According to studies a large number of accidents is related with driver’s physiological state. Such states related to accidents are mainly fatigue (drowsiness or mental fatigue) as well as stress. In the USA a series of studies by the National Transportation Safety Board (NTSB) have pointed to the significance of sleepiness as a factor in accidents involving heavy vehicles.

The NTSB came to the concluded that 52 per cent of 107 single-vehicle accidents involving heavy trucks were fatigue-related; in nearly 18 per cent of the cases, the driver admitted to falling asleep. Summarizing the US Department of Transportation's investigations into fatigue in the 1990s, the extent of fatigue-related fatal accidents is estimated to be around 30%. Research shows that driver fatigue is a significant factor in approximately 20% of commercial road transport crashes and over 50% of long haul drivers have fallen asleep at the wheel.

In the UK alone, almost 45,000 people are killed, or seriously injured in road accidents every year, and road safety experts consider driver fatigue is a major cause. Driver fatigue is shown to be responsible for more than 20% of traffic accidents in UK. Stress is also considered to be an accident factor. In routine driving tasks, stress may cause failures of memory and attention. For high-demanding driving tasks, stress may result in mistakes due to reduction of processing capacity, or changes of attentional selectivity and narrowing of attentional focus . On the other hand, the significant correlation may imply that errors could result in additional stress, an argument consistent with the transactional stress approach.

1. Based on Video vigilance of driver’s and extraction of eye blinking and head movement.
2. Based on driver’s physiological signal acquisition (ECG, EEG, Respiration rate, Skin Conductivity, blood pressure e.t.c.) and extraction of features correlated with states under investigation.
3. In the combination of the abovementioned approaches.
   

Figure 1. Basic steps towards Driver State Assessment

In the first stage we acquire the physiological signals. In our study we acquire ECG, Respiration, EMG and Skin Conductivity. Those signals are extensively used in literature and proved correlated with many physiological states. A very challenging task is to define an experimental protocol that will provide necessary data for the method you are going to derive. In physiological measurements involving driver’s the first question is whether to use a simulation environment or to perform experiments on real driving conditions. The first approach is a controlled safe one, while the second can give more realistic driver’s reactions to different stimulus. We selected the second one and we have performed a large number of experiments so far with a “portable” data acquisition system (Biopac MP100) installed on a car (Figure2).

Figure 2. Data Acquisition on Real driving experiments

The next step is the preprocessing and the feature extraction. Our main effort lies on the investigation of new features both on time-domain and frequency-domain correlated with different states. Some of the approaches used are time-frequency analysis, wavelets, time-varying AR processes and others. An important issue is the window length, referring to the size of the data in which we apply the above mentioned methods to extract features. This length is related to the type of state that we are investigated. In emotion recognition methods, the length is about 10 sec, the time that a stimulus-driven emotion is expected to last. A physiological state such as fatigue and stress, are considered permanent states and larger windows (2-5 mins) can be applied.

The last step is the classification of the features. There is a lot of classification method and there is no golden standard. A common practice is to investigate the performance of different classifiers and find the best one.

However there is a rising issue on the classification of features extracted from physiological signals. Each subject has a different baseline. For example each subject has a different normal heart rate and we cannot define thresholds on the heart rate value itself. Instead, it is a common practice to first apply a normalization method which removes those individual baselines.

There are two similar in concept approaches to deal with changing problem environment (a new driver).

So far we considered the permanent driver’s physiological states (fatigue, stress) . We are also interested in the driver’s response to specific stimulus (similar to emotional recognition studies) . The main goal is to derive real time measures of stress and correlate those metrics whith specific driving events. Such knowledge could allow smart driving systems that infer the type of events that cause stress to drivers and adapt their functionality. Some methods applied on the problem are Dynamic Bayesian Networks and Extended Kalman Filters. In Figure 3 we demonstrate the impact of different event (annotated by the driver as low-medium-high stress events) on the Heart Rate and the Skin Conductivity.

Figure 3. EDA and Heart rate of a driving session. Red lines indicate High Stress Events, Black lines indicates medium stress events and green line low stress events

This work is funded by the Greek Secretariat for Research and Technology (PENED: project 03ED139, Title:"Intelligent System for monitoring driver's emotional and physical state in real conditions").