Seeing Around Street Corners: Non-Line-of-Sight Detection and Tracking In-the-Wild Using Doppler Radar

In this section, we describe the recorded data set and how it was generated. Data Set Description Our data set consists of 21 different scenarios including 2-8 repetitions each. In total the data set amounts to 100 sequences including over 32 million radar detection points. For our experiments we focus on two kinds of road users: pedestrians and cyclists. This choice is reasoned by the high effort for data recording and labeling. Also, the detection of bigger, faster, and more electrically conductive objects such as cars is in general much easier with radar systems. Therefore, we postulate that our results can be generalized and should even improve for such objects. For both kinds of recorded road users we measured and labeled roughly the same amount of sequences. We are mainly interested in the data from the two radar sensors installed in the front bumper of the car. Furthermore, we also record lidar data for reflector geometry estimation, camera images for scene documentation, as well as global navigation satellite system (GNSS) and inertial measurement unit (IMU) data for annotation purposes. The number of different scenarios was 21 for cyclists and 20 for pedestrians due to erroneous data in one scenario for all pedestrian repetitions. All measurements were recorded over five different days with five different people as rider and pedestrian, respectively. In each scenario we had the observed road user start at the ego-vehicle and drive alongside the reflector until out of range, cf. Fig. 1 . Then, the cyclist or pedestrian turned and approached the test vehicle again on a similar trajectory. This allows us to effectively double our training data per scenario and, also, eases the quality checking procedure as the road user always start in the field of view of the documentation camera. A complete data set overview is given in Tab. 1, where each scene is shortly described and the amount of repetitions for both object classes is indicated. Furthermore, in Fig. 3 some example pictures of used reflectors and road users are depicted. In order to provide a better impression about the distances between the test vehicle, reflectors, and VRUs in the different scenarios, we plot their distribution in Fig. 4 . We use the distance of labeled NLOS detection points for the estimation. While this approach yields an comprehensible way for determining the distances, it is affected by the irregular distribution of radar detections which is generally higher for objects in close distance. Thus, the shown graphs indicate rather a lower bound than absolute values of the actual distances. Moreover, a deeper insight in the data distribution is given in Fig. 5 , where we compare the distributions of direct-sight and NLOS detection points on the VRUs in our data set averaged for every individual scenario. Despite the high number of NLOS detection points for the observed road users, the reduction compared to the direct sight object are obvious, ranging from a factor of roughly 2× up to 50× less detections. In comparison, the average total number of detections per measurement scene (3.2 • 10 5 ) is three orders of magnitude higher than the combined number of direct-and non-direct sight detections, indicating the immense amount of clutter, the tracking algorithm has to cope with. * Equal contribution.