Intersection Navigation Algorithms for Autonomous Vehicles (AV)
In 2012, SwRI developed a framework that would allow an autonomous (driverless) vehicle to safely and appropriately navigate dynamic intersections where other vehicles and pedestrians are present. This framework included a behavior architecture to manage the various stages of approaching and traversing an intersection, as well as vision-based algorithms to detect objects in or near the intersection and estimate their states.
SwRI developed a generic intersection behavior architecture that used the Robot Operating System (ROS) and Boost Statechart Library to implement a modular, extensible behavior framework with behavior states for approaching and traversing a generic intersection. A vision-based, cascading-window classifier was also developed that can be trained to detect and classify various objects, including vehicles and pedestrians. The algorithms implement a combined HOG/Haar-based classifier that achieves state-of-the-art performance. To estimate the velocities and accelerations of the detected vehicles, an optical flow-based optimization technique tracks the motion of key features of the vehicles over time. Using the estimated positions, velocities, and accelerations, the behavior state machine allows the vehicle to safely traverse the intersection.
The developed algorithms for autonomous intersection navigation have been integrated and tested on SwRI's Mobile Autonomous Robotics Technology Initiative (MARTI) autonomous ground vehicle (UGV) platform and evaluated at SwRI's on-site test intersection. The MARTI® vehicle successfully managed various intersection scenarios with other manned vehicles present, including cases where other vehicles violated intersection rules and MARTI was forced to stop to avoid a collision.