MIT researchers developing algorithms to predict more accurately which cars are likeliest to run red lights
Researchers at MIT are devising algorithms for more accurately estimating driver behavior at road intersections—i.e., whether or not an oncoming car is likely to run a red light—and validating them using real traffic data.
The team has introduced two classes of algorithms that can classify drivers as “compliant” or “violating”. These are based on i) Support Vector Machines (SVM) and ii) Hidden Markov Models (HMM)—two very popular machine learning approaches that have been used successfully for classification in multiple disciplines. However, existing work has not explored the benefits of applying these techniques to the problem of driver behavior classification at intersections, the team says.
A paper describing work will be published in the journal IEEE Transactions on Intelligent Transportation Systems. The team presented the work earlier this year at the 2011 IEEE Intelligent Vehicles Symposium in Germany.
In 2008, road accidents in the US caused 37,261 fatalities and about 2.35 million injuries. An estimated 45% of injury crashes and 22% of roadway fatalities in the US are intersection-related, according to NHTSA. A main contributing factor in these accidents is the driver’s inability to correctly assess and/or observe the danger involved in such situations.
Half of the people killed in such accidents are not the drivers who ran the light, but other drivers, passengers and pedestrians, according to the Insurance Institute for Highway Safety (IIHS).
This research is focused on developing algorithms that infer driver behaviors at road intersections, and validating them using naturalistic data. The resulting algorithms can be applied to either vehicle-based systems or infrastructure-based systems. Inferring driver intentions has been the subject of extensive research.
...More specifically, the modeling of behavior at intersections has been studied using different statistical models. These studies showed that the stopping behavior depends on several factors including driver profile (e.g., age and perception reaction time), yellow-onset kinematic and geometric parameters (e.g., vehicle speed and distance to intersection).
...This paper develops two novel classes of algorithms based on distinct branches of classification in machine learning to model driver behaviors at signalized intersections. It also successfully validates these algorithms on a large naturalistic dataset. First, it describes the driver behavior inference problem and the different factors involved in the decision making. Then, it introduces the two classes of algorithms, a discriminative approach based on Support Vector Machines (SVM), and a generative approach based on Hidden Markov Models (HMM), along with the traditional approaches that they are compared to. Next, it describes the implementation process of the different algorithms. Finally, it evaluates their performance on intersection data collected in Christiansburg, VA as part of the DOT Cooperative Intersection Collision Avoidance System for Violations (CICAS-V) initiative.—Aoude et al.
The first algorithm (SVM-BF) combines a Support Vector Machines classifier with a Bayesian filter to discriminate between compliant drivers and violators based on vehicle speed, acceleration and distance. The second, an HMM-based classifier, uses an Expectation-Maximization (EM) algorithm to develop two distinct Hidden Markov Models for compliant and violating behaviors.
The MIT team designed the algorithms to maximize true positive rates while keeping false alarm rates below a 5% threshold. They validated the two algorithms on more than 10,000 intersection approaches collected in Christiansburg, VA as part of the US Department of Transportation CICAS-V initiative.
Compared to three popular traditional approaches (time to intersection (TTI)-based; required deceleration parameter (RDP)-based; and speed-distance regression (SDR)-based), the MIT algorithms showed consistent and significant improvements, ranging from a minimum of 10% increase in true positive rates to more than 20% increase when issuing a warning 1 and 2 seconds in advance, respectively.
Compared to similar safety-prediction technologies, the group found that its algorithm generated fewer false positives. Jonathan How, the Richard Cockburn Maclaurin Professor of Aeronautics and Astronautics at MIT, says this may be due to the algorithm’s ability to analyze multiple parameters. He adds that other algorithms tend to be “skittish,” erring on the side of caution in flagging potential problems, which may itself be a problem when cars are outfitted with such technology.
The challenge is, you don’t want to be overly pessimistic. If you’re too pessimistic, you start reporting there’s a problem when there really isn’t, and then very rapidly, the human’s going to push a button that turns this thing off.—Jonathan How
How says “smart” cars of the future may use such algorithms to help drivers anticipate and avoid potential accidents.
If you had some type of heads-up display for the driver, it might be something where the algorithms are analyzing and saying, “We’re concerned”. Even though your light might be green, it may recommend you not go, because there are people behaving badly that you may not be aware of.—Jonathan How
How says that in order to implement such warning systems, vehicles would need to be able to communicate with each other, wirelessly sending and receiving information such as a car’s speed and position data. Such vehicle-to-vehicle (V2V) communication, he says, can potentially improve safety and avoid traffic congestion. Today, the US Department of Transportation (DOT) is exploring V2V technology, along with several major car manufacturers—including Ford Motor Company, which this year has been road-testing prototypes with advanced Wi-Fi and collision-avoidance systems. (Earlier post.)
The researchers are now investigating ways to design a closed-loop system to give drivers a recommendation of what to do in response to a potential accident and are also planning to adapt the existing algorithm to air traffic control, to predict the behavior of aircraft.
The research was funded in part by the Ford-MIT Alliance whose initial funding led to the seedling of the ideas in this paper, and the continued funding of Scientific Systems Company, Inc. (SSCI) and Le Fonds Québécois de la Recherche sur la Nature et les Technologies (FQRNT) Graduate Award.
Georges S. Aoude, Vishnu R. Desaraju, Lauren H. Stephens, and Jonathan P. How (2011) Behavior classification algorithms at intersections and validation using naturalistic data. Intelligent Vehicles Symposium (IV), 2011 IEEE doi: 10.1109/IVS.2011.5940569