Page - 84 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 5. Distributed setup of sensors including 2 laser scanners and one Pico Flexx ToF camera around the workplace. The sensors we have chosen are not supposed to be used for object/human detection and localization per se, but mainly for distance measurement. Therefore, with both types of sensors,weneed toperformsomepost-processing inorder to be able to detect and perceive an approaching object. In order to have a safety-eligible sensory data analysis and decision making, we need to reduce the chance of false positive and false negative in our perception system. Safety- wise, perception scenarios with false negative (i.e., a human approaching the robot isnotdetected)are farmoredangerous compared to scenarios with false positive. In case of false positive, on the other hand, we may observe instances of unwanted robot speed reduction or even a complete stop, which is affecting the system performance but not the safety property. For instance, in case of ToF camera the following steps are being performed to robustly detect a moving object: • Filtering the depth image: it is performed by using various filtering method (e.g., median filter in both spatial and temporal domain) which mitigates the false detection. Filtering steps are shown in Figures 6 and 7, and resulting filtered depth image is shown in Figure 8. • Background image: recording filtered depth image at startup.Thebackground image is refreshed if there is no movement detected for a specific period of time (Figure 8). • Difference image: subtraction of background and cur- rent depth image (Figure 8 – Figure 10 = Figure 12). • Blob Detection in binary difference image (Figure 13). To avoid detecting changes produced by noise and also using the prior-knowledge of the size of an approaching object (e.g., human) we adjust the parameters of our blob detector (such as expected shape and size) in a way todetectonly the intendedmoving targets.Whenat least one blob is detected, it means that there is a movement in the workspace, and therefore we can proceed with the next two steps. • Masking the original depth image: binary difference image is used as a mask in order to have real depth data of each pixel of the blob that is assumed to be a moving object (Figure 11). • Final depth information of detected moving object: is a result of using median value of depth info from the masked image. Higher importance is given to closer distances that still have a smaller covering area in the depth image, such as an intruding arm of a human. Fig. 6. Original depth images Fig. 7. Filtered depth images Fig. 8. Final background image Fig. 9. Original images with hu- man Fig. 10. Final fil- tered depth image Fig. 11. Masked original depth image Fig. 12. Difference image Fig. 13. Blob de- tection in BW diff. image For the laser scanners, which provide 2D data (scan a plane or a cross-section out of the 3D space), the process of extracting distance of a detected object and its coordinates is aligned with the one of the ToF camera: • Background data: resulting background data is median filter applied on temporal domain of data collected in initialization step. • Difference data is calculated as subtraction of back- ground data and current data every time stamp. • Movement in the workspace is detected if the percent- age of not moving points is less than 98.5%. • Transformation of depth data from the laser coordinate system to the robot’s coordinate system is done using Euclidean distance, taken into account the fixed position of the laser scanner relative to the robot. Every time stamp we have the result of our sensor fu- sion as the final danger zone. From each sensor, regarding distance of a human, or any other moving object in robot’s workspace, it is decided in which danger zone the detection happened, and the final danger zone is the worst case of all three. Measuring the separation distance between the ob- ject/human and the robot, in constant speed setting situations with worst-case value taken into account, it is ensured that the robot system never gets closer to the operator than the protective separation distance [12]. 84