Spatio-temporal Data Association for Object-augmented Mapping

Traditionally, visual SLAM methods make use of visual features for mapping and localization. However, the resulting map may lack important semantic information, such as the objects (and their locations) present in the location. Since the same objects may be detected several times during the mapping phase, data association becomes a critical issue: objects viewed from different angles and in different time instants must be fused together into a single instance on the map. In this paper, we propose Spatio-temporal Data Association (STDA) for object-augmented mapping. It is based on expected similarities between consecutive frames (temporal association) and similar non-consecutive frames (spatial association). The experiments suggest that our system is capable of correctly fusing together multiple views of several objects, resulting in only one false positive association in more than 130 detected objects across several datasets. The results are competitive with the state-of-the-art. We also generated object location ground truth annotations for 3 simulated environments to foster further comparison. Finally, the annotated map was used for an object fetching task.