New Methodologies for the Tracking of Low-Observable Objects

Multiobject tracking refers to the problem of estimating the time dependent number and states of multiple objects from measurements provided by one or multiple sensors. Applications include surveillance, autonomous driving, indoor localization, oceanography, robotics, and biomedical analytics. The sensor measurements are either preprocessed to reduce data flow and computational complexity, which leads to an approach known as detect-then-track (DTT) multiobject tracking, or unprocessed, which leads to an approach known as track-before-detect (TBD). A particularly challenging task in multiobject tracking is the tracking of low-observable objects. Low- observable objects give rise to sensor measurements of high uncertainty. Possible causes are limited sensing and detection capabilities of the sensors or objects that maneuver in areas of reduced observability, such as in underwater environments. The goal of the proposed postdoctoral stay is to devise high-performing yet efficient methodologies and algorithms for the tracking of multiple low- observable objects. In the proposed research, we focus on the TBD paradigm, which is especially suited for the tracking of low-observable objects because it does not discard relevant sensor information. In particular, for the problem modeling and the corresponding estimation process, we combine the techniques of random finite sets and belief propagation, which is expected to lead to powerful yet low-computational tracking algorithms. Additionally, we contribute to DTT multiobject tracking using multiple sensors. We will address distributed multisensor scenarios, i.e., without a central processing unit, where each sensor is only able to communicate with its neighbors. Most state-of-the-art algorithms fuse probability distributions using a hard association of tracked objects. To avoid an incorrect association, which often occurs in the case of low-observable objects, we propose to use a soft (probabilistic) association of objects.

Multi-object tracking refers to the problem of estimating the time-dependent number and states of multiple objects from measurements provided by one or multiple sensors. Applications include surveillance, autonomous driving, indoor localization, oceanography, robotics, and biomedical analytics. The sensor measurements are either preprocessed to reduce data volume and computational complexity, yielding the so-called detect-then-track (DTT) multi-object tracking approach, or left unprocessed, yielding the so-called track-before-detect (TBD) approach. A particularly challenging task in multi-object tracking is tracking low-observable objects. Low-observable objects give rise to sensor measurements of high uncertainty. Possible causes include limited sensing and detection capabilities of sensors or objects that maneuver in areas of reduced observability, such as underwater environments. The goal of the underlying Schrödinger project was to develop new tracking methodologies that yield both high tracking accuracy and low computational complexity for tracking multiple low-observable objects. Within the project, we focused on the TBD paradigm, which is particularly well-suited to tracking low-observable objects because it does not discard relevant sensor information. More precisely, we developed a method based on the frameworks of random finite sets (RFSs) and belief propagation (BP). The proposed method performs even well even at low signal-to-noise ratios and in the challenging scenario of multiple closely spaced objects. Since the amount of data is typically high in TBD scenarios, we developed a second method that uses only a fraction of the available data. The method is based on a pre-detector that selects only measurements for the tracking task that meet a specified quality criterion. We demonstrated that the proposed method achieves significantly lower complexity while incurring only a slight loss in tracking accuracy. Additionally, we contributed to DTT multi-object tracking using multiple sensors. In fact, we developed a distributed multi-sensor algorithm, i.e., an algorithm without a central processing unit, in which each sensor communicates only with its neighbors. Our proposed method is based on the concept of soft (probabilistic) association among objects tracked by different sensors. This has the advantage of avoiding an incorrect association, which often occurs in the case of low-observable objects. In developing this algorithm, we again used the RFS and BP frameworks. Finally, we applied multi-object tracking algorithms to real-world underwater tracking applications. More precisely, we successfully tracked the position of multiple humpback whales in a challenging underwater environment.