Page - 158 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Operations 1 and 6 are specific to multi-camera sys- temsandare thereforedescribedinmoredetail in the following sections. The remaining CPU/GPU per- formance on the hub is available for AI and deep- learningapplications. 3.Synchronization Imaging systems that rely on multiple active sen- sors inevitably requirea synchronization. In thecase of the multi-ToF platform, synchronization serves two purposes: On the one hand, it avoids interfer- ence effects between the sensors, and on the other hand, it simplifies the registrationof thepointclouds produced by the individual frontends. The hub can synchronize multiple frontends by using a hardware triggertostart theacquisitionofindividualfrontends. This could be done in a round-robin scheme or by triggering opposite sensors at the same time to avoid interference. 4.Pointcloudregistration Each ToF-sensor frontend produces a 2D depth map which can be converted into a 3D point cloud. A consistent 360° view of the environment necessi- tates the registration of these individual point clouds in a common world coordinate system. Through an extrinsic calibration all sensors of the ring can be combined inasinglepointcloudwhichcanbe trans- formedinarobotorworldcoordinatesystemgivena known positionof the robot’s joints. 5.Advantagesandperformance Withtoday’sToFtechnology,camerasarecapable ofdetectingobjectswithhigh frameratesand lowla- tencies. Active lighting ensures that data quality is independent from ambient conditions to a high de- gree. The exact distance measurement accuracy is dependent on the target’s reflectivity and distance, but the user can expect a relative accuracy of 1% basedon thedistance. Considering the system’s performance in the con- text of machine learning, ToF cameras provide use- ful additional information compared to, for exam- ple, 2D RGB cameras: Objects can be more easily spatially separated using the 3D point cloud and the corresponding IR greyscale image can be employed when training a network. Training labels can easily be transferred between the four ToF channels (X, Y, Z,andamplitude)atpixelprecision. Asaresult,ToF cameras reveal more information about the observed Figure 2. Depth image (left), with red indicating smaller and green larger distances, and the corresponding IR greyscale image (right). scene, but labelling the data does not require addi- tional effort. The recognition performance of deep learning algorithms in particular benefits from an in- crease in theamountofavailabledata. 6.Conclusion In this paper we have presented a hardware plat- form which uses multiple ToF Sensors and a central processing hub to generate a high-resolution point cloudaroundanautonomousmachinewhichenables collaborative and safety functions. Further work will include a synchronization of multiple machines working in close proximity using a clock synchro- nizationmechanismoverstateoftheartwirelesscon- nectivityhardware. References [1] U. Behrje, M. Himstedt, and E. Maehle. An au- tonomous forklift with 3d time-of-flight camera- based localization and navigation. In 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), pages 1739–1746. IEEE,2018. [2] S. Kumar, S. Arora, and F. Sahin. Speed and sepa- ration monitoring using on-robot time-of-flight laser- ranging sensor arrays. In 2019 IEEE 15th Interna- tional Conference on Automation Science and Engi- neering (CASE), pages 1684–1691. IEEE, 2019. 158