Driver fatigue detection using image sensors

Abstract

The overall purpose of this study is to develop a software to detect driver fatigue using image sensors. The basic design of the study includes both self-built CNN model and also Transfer Learning using pre-trained model, specifically Inception-V3 and ResNet-50 to detect blink using classification of open and closed eyes which is eventually used to detect blink rate. Eyes is used in this project since it is the most significant symptom of fatigue. Higher blink rate usually associates with fatigue. Dataset was originally trained on ResNet-50 before being replaced by Inception-v3, giving an opportunity to compare results between the two models. Overall, ResNet-50 has proven to give higher accuracy. An important pointer to note is the biasness of the data which always has to be rectified. With the better pre-trained model in my research being ResNet-50, this diverges with my research of comparison between the two models which states that Inception-v3 is more accurate. In future, further development of the product will leverage on this model. In this project, implementation is reported in detail, together with evaluation of the results obtained. Finally, potential of further experiments to venture into could be done before concluding the project.