Rozantsev, ArtemChristoudias, Christos MariosLepetit, VincentFua, Pascal2013-03-072013-03-072013-03-072013https://infoscience.epfl.ch/handle/20.500.14299/90169We have developed a method for the detection of both generic flight neighbouring aircrafts and those on a collision course. Our approach employs a sliding window linear Support Vector Machine (SVM) classifier with a Histogram of Oriented Gradients (HOG) feature representation. An extension of this approach to the spatio-temporal domain is also considered and we demonstrate its advantage for the detection of aircrafts on a collision path. We evaluated our approach for the detection of both small rotorcraft and larger fixed-wing aircrafts in challenging video sequences. Our results show that aircrafts on a collision course can be detected more reliably than when assuming a generic flight path. This is very interesting in practice, since this case is of critical importance. We also show that our spatio-temporal approach improves the detection accuracy with respect to conventional single- frame approaches.UAV detection3D HOGaircraft detectioncollision avoidancetext::report