Honda Intersection Interaction Project
| Introduction | Simulator | Vehicle Tracker | Experimental Results |
Introduction
The underlying goal of this project is to minimize the severity of a side impact collision occurring at a standard right angle intersection as seen in the figure below.
In this scenario the host vehicle uses a lidar sensor to track an oncoming vehicle referred to as the other vehicle. The basic idea of the project is to track the other vehicle and determine if the host vehicle should accelerate, brake, or maintain constant velocity during the last 100ms before a collision occurs to minimize the severity of the collision on the host vehicle.
Simulator
In an effort to greatly reduce cost, OSU developed a vehicle intersection interaction simulator to imitate real world collisions. The simulator provides the user with emulated lidar returns from the other vehicle and gps data from both vehicles. The data from a simulated collision can then be used to test a vehicle tracker over several Monte Carlo runs. To make the lidar returns realistic zero mean Gaussian white noise is added to the range measurements and a certain percentage of the returns are dropped during each run. The percentage of points dropped and variance of the range noise are adjustable; however, most simulations are completed with values based on experimental data from Honda.
Vehicle Tracker
The current vehicle tracker can be summarized as follows. First, raw lidar data is separated into clusters of returns which are believed to belong to the same object. Next, the clusters are sent to a shape fitting algorithm which attempts to label each cluster as a line, corner, or blob.
Ideally all vehicles would be labeled as corners; however, due to vehicle orientation or various other parameters that affect lidar returns, this is not always the case. Consequently, the vehicle tracker attempts to classify clusters that fit a line as the front or side of a vehicle. If this is done, a psuedo-corner can be created and tracked. Incidently, simulations performed thus far have shown that in order to track vehicles with the precision necessary in this project it is imperative that the points on a vehicle being tracked are stable, i.e. a corner.
A tracking filter is then attached to each potential vehicle. Thus, at each consecutive time step a data association algorithm must determine which current lidar clusters correspond to existing filter tracks. Moreover, the data association algorithm also determines if additional filter tracks need to be initialized or existing filter tracks need to be deleted.
The current tracking filter used consists of a dual UKF (Unscented Kalman Filter) system. In this system the master UKF is responsible for tracking the x and y position, velocity, and acceleration of the vehicle while the slave UKF attempts to update the process noise.
Once vehicles are tracked their trajectory can be predicted. Moreover, if it is predicted that a collision will occur the potential severity of the collision can be calculated.
The following video portrays one run of the host vehicle traveling along the x-axis, tracking the other vehicle traveling along the y-axis. In the video the red rectangle, green rectangle, blue Xs, and black X refer to the host vehicle location, actual other vehicle location, lidar returns, and predicted other vehicle corner location respectively.
The following figure is the resulting normalized predicted collision severity for multiple Monte Carlo runs of the above scenario with a Gaussian noise variance of .1m and a .1 probability of a lidar return being dropped.
