Driver Drowsiness Detection Using Hybrid Model CNN and LSTM

keyword: Driver drowsiness detection, Convolutional Neural Network (CNN), Long Short Term Memory (LSTM), Real Time Monitoring, Deep Learning, Computer Vision, Fatigue Detection, Hybrid CNN-LSTM model Abstract - Driving fatigue is one of the listed causes of road accidents, statistics reveal that driving drowsiness accounts for 25% to 50% of all road traffic crashes. Detection of driver fatigue on time is one key area to increase the safety in road. Various modalities have been studied for drowsiness detection, including EEG, vehicle driving dynamics and eye movements. EEG is incredibly precise but an impractical real world application as it is invasive. Hard to Get but Not AccurateTo track driving behavior, the detection of eyeball movement is used, which is a good balance between speed and effectiveness, but the current system utilizes a high-speed camera and a complex algorithm, making it difficult to implement on embedded platforms. To make progress on these issues, recent work has proposed the use of deep learning solutions trained using extensive datasets with a standard video camera in order to deliver more reasonable substitutes to costly eye-tracking systems. These recurrent neural networks (RNN), in particular the Long Short-Term Memory (LSTM) networks, are especially suited for this purpose. With one approach, the authors examined 48 × 48 eye patch images obtained from a simulated driving study and obtained 82% accuracy with LSTM and up to an 97% C-LSTM model, respectively. In contrast, other models use CNNs on top of LSTM or BiLSTM network to take advantage of spatial and temporal characteristics of eye movements and facial expressions. These hybrid models can classify eye blinks, closure duration and landmarks of the eyes and register drowsiness, and some of them have been able to achieve 99% accuracy. Contrarily, state-of-the-art techniques from paper [14] apply FaceMesh for facial landmarks localization and yawning detection as IOU and finally, the head pose estimation techniques are implemented for classifying driver attention. The highest reported accuracy of up to 99.71% achieved using ResNet50V2 as a base neural network architecture and trained using NITYMED and NTHU datasets. Face check using OpenCV, alarm system, and graphical signals