Page - 55 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 4. ROS nodes and message subscriptions in rqt graph Fig. 5. Edge8 in a simulated environment using GazeboSim. to the left and right of this vector are identified, then the center point identified, and a vector projected forward again. Since the left and right sides can be distinguished, a situation where the cut to the left of the vector which contains a right- side cone, or vice-versa, is manageable. The vector is then rotated accordingly so that the sides are split correctly. But following the center path would not be very efficient, so the model predictive [4] motion controller computes its own optimal local trajectory. The MPC [10] used optimises the trajectory using the latest perception state and incor- porates the mechanical constraints of the vehicle’s motion model, such as maximum linear acceleration and decelera- tion, actuator delays and disturbances resulting from vehicle dynamics, e. g . sideslip angle. V. RESULTS In order to build a foundation for simulation of the race car in software that integrates well into ROS, TUW Racing extended its approach to the Pioneer 3AT and used gazebo simulations to test marker mapping, odometry corrections based on simulated IMU data, trajectory planning, as well as a basic self-developed motion controller. Fig. 5 shows the robot following a narrow turn in the simulation, with obstacles of width and height comparable to those of cones on the track. It successfully navigates a 180deg turn, detects shapes based on simulated laser scans and determines the center path accurately. In Fig. 6 results with a cone detection algorithm on a still from a video of last year’s autocross race at the FSG Fig. 6. Cone detection on still of last year’s race recorded using a GoPro camera are shown. The cones detected in the image are enclosed by yellow rectangles. The algorithm is based on a trained cascade classifier using a histogram by an oriented gradients feature descriptor. Photos of the specific cones seen in the image, from angles of 0deg horizontal to about 40deg turned downwards, served as training data. The algorithm turned out to be computationally expensive and susceptible to variations in lighting and weather conditions. With the trained algo- rithm on last year’s video we achieved an unstable detection rate of 30% to 90% depending on the track section. The algorithms explained above, evaluated on recordings with the stereo camera, detected 75% of the traffic cones in the image regardless of lighting. Most of the missed cones were affected by very strong shadows cast by large nearby obstacles, wear of the cone itself or cones that partially covered each other. The location, color and shape of the cones, the point of view, background, the vehicle itself and lighting are very similar to the scene in the image. Evaluation of detection on a recording of last year’s race has shownthatdifferent light conditionscanbeaproblemfor the car’s visual sensor. Extreme light exposure to the camera or even driving from shade into sunlight can be disturbing, as the camera needs some time to adjust to the brightness. Similar issues emerged in evaluation of the stereo camera mounted on the car driving a short sample track. In Fig. 7, the camera adjusts to increasing brightness using integrated exposurecontrolwhen thecarentersa sunlight-drenchedpart of the track. As the more distant cones become visible by the aperture’s adjustment, the cone marked with the orange 55