Page - 20 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

X in m VO GPS −50−40−30−20−10 0 10 20 30 40 −70 −60 −50 −40 −30 −20 −10 0 Fig. 5: Scenario 1 – Comparison of the estimated trajectory with GPS As shown in Fig. 5, the estimated pose of the first 75m fits very well with ground truth. The estimated trajectory of the return slightly deviates from the GPS reference. The inaccurate estimation of the orientation happens due to occlusions of the left camera like it is shown in Fig. 4. However, our robust VO prevents a total failure and still allows a valid but slightly inaccurate pose estimation via using images of the right camera instead. The third track of the logging road is estimated well as a straight line again. Using the mentioned laptop, the computation time of our off-line VO of this scenario is about 0.529s per stereo pair. This timeduration is increaseddue to theocclusionof the left camera, which acquires the additional processing of the right image instead of just the left one. This problem especially happens at the return of the vehicle because the cameras are mounted on the back of the driver‘s cab. The second scenario of our dataset is a 3×2169m long drive of the presented vehicle on a logging road. This scenario should deliver an answer about the drift behavior of our implemented VO. Figure 3b represents the first image of this sequence. The results are shown in Fig. 6. The estimated trajectory is inconsistent with the ground truth and just the first 2169m long loop can be identified somehow. Then, the trajectory continuously deviates from the driven track. If we look closely at the start of Fig. 6, it shows that distances are estimated too large in general. The whole trajectory seems to be scaled compared to the original track. For a better understanding of the results, it is helpful to further investigate the 3D-trajectory illustrated in Fig. 7. Referring to the estimated VO path of this figure, from the beginning the truck starts to move downwards and also to twist sideways. This results in a distorted trajectory instead of a more or less planar movement of the truck. The explanation of the occurrent problem can be found with a closer look at the features, which are used for the pose estimation. Figure 8 shows the detected A-KAZE features of Fig. 3b, and the sweeping area only contains a few key points. Most of them are found at the treetops in the upper half of the image. In the worst-case scenario, for example if all trees have the same height, all features are just on one line X in m VO GPS −1000 −800 −600 −400 −200 0 −200 −100 0 100 200 300 400 500 600 700 Fig. 6: Scenario 2 – Comparison of the estimated trajectory with GPS Y in m X in m VO GPS −1000−800 −600 −400 −200 0 200 −400 −200 0 200 400 600 800 −1000 −800 −600 −400 −200 0 200 Fig. 7: Scenario 2 – Comparison of the estimated 3D-trajectory with GPS instead of being well distributed in the image. The outcome of this is an ill-conditioned pose estimation and hence an inaccuratelyestimateddistanceandpitch-angle.Theproblem of this scenario is that image positions hardly change by a further increase of the distance. However, as shown in Fig. 9, the yaw angle can be estimated well because a planar rotation definitely changes the image positions of these features. The figure clearly illustrates every turn of the track and the good consensus of the yaw angle for each loop. Just some minor deviations due to different drive behavior and drift can be seen. This means that Fig. 9 shows the potential of our implemented VO for applications which mainly rely on a good estimation Fig. 8: Features of two left images used for pose estimation 20