The first data set used in this PhD project was obtained from the TeleFOT project which is a large-scale collaborative European FOT funded under the seventh European Commission framework research programme. The objectives of the TeleFOT project were related to safety and mobility as well as economic/fuel-efficient driving and user-acceptance aspects of aftermarket and nomadic devices (e.g. SatNav, Speed Alert etc.) that can be introduced into the vehicle once it has ‘left the showroom’.
As part of the experimental work in TeleFOT, the participants were asked to drive along a specific 16.5 km long route in the Leicestershire area of England, as depicted in Figure 3.5a, after driving for a couple of hours to familiarize with the instrumented car and their behaviour was captured, monitored and analysed using a software programme developed by Race Technology Ltd. The route was carefully chosen to have a good mixture of different road elements such as roundabouts, T-junction, cross-junction, traffic light, mid-block crossings and the existence of dynamic obstacles (e.g. other vehicles, pedestrians, cyclists). This was to capture braking behaviours that significantly vary due to the road element. There were 44 trips conducted by 25 drivers in which data-logging occurred, as 19 of the drivers performed multiple trips.
An instrumented vehicle capable of recording driver behaviour, vehicle kinematics and driving environment (e.g. traffic density, road elements) was employed. Since a single vehicle was used in the experiment, the influence of vehicle-related factors (e.g. engine size, vehicle power and braking performance) need not be considered. The vehicle was equipped with four video cameras (forward road view, driver face, backward road view, driver reaction from the passenger seat), GPS, speedometer and accelerometer (see Figure 3.5b). The sampling frequency was 100 Hz for the duration
49
of the entire trip with an average driving time of 30 minutes per trip. This resulted in a total of 10.8 million observations. The data was processed by software (with a built-in noise filter) developed by Race Technology (Figure 3.5) (Fruttaldo, 2011).
Figure 3.5: The route of the field test (a), the view from the 4 cameras (b) and the acceleration-time and speed-time diagram for the whole trip from the Race Technology
programme (c).
In order to extract the data of interest from this project, the Race Technology V8.5 software was used. For each of the trips, available time-variant variables were: time, longitudinal acceleration, speed of the car, travelling distance, the video frame and the GPS coordinates. Traffic characteristics data for the study area were not available from the traffic management centre. However, traffic density is an important variable that might affect driving behaviour along with other variables that were not included in the available time-series, such as the reason for braking, the existence of congestion etc.
To obtain these variables that might play a significant role in the braking behaviour, the videos were watched, and the desired variables were extracted qualitatively. The procedure was as follows: first, the detection algorithm extracted all the deceleration
50
events of the dataset and saved in a different file the video frame number that corresponds to the beginning of each deceleration event. Then, the video of each trip was initiated and by tracking the video frame number of each event, the necessary variables were obtained.
Specifically, the traffic density was measured by counting the number of vehicles and taking into account the length of the visible roads. Since it was calculated qualitatively, it was included in the models as a categorical variable (i.e. low, medium or high traffic density). Moreover, the situational factors (i.e. the reason for braking) were also determined qualitatively by viewing the videos related to the deceleration events (Table 3.3 and Table 3.4). The situational factors that were examined were: the presence of a traffic light, whether the car stops in car blocks, which indicates the existence of congestion and the cause for braking (i.e. if the car decelerates because it approaches a roundabout, a cross or T-junction, a pedestrian crossing or because of an obstacle like pedestrian, bicycle or road jump). In detail, by watching multiple times the video frame starting 5 seconds before the beginning of the deceleration event it was possible to recognise the most challenging variable, the reason for braking.
