Page - 93 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure 2. Template of the GNSS antenna (left) and best position found in the right image. proved due to bad GNSS quality or if there are an- tenna positions missing due to GNSS outages and need to be derived from VO solely. It is important that there are GNSS positions available before and after an outage has occurred to ensure that the VO trajectory is correctly oriented and placed w.r.t. the definedcoordinatesystem(UTM).Oneissuethathad to be solved is to mask out all areas in the stereo im- ages covered by the tractor or milling machine itself as thiswoulddeteriorate theVOprocesssignificantly and oftencausedcomplete fail of theVOsolution. The GNSS antenna is mounted straight above the positionwherethecableis laidataknownheightoff- set. It is therefore necessary to determine the exact 3D position of the GNSS antenna which can move relatively to thecamerasystem. This issolvedbyau- tomateddetectionandmeasurementof theGNSSan- tenna in both stereo images using an advanced tem- plate matching process (see Figure 2). GNSS posi- tions of good quality are introduced as ground con- trol points (GCP) in the adjustment process. If the GNSSpositionis inaccurateorevenunknownthe3D position of the antenna is reconstructed by using the stereo imagemeasurementsof that event. The VO workflow has been implemented by us- ing the Agisoft Metashape v1.5.4 software [1] and its Python scripting capabilities. Importing of the images,correctingimagedistortions,applyingimage masks, feature point extraction, image matching and photogrammetric triangulation are fully automated bythescript. Thereconstructedcamerapositionsand derivedGNSSantennapositionscanbeinspectedus- ingtheMetashapeGUIandQCreportingtools. If the expected accuracy level has been reached the GNSS antennapositionareexportedtoanASCIIcoordinate file. 3.ResultsandConclusion The VO workflow has been tested with data from aLAYJETproductionruncollectedinGermany. The roadpasses throughaforestwhichcausesbadGNSS signal quality and a low number of visible satellites (inaddition theRTKcorrectionsignalhasbeenlost). Theestimatedpositionaccuracy is thereforestrongly reduced to about± 2m. The photogrammetric bun- dle adjustment uses the GNSS solution as approxi- mate positions and the well-defined relative geome- try of the stereo pairs and consecutive stereo models to improve theaccuracyat leastbya factorof10-20. Figure 3 shows the reconstructed trajectory of the stereo rig and the derived GNSS antenna positions for 50 trigger events (section of 100m length). The sparse 3D point cloud generated during the VO pro- cess can be easily improved by an additional dense matching step which allows to inspect the environ- mentandcable routingmoreclosely. First test evaluations have shown a throughput of about10stereomodelsperminute(50minperkm)on an Intel workstation equipped with 16GB RAM and aNVIDIAGeForceGTX1660TiGPUwhichshould allow for overnight processing of the data collected ononeday. Figure 3. Path of the LAYJET tractor reconstructed using aVOworkflowimplemented inMetashape. The VO system described in this paper derives absolute orientation angles solely from GNSS posi- tions, which is straightforward for the heading angle but also works for roll and pitch as long as there are turns included in the trajectory. In case of exactly straight road sections the pitch angle is not defined and has to be set to zero. As the road cross profile inclination can be assumed to be in the range of± 3degthiscausesa lateralpositionerrorofupto15cm (GNSSantennaheight∼2.5m). ForlongerGNSSoutagestheestimationoftheroll angle degrades with distance which can lead to sig- nificant height errors in case of steep descents. It is therefore recommended to integrateanadditional in- clinometer tomeasurerollandpitchanglesatapreci- sionofabout±1deginanextversionoftheLAYJET VOsystem. 93