IP2: Exploiting Repeated Data Acquisitions for Improved Long-term Monitoring Capabilities
Lead: Cyrill Stachniss
PhD student:
Fields and orchards are monitored repeatedly to assess the status quo and to trigger management decisions. To assess growth stages or compute phenotypic parameters of plants, knowledge about the plant geometry and further semantic information is key. Thus, estimating 3D geometric and semantic models of plants plays a key role in the automated status quo assessment. Most sensor-based monitoring systems today assume that the sensor platform is observing a new scene whenever starting the data acquisition process. Few approaches take prior maps from previous data acquisitions into account to extend models or automatically track changes such as growth over time. The fact that the same scene or objects is re-observed, potentially after undergoing some changes, is often not exploited to its full extent.
Central Questions:
-
- How to build accurate plant models and exploit the fact that the same, but growing and changing objects, are being monitored repeatedly to,
- improve and achieve consistent modeling in the spatial and temporal dimensions (4D)
- estimate semantic information more precisely and consistently over time, and
- improve the involved learning approaches in a self-supervised or unsupervised way by exploiting prior knowledge about the scene?
- How to build accurate plant models and exploit the fact that the same, but growing and changing objects, are being monitored repeatedly to,
We will develop new approaches and will extend current systems for robot mapping/SLAM, filtering approaches for dealing with change, as well as contrastive learning in combination with deep neural networks to tackle the three research questions.