Table 3.3: Percentage of deceleration events by reason for braking Reason for braking Deceleration event Percentage (%)
Roundabout 16.86 T junction 30.07 Cross junction 8.37 Pedestrian Crossing 5.20 Dynamic Obstacle 39.50 Existence of traffic light 15.33
Table 3.4: Percentage of deceleration events by traffic density Traffic Density Deceleration event Percentage (%)
Low 65.64
Medium 27.85
High 6.51
The rest of the data that are important for the analysis were obtained by developing a data extraction algorithm which was implemented in Matlab. Specifically, the trip
51
duration, the maximum and the mean deceleration of the car during the event, the mean and the initial speed of the car during the event, the duration of the event and the travel distance during the event were calculated. Furthermore, the sample was composed of 25 drivers (14 males and 11 females) with an average age of 40 years, varying from 23 to 59 years old. Information on the driver (subject number, age, gender, driven miles per year) was reported in the summary sheet of the file.
Table 3.5: Descriptive statistics of the variables during deceleration events
Variable Mean SD Minimum Maximum
Max deceleration (m/s2) -2.38 0.4 -4.885 -2.00 Duration (sec) 4.26 1.98 0.74 14.95 Mean deceleration (m/s2) -1.32 0.34 -3.46 -0.50 Final speed (km/h) 13.57 11.84 0.00 78.49 Initial speed (km/h) 34.66 14.99 4.10 107.51 Mean speed (km/h) 24.71 12.75 1.88 88.2 Distance covered (m) 2.31 2.27 0.8 18.52
Trip duration (min) 33.02 4.70 22.58 43.05
Given the range of data types captured by the instrumented vehicle, it was possible to analyse the deceleration events (i.e. the deceleration value and the duration) based on different influencing factors related to the driver (e.g. age, gender and experience), vehicle kinematics (e.g. the initial speed before the event), traffic (e.g. low, medium and high traffic density) and road infrastructure. Various descriptive statistics were generated to understand these factors (Table 3.5). The average max deceleration value was found to be -2.38 m/s2 and its absolute maximum value was -4.85 m/s2,
while the average duration was 4.26 sec, with a maximum value of 14.95 sec.
Most of the deceleration rates observed in this study are relatively low as can be seen in Figure 3.6(a) and this may be due to the nature of the FOT which reflects the driver’s normal braking and does not include any safety-critical events. Therefore, the threshold was set at 2m/s2 for this study, which is the lowest value found in the
literature to detect deceleration events (Wu et al., 2009), as it was explained in the methodology section.
52
Figure 3.6: Characteristics of extracted deceleration events.
The beginning and the end of the deceleration event are defined according to the threshold that was described in the Methodology chapter. The algorithm, that was developed in the Matlab software package R2016a to detect the deceleration events resulted in a total of 937 events. That algorithm also computes some essential parameters i.e. the duration (sec) of the deceleration event (see Figure 3.6b), the maximum and the mean deceleration rate (m/s2), the speed at the beginning and at
the end of the event, the video frame of the start of the event in order to detect it in the videos and receive more information and the travelled distance (m) of each event. Moreover, the algorithm splits each event into two parts and calculates the best fitted braking function for each event, which is added as an explanatory variable.
It should be noted that after detecting the deceleration event, a dataset that includes the dependent variables, i.e. the deceleration value and the deceleration duration, and all the explanatory factors that were obtained directly from the Race Technology V8.5
53
software, from the algorithm and from the videos was formatted as depicted in Table 3.2. The next step was the detection and the deletion of any possible outliers, that might have resulted from the computational procedure or the data gathering procedure. Since, the deceleration events are of interest, univariate detection of outliers considering the duration of the event was performed. Specifically, the percentiles of the duration were calculated in SPSS and the upper and lower threshold were computed by these equations:
Inter Quartile Range (IQR) = Q3 – Q1 Upper Threshold (UT) = Q3 + 2.2 IQR Lower Threshold (LT) = Q1- 2.2 IQR
Moreover, since the values of the duration were small, the lower threshold came out to be negative but a deceleration event with really small duration, for example, 0.3sec should not be taken into consideration since it is really short to be analysed and modelled, a lower limit of 0.5sec was set for the duration. The same procedure was following for the adjusted R2 of the fitted functions that were calculated. Therefore, the
observations whose values are outside the limits were excluded from the database, resulting in 869 deceleration events.
