Page - 19 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 2: Vehicle with measurement setup and DGPS-receiver IV. EXPERIMENTAL VALIDATION Ourrealisticdataset shows theperformanceofourVOona track through a forest. It contains GPS data as well as images during a drive of a truck on a logging road. The vehicle used for the measurement is further discussed in Section IV-A and sample images of the road can be seen in Section IV-B. A. Setup A small truck, equipped with a stereo-camera system, was used for the measurement. The cameras are mounted on the back of the driver‘s cab via aluminum profiles and magnets. This mounting position guarantees a good viewpoint back- ward without having unwanted objects within the field of view. In addition to the cameras, a DGPS-unit (Differential Global Positioning System) is used for ground truth although the signal strength lacks inside the dense forest. The vehicle and its measurement setup is shown in Fig. 2. The cameras are mounted parallel on an aluminum profile at a distance of approximately one meter. The 12V battery of the truck powers both cameras inside the wired box. Two Gigabit Ethernet cables facilitate the data transfer of the stereo-camera system, which operates at 10Hz. A higher sample rate of the cameras is unsuitable due to the high computing time of the VO. We used the following sensors and devices: • 2× JAI GO monochrome-cameras (JAI GO-5000M- PGE) with a maximal sampling rate of 22Hz with the full resolution of 2560×2048 pixel • 1× DGPS-system with open sky localization error of ca. 2cm/0.1◦ • 1× Xsens MTi-30 IMU with 400Hz sampling rate (additional sensor for further experiments) • Lenovo Thinkpad S540 with Intel Core i7-4510U CPU @ 2.00GHz and 16GB RAM Windows 7 Professional SP1 - 64 Bit The JAI GO cameras allow a maximum resolution of 2560×2048 pixel. Due to lots of bumps on the logging road, the long exposure time of the cameras might blur images at darker areas of the forest. Therefore, we use 2×2 pixel (a) (b) (c) (d) Fig. 3: Sample images of the driven logging rode (a) Left stereo image (b) Right stereo image Fig. 4: Stereo image pair with branch occlusion binning and a resulting resolution of 1280×1024 pixel to decrease the exposure time. The resolution of the images is furtherdecreased to640×512pixelbysoftware to reduce the computing time of feature detection and description. After the decrease of the resolution, a rectification of these images is also done. B. Experiments Our dataset contains two different drives of the presented vehicle on a logging rode and illustrates a realistic perfor- mance of our VO system. Figure 3 shows some road sections of our scenarios. Widespread areas and overexposed images may result in an inhomogeneous distribution of features, which is a big challenge for the VO. The first scenario of our dataset is a 3×75m long test drive on the part of the logging road, which is shown in Fig. 3a. In this dense forest area, our proposed VO demonstrates its robustness against overexposed images and occluded cameras like shown in Fig. 4, where a branch occludes the left camera entirely. The implementation is robust enough to handle such situations and still estimates a valid pose. The results of our test drive are presented in Fig. 5. The starting point is marked with a circle. Due to the low signal quality of the GPS, the reference position exhibits some inconsistencies. The plotted coordinate system is the one of the GPS withX pointing to East,Y to North andZ upwards. 19