Car detection in aerial thermal images by local and global evidence accumulation

Abstract

This paper introduces an approach to automatic vehicle detection from aerial infrared images of approximately 1 m resolution. On one hand, the extraction relies on a local description of cars, i.e. blob-like structures; on the other hand, because many other objects in urban areas have a similar appearance, the local model is extended by a more global description that incorporates knowledge about the appearance of cars as repetitive patterns in dense traffic situations or in filled parking lots. The implemented system is intended to detect vehicles independent of their current state (stationary or moving). Since vehicle velocity is not included as feature for detection, the system can be used as first module for vehicle tracking but also for analyzing traffic congestions or counting cars in parking lots. The only requirement to be met is that cars appear as repetitive patterns which consist of blob-like structures.