Focus on... The DATASAFE project: Understanding data accidents for traffic safety

on the December 17, 2019

The project DATASAFE funded by the Data Institute for two years focused on understanding from real traffic data the behavior of traffic in the moments preceding an accident.
Traffic congestion is a major concern in the modern society in terms of loss of productivity, waste of time, pollution and city noise management. Two main quantities are used to describe traffic flow on a network: traffic density and average speed of cars. In the transportation literature, the graph that links the flow and the density is called fundamental diagram. The fundamental diagram is thought to be a description of the drivers’ characteristic behavior and is usually assumed constant. In this DATASAFE, we studied how the fundamental diagram depends on time and day of the week. In particular, we are interested in understanding if certain particular road conditions can be linked to generation of accidents.

A non-negligible part of the risk in vehicular traffic is due to individual behavior such as the driving style or the level of attention, which have a strong impact on the resulting traffic flow. It is necessary to understand if there are certain traffic situations that can cause the onset of these risks. During her Master 2 internship, Aleksandra Malkova worked to understand from data if there exists a particular link between speed and density of cars at which collisions are more likely to occur.

We used the raw sensor data collected (see Fig. 1, for an example) on the Rocade Sud in the framework of the Grenoble Traffic Lab (https://gtl.inrialpes.fr/status) using data for different days of the week and different day time.

The sensor data consists of these quantities collected every 15 seconds:
• Volume: Number of vehicles that passed the sensors position in the last 15 seconds
• Occupancy: The proportion of time that the sensory was “occupied” by vehicles in the last 15 seconds.
• Speed: Average speed of the vehicles that have passed the sensor position in the last 15 seconds.
Figure 1 Speed raw data represented according to day of the month VS time of the day, with color code from black-red for low speed to green for high speed.

The Rocade Sud is equipped with 300 sensors along 13 km in the direction Meylan - Rondeau. Moreover, we collected accident data from the DIR (Direction Interdépartementale des Routes) Centre-Est. These data contain the time stamp and the location of the occurred accident. Firstly, we looked at recognizing accident days by looking at the fundamental diagram for each sensor for each day, see the results in Figure 2. One can immediately notice that there is a minimal difference between days with and without accidents.

 
Figure 2 Fundamental diagrams. Left the day with an accident. Right the day with accident

However, if we look at the speed plots for those days (Figure 3) one can immediately spot a difference.

 
Figure 3 Speed VS time. Day with accident (left), day with no accident (right).

Then, we learned the road capacity of the Rocade Sud using a Bayesian learning technique based on particle filtering. In Figure 4, one can see that it is possible to learn the road capacity during a day were accidents are present as well as, during a day with no accidents.

Figure 4 Result for the day with a car accident (below), result for the day without a car accident (above). The 90% confidence interval in light-blue color, the learned capacity in blue, the beginning and the end of the accident is indicated by a dotted line


Report written by Aleksandra Malkova, research intern at Inria supervised by Maria Laura Delle Monache (NeCS) and Julyan Arbel (Mistis). Her internship was financed by the Grenoble Alpes Data Institute.
Published on January 9, 2020