Recognition of Vehicles in 3D Point Clouds

In both civilian and military contexts it is sometimes necessary to search for people and objects with an unmanned aerial vehicle that can carry a visual, infrared or laser sensor. The laser sensor generates point clouds of the environment and one important issue is the requirement of resolution for users to recognize targets. An experiment was carried out with the aim to identify how many points that are required for the recognition objects. Five civilians and five military vehicles where used as target object, that where filmed by a simulated UAV that flew in an arc around them. The participants' task was to watch these video sequences and to try to recognize which vehicles that where visualized and also to estimate how confident they were in their answers (0-100%). The results show that the ability deteriorate as the number of dots decreases and that the variation between vehicles are large. The results also show that the participants become more precarious (lower confidence estimations) and that it takes longer time to respond the lower the resolution is. The results also show that there was no difference whether the participants had civil or military background regarding recognition of military vehicles, i.e. both groups performed equally well. The conclusion of the results show that it takes a resolution about 2.8 dots/m2 to recognize vehicles that have a distinct appearance, or do not have other vehicles that are similar, which is about 100 actual dots. To recognize vehicles when there are multiple vehicles with similar appearance and size required a resolution about 25 dots/m2, which is about 500-1000 dots. The conclusions apply provided that the information is collected from a moving platform like a UAV with similar speed as used in this experiment and for participants with no prior training for this situation.