Seite - 54 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

• Path extraction: A path must be computed which guides the vehicle between the traffic cones to the destination and marks the drivable area. • Motioncontrol:Themotioncontrolleruses thedrivable area detected between the cones, computes an optimal trajectory, and dispatches the actual motion commands for steering and accelerating the car. In Fig. 4 one can see an overview of the messaging system among the ROS components. It shows the connection of the actuators to the sensors, as well as the processing steps needed to achieve desired control sequences. Some of them are implemented as nodelets to minimise messaging delays. In the following section techniques used are described in more detail. IV. APPROACH The EKF-SLAM implementation used is based on Mac- sek’s [6] Master’s thesis. The filter uses perception of cone measurements with uncertainties and generates a map of the cones in 2D. It also determines the car’s own position on the map, computes its position relative to previous laps, and performs motion control. It is necessary to predict the vehicle’s pose ahead of time in order to send proper motion commands, since actuators have a considerable reaction delay. For sensing, it is vital to detect traffic cones in the environ- ment and estimate their location relative to the sensors. The known location of the sensors on the vehicle enables one to create a global map of landmarks for the trajectory-planning algorithm. In the first development phase, TUW Racing used Pio- neer 3AT in a simulated and real world scenario in order to generate measurements, try out sensors and plan algorithms, as shown in Fig. 5. A. Traffic cone detection Becauseof thevarietyof sensorsused (lasersandcameras) various techniques have to be implemented accordingly. Two approaches are used to perform cone detection and position estimation on the stereo camera images. In the first approach, a depth image based on a block-matching algorithm is used. This allows for extraction and removal of planes in a 3D data set. The track floor that is geometrically known can be pre-segmented, and cone candidates appear as isolated objects. These candidates are then further evaluated using classical image processing steps on the estimated location in the image plane, similar to the approach of Yong and Jianru [11]. However, a block matching algorithm cannot be easily adapted to use the known structure of the environment for improved perception of target objects. Moreover, the computation effort and thus hardware requirements would be unnecessarily high, since one would have to operate on an entire image. In the second approach, cones can be detected in both images separately. The algorithm computes the disparity, the z-depth and consecutively the 3D position of the cone, based only on the UV center points of the same object determined in the left and right images. In both approaches, the program determines the colors heuristically with RGB color thresholds. If the estimated center point is within a white or a black stripe, the color sample point is moved downwards until another color match is determined. In addition, the planar laser scan is searched for point clusters of distances that match the radius of the cone at the height scanned. If the total distance spanned by coherent scan points is above this radius but below the total radius of the cone base, the algorithm considers the center point between the cluster to be the center point of a cone. In order to avoid detection of obstacles other than cones, the detector filters long connected components such as walls. The detection results of both algorithms are then fused with a feature-based EKF-SLAM [6]. Thus the team benefits from the advantages of both of the detection approaches, because the fusion algorithm respects the reliability of each detection result and takes the more reliable detection result intogreateraccount.Forexample,distanceestimationathigh distances is less accurate at image-based detection, due to the fact that pixel disparity is smaller. On the other hand, laser- based detection cannot reliably detect cones that are behind other cones. B. Mapping For TUW Racing, it is essential that the mapping com- ponents be able handle failures in the detection component. The camera or the laser could have short periods of blindness due to sunlight or uneven road conditions. In these cases a cone might be missed. Since cones are guaranteed to be at most 5 m apart, missing ones can be filled in. Because all cones on the left side have a different color than those on the right, heuristics can fill in cones missing from one side based on the cones observed on the other side. Since the trafficcones themselvesare thebest reference for position and orientation on the track, the team uses an EKF- SLAM that maps them and corrects the vehicle’s position based on their perceived change of position. An additional challenge here is that the same track sections are driven ten times throughout the competition. Thus, the mapper must recognize previously mapped cones by ID, even though they are of the same color and shape. An index counter from the starting line is insufficient due to potential perception failure. Nor does the vehicle have a full 360deg line of sight. In a narrow left turn, due to the vehicle’s pose in an optimal trajectory, it is possible that only the outer cones might be seen. If the mapping algorithm performs corrections to positions on the map on those right lane cones, the left lane cones must also be updated automatically. C. Trajectory Planning From the mapping one can derive a path which lies in the center of the travel area; path-planning is therefore obsolete. To start, a normal vector is projected from the center point between the first left and right cones forward. The first cones 54