AI Based Novel Approach to Detect Driver Drowsiness

Abstract

Vehicle accidents due to driver drowsiness are the biggest problem all over the world. It leads to millions of road accidents result in fatal and non-fatal accidents and huge financial losses not only for drivers but also for their victims and their families. According to various countries road accident reports, driver drowsiness represents up to 30% of all vehicle crashes [1]. This is the major concern for all the countries where huge accidents occur on the roads. To overcome vehicle accidents and provide safety to all commuters on the roads, a smart driver drowsiness detection system is required to detect the drowsy state of the driver in the early stage. Lots of research are going on in the field of driver drowsiness detection system by the researchers all over the world. Many techniques and methods proposed by various researchers. The three main measures that detect driver drowsiness: vehicle- based, physiological and behavioural [2]. Vehicle-based and physiological measures are traditional/conventional approaches that are used to detect driver drowsiness. But these approaches have number of limitations like road conditions, driver expertise and intrusiveness which is very difficult to implement in real driving conditions. The behavioural measure is the new way to detect driver drowsiness. It is based on the face recognition technique where major signs i.e. eye state, head movement and frequently yawning can be captured by the camera. Further, these images that are recorded by the camera can be analysed by AI or its subsets. i.e. machine learning (ML) and deep learning (DL). ML and DL play a vital role to develop the visual-based driver drowsiness detection system. There are number of algorithms that are used to detect the drowsy state of the driver [3]. It is challenging to select the appropriate algorithms that detect drowsiness with high accuracy.

Fig 1. represents the DL and ML that is subsets of AI

ML is a subset of artificial intelligence that serves to provide the machines with the ability to automatically learn and act based on previous experience. DL, in its turn, is a subset of machine learning. DL is the machine learning method that allows computers to mimic the human brain to complete classification tasks on images or non-visual data sets. Deep learning uses the only algorithm-neural network similar to the human neural system to data mining and analyses various factors.

In behavioural measures, the facial features are extracted by the computer vision algorithm i.e. landmark localization (LL), Histogram of orient gradient (HOG) and Local Binary Pattern (LBP). After extraction of various facial features, the ML or DL based algorithms are analysed the three states of the driver. i.e. eye state, yawning analyses and head position.  Various algorithms for filtering the states of the driver i.e. SVM, CNN, HMM etc. ML-based and MTCNN, DSST, KCF+CNN etc. DL based algorithms that are used. Each algorithm used a different dataset to analyse the state of the driver and most of the datasets has a limited set of images and are unable to provide complete results [4]. A meta-analysis has been conducted on 20 papers that use ML and DL based algorithms to detect driver drowsiness. The analyse reveals that DL based algorithms perform better to detect the drowsy behaviour of the driver as compare to ML. In addition, modified CNN based algorithms have a high detection rate and less false detection.