Spatio-temporal fusion of object segmentation approaches for moving distant targets

Object segmentation can be an important step between object detection and tracking. Especially in image processing applications, where object detection is difficult due to high object distance, camera motion, or noise, the detection result might not be precise enough to robustly initialize tracks and perform multi-target tracking. In this paper we present the detection and segmentation of moving objects in image sequences coming from a small Unmanned Aerial Vehicle (UAV). Based on the detection and tracking of local image features, camera motion is compensated and independent motion created by moving vehicles and people on the ground is found. By clustering the independent motion vectors initial object hypotheses are generated which may be affected by over- and under-segmentation. For improvement, several object segmentation approaches are introduced and tested. Best results are achieved with a spatiotemporal fusion of some approaches. Both spatial and temporal information is provided by the local image features. The object segmentation approaches and the fusion methods are evaluated for their completeness and precision.