪/疲劳识别的一个应用是根据一个人的状态来确定他/她即将发生的身体反应。恐惧、焦虑、分心、愤怒、注意力不集中和疲劳都会损害身体的平衡，影响一个健康人站立和行走时的稳定性。它们还会对个体的反应时间产生负面影响。疲劳是许多导致伤亡的驾驶事故的已知原因。因此，通过人类面部表情检测到的疲劳和情绪变化可能会对即将发生的事故发出警报。在本研究中，我们通过分析驾驶员的面部表情来识别潜在的危险状况，并向现场的驾驶员和行人发出警报。通常，识别算法包括三个主要步骤：1)获取步骤，其中检测到表达状态的个人工件；2)特征提取与表示，根据所选择的特征提取方法，将提取的成分以几种不同的方式表示；3)表情分类步骤，算法利用提取的特征，确定最适合参与者的情绪。这些算法面临的挑战之一是实现高水平的识别率，低水平的误检（虚警）率，以及高灵敏度和特异性。噪音可能导致错误识别；例如，一张中性的脸可能被误认为是一张悲伤的脸，或者一个平静的声音可能被误认为是一个不安的声音。字典学习，特别是当与其他信号处理算法相结合时，在特征提取方面已经被证明是强大的。我们已经证明，当将其他非线性原子引入DCT（离散余弦变换）字典时，字典算法产生了更好的结果。然而，这种增强的方法尚未应用于情绪/疲劳检测，而这正是本研究项目的目标。本研究项目的成果包括：1)一种具有高分类率的增强图像处理算法，能够检测车内驾驶员的情绪状态；2)识别车辆方向和行人注意力不集中的交叉口的目标方向检测；3)车载摄像头系统与交叉口智能摄像头系统之间的通信协议；通过开发一种技术来检测驾驶员的不安全行为并向他们提供适当的警告，该项目促进了道路安全和事故预防。减少事故导致降低治疗费用（包括康复和事故调查）和减少生产力损失（或缺勤）。

Accidents happen on the city and highway roads for many reasons. Some of the factors contributing to road accidents include the emotional state, fatigue, and inattentiveness of the drivers and pedestrians. The short-term goal of this research program involves with driver’s facial expression recognition by analyzing their eyes, eyebrows, and lips. A camera network installed inside the car is utilized to take still images. Image processing techniques are used to extract the facial objects and features necessary for emotion detection. The long-term goal of this research program is to expand the in-car camera system with smart-camera networks installed on intersections to alert the drivers and pedestrians of potential dangers due to fatigue or inattentiveness. The novelty of this research is the use of an enhanced dictionary approach in image processing where new atoms are introduced. Emotion/fatigue recognition is the identification of various expressions based on a database of input images when an image is passed through a trained algorithm. One application of emotion/fatigue recognition is in identifying the upcoming physical reaction of an individual based on his/her state. Fear, anxiety, distraction, anger, inattentiveness, and fatigue could compromise the body’s balance, and impact a healthy individual's stability during standing and walking. They also negatively affect the individual’s reaction time. Fatigue is a known cause of many driving accidents resulting in injuries and death. Therefore, fatigue and changes in emotions detected via human's facial expression could alarm an upcoming accident. In this research, we analyze drivers’ facial expressions to identify potentially dangerous conditions and alarm drivers and pedestrians involved in the scene. Typically, recognition algorithms consist of three main steps: 1) the acquisition step in which the artifact of an individual while expressing a state is detected; 2) feature extraction and representation in which the extracted components are represented in several different ways based on the selected feature extraction method; and 3) expression classification step in which using the extracted features, the algorithm determines the best suited emotion of the participant. One of the challenges in these algorithms is to achieve a high level of recognition rate, a low level of misdetection (false alarm) rate, as well as high sensitivity and specificity rates. Noise could result in false recognition; e.g., a neutral face could be mistaken as a sad face, or a calm voice could be identified as a disturbed voice. Dictionary learning, especially when combined with other signal processing algorithms, has proven powerful in feature extraction. We have shown that dictionary algorithm yields better results when other nonlinear atoms are introduced into DCT (discrete-cosine transform) dictionary. However, this enhanced approach has not yet been applied to emotion/fatigue detection and that is what this research program aims at. The deliverables of this research program are: 1) An enhanced image processing algorithm that detects the emotional state of the driver inside a car with high classification rate, 2) object orientation detection for intersections where cars direction, and pedestrian inattentiveness are identified, 3) communication protocol between in-car camera system and intersection smart-camera system, and 4) warning system for both drivers and pedestrians. By developing a technique to detect the unsafe behaviors of drivers and provide them with proper warnings, this project promotes road safety and accident prevention. Fewer accidents leads to reduced cost of treatment (including rehabilitation and incident investigation) and reduced loss of productivity (or absenteeism).