Real-time driver behavior detection system for smart vehicles

• Main Author: Yu, Yingchen
• Other Authors: Huang Guangbin
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/64739
• Institution: Nanyang Technological University

As most accidents are caused by unsafe driving behaviors such as distraction and fatigue, this system is designed to detect and recognize those unsafe driving behaviors using computer vision techniques. In this project, the main objects of interest are the head whose movements most perceivably reflect whether the driver is keeping attention. There are three key steps to achieve the recognition of the driving behaviors, firstly the videos of driver behavior are captured by the normal camera, and then the state-of-art face detection and head-pose estimation algorithms are applied to obtain the results about the driver’s head orientation. In this process, two most suitable head-pose estimation algorithms were implemented and compared, and the Constrained Local Neural Fields (CLNF) model is chosen to be used in this system due to its excellent robustness and accuracy. Lastly, based on the training data which are collected from the driving simulation device, the machine learning algorithm is used to classify the head-pose results and achieve the driver behavior recognition. Among various state-of-art machine learning algorithms, the Extreme Learning Machine (ELM) algorithm achieved the best accuracy of 88.57%. It is the first time that CLNF and ELM algorithms are applied in driving scenario, which contributes the novelty of this project. In this project, the driver behavior detection system was successfully designed and implemented, it is able to detect some basic driver behavior from the driving videos and achieve a promising system performance. While this project sets out to detect unsafe driving behaviors robustly, it is recommended to add more driving behaviors into the dataset and meanwhile improve the accuracy of this system.