One more thing that should be taken into consideration in the pre-processing is the possibility of any two explanatory variables to be correlated. If there are two or more correlated variables, only one should be included in the statistical analysis. Therefore, the correlation table was computed using the SPSS software and if the Pearson- correlation value was more than 0.8 indicating a high correlation between the variables, then only one of them was included in the model. For the TeleFOT data only the event_id, the trip_id and the driver_id were highly correlated.
For the deceleration events, the average initial speed was 40 km/h (25 mph), with 80% of the events starting at speeds between 5 m/s (11 mph) and 15m/s (33 mph). The mean duration and mean deceleration value for different initial speeds are presented in Figure 3.6 and Table 3.6 and it can be concluded that the higher the initial speed the longer and harder the deceleration event.
54
Table 3.6: Average deceleration statistics based on different factors Mean values
Maximum deceleration value (m/s2) Duration (sec)
Gender: male -2.447 2.64 female -2.423 2.83 Initial speed: 0-5 -2.385 0.94 5-10 -2.379 2.08 11-15 -2.442 3.04 16-20 -2.399 4.48 21-25 -2.570 6.01 >25 -2.697 8.23 Traffic density: low -2.426 2.70 medium -2.422 3.17 high -2.379 2.82 Age: 20 - 30 -2.490 2.48 30 - 40 -2.486 2.39 40 - 50 -2.374 2.91 50+ -2.388 2.80 Traffic light: Signalised -2.480 3.44 Unsignalised -2.380 2.57 Reason of braking: Roundabout -2.375 3.94 T-junction -2.407 3.34 Cross- junction -2.355 2.54 Mid-block crossing -2.500 1.53 Dynamic-obstacle -2.406 1.91
As far as the traffic density is concerned, most of the deceleration events (66%) occurred in low traffic density conditions, 29% in medium traffic conditions and only 5% in high-density conditions. The mean of the maximum deceleration values for different traffic densities (-2.43 m/s2 for low traffic density, -2.42 m/s2 for medium traffic
density and -2.38 m/s2 for high traffic density) did not indicate that a relationship exists
between the observed rates and the traffic density (see Table 3.6). Moreover, it can be noted that gender does not affect the deceleration value but affects the duration as males seem to decelerate in a shorter time than females. However, as can be seen in
55
Figure 3.7 the male drivers have a bigger percentage of hard deceleration events (deceleration value<-3m/s2) comparing to female drivers.
Figure 3.7: Percentage of different deceleration values based on the gender of the driver.
Also, younger drivers seem to decelerate in a harder way, both greater deceleration value and shorter duration. For example, drivers aged between 20-30 years old had a mean deceleration value of -2.49 m/s2, which is 5% lower than the deceleration value
of the 40-50-year-old drivers and the deceleration event lasts on average 15% less. The deceleration events for each reason of braking are 143 in roundabouts, 255 in T- junctions, 71 in cross-junctions, 44 in mid-block crossing and 335 for obstacles (Table 3.6). As can be concluded from Table 3.6 the reason for braking affects slightly the deceleration value and more the duration of the event, with the durations for mid-block crossings and dynamic-obstacles being relatively shorter. Finally, some influence is noted between the deceleration event and the fact that a road element is signalised or not, which is that for signalised road elements the deceleration value is greater than for non-signalised (Figure 3.8). Moreover, the duration time of the deceleration event for the non-signalised elements is smaller indicating harder braking (see Table 3.6).
male female Percentage of the deceleration events (<-2.5) 24.72 27.06 Percentage of the deceleration events (<-3) 8.44 5.87 0.00 5.00 10.00 15.00 20.00 25.00 30.00 Per ce n tage (% )
56
Figure 3.8: Percentage of different deceleration values based on signalised or non- signalised road elements