Agent Localization and Inference of Dynamic Environments

Providing a team of mobile agents with reliable knowledge of their locations and of salient features of the environment (map) is a fundamental requirement for autonomous vehicles and mobile robots. Our goal in this project is to develop improved models and inference methods for the simultaneous localization and mapping (SLAM) problem in the challenging but important case where the environment may change over time. We propose to describe the different parts of the environment by so-called extended objects (EOs). Although EOs appear to provide a better compromise between accuracy and scalability than existing map models, they have not been used for SLAM so far. For a statistical characterization of the EOs, we propose a new combination of random finite sets (RFSs) and Bayesian nonparametrics (BNP). Our approach is to describe the time-variant kinematic states of the agents and EOs such as their positions by RFSs and the time-invariant attributes of the EOs such as their geometric extents by BNP marks that are associated with the RFS elements. Building on this RFS BNP model, we furthermore propose a new SLAM inference methodology that is able to leverage an inherent class structure of the environment for improved performance or reduced complexity. This methodology does not presuppose prior knowledge of the classes; rather, the number and properties of the EO classes are continually learned, and statistical beliefs of the EOs` class affiliations are continually estimated. We expect that these affiliation beliefs will facilitate and improve inference of the kinematic states of the EOs and agents. The main goal of the project is to develop sequential SLAM inference algorithms based on the new RFS BNP model. For SLAM scenarios with multiple agents, we will devise both centralized and distributed (decentralized), cooperative algorithms. We expect that the developed methods will outperform existing SLAM methods with regard to accuracy, computational efficiency, robustness, scalability, and/or versatility. Potential applications of our results include autonomous driving, intelligent transportation, surveillance, logistics, assisted living, exploration, farming, search and rescue, and environmental monitoring. The proposed research will be supported by collaboration partners based at research institutions in Austria and the United States. The participating researchers have extensive competencies and track records in relevant scientific fields.

In today's era of autonomous cars, drones, and mobile robots, providing reliable knowledge of the positions and other properties of moving objects is an important aspect of situational awareness. Accordingly, a major focus of the FWF project "Agent Localization and Inference of Dynamic Environments" was the problem of multi-object tracking. Multi-object tracking aims at estimating the states-such as positions and velocities-of moving objects over time, based on measurements provided by sensing devices such as radar, sonar, lidar, or cameras. This problem is important in a wide range of applications including air traffic control, maritime surveillance, autonomous driving, environmental monitoring, robotics, security, and biomedical analysis. However, a major difficulty is that in addition to the states of the objects, also their number and their association with the sensor measurements are usually unknown. We addressed these challenges by developing new multi-object tracking methods that employ advanced statistical inference techniques to achieve excellent performance at moderate computational cost. A substantial result of the project was a new method for "simultaneous localization and mapping" (SLAM), which jointly estimates the positions of multiple moving objects and certain features of the surrounding environment (defining the "map"). Another noteworthy result was a new methodology for "class-aided" multi-object tracking, which leverages an unknown class structure of the objects for improved tracking performance. Further results include methods allowing the fusion of sensor measurements with contextual information or with the output of a classifier. A significant part of our research on object tracking emphasized distributed methods for use in decentralized sensor networks. Distributed methods have the advantage of not requiring a central data collection and processing unit or communication between distant sensors. For single-object tracking, we devised a distributed method that uses an advanced consensus technique for fusing the statistical information provided by the sensors. For multi-object tracking, we devised distributed methods in which the statistical information is represented by random finite sets. We also devised distributed statistical techniques for associating throughout the sensor network the identities of objects. Another line of our research was the estimation of the optical flow in image sequences, which is an important problem in image processing with a wide range of applications including autonomous navigation and medical diagnosis. We proposed a unified statistical methodology for optical flow estimation that is based on a variational analysis. Our approach supports ultrasound-specific statistical models and is thus well suited to ultrasonic imaging. The results of this project were published in a book chapter, in 13 papers in high-quality journals, and in five papers in the proceedings of international conferences. The project results also led to applications for two FWF projects and an Erwin Schrödinger Fellowship.