Autonomous Real-time Vehicle Detection from a Medium-Level UAV

Detection vehicles from a mid-range (40 m target altitude) aerial platform (using Haar classifiers). It's still somewhat relevant because it detects vehicles from a perspective aerial view, from an acute camera angle. Claims to be able to detect both static and moving vehicles, invariant to orientation, color, type and configuration. It uses a 2-stage detection: the primary stage is detection using the trained Haar classifiers, the secondary stage is a verification based on altitude driven vehicle size constraints. They trained 4 separate Haar classifiers based on the orientation of the vehicle (at 0, 45, 90, 135 degrees - no more is needed as the vehicles are assumed to be symmetric).

The method was mostly tested in rural areas. It had relatively good results, though it detected some false positives and also some vehicles were missed.

kundralaci / JAMResearch

Autonomous Real-time Vehicle Detection from a Medium-Level UAV #133