Multiobject Tracking Using Multiple Sensors

Multiobject tracking using multiple sensors (MTMS) has found a wide variety of applications over the last years. For instance, in todays driver assistance systems, measurements provided by multiple cameras and automotive radar sensors are automatically fused in order to detect and track other road users and obstacles. Further applications of MTMS include biomedical analytics, robotics, remote sensing, oceanography, air traffic control, and computer vision. Investigating the performance limits of MTMS and developing accurate and reliable MTMS algorithms is difficult since the number of objects to be tracked and the associations between measurements and objects are unknown. Another, more recent challenge in MTMS results from the paradigm of agent networks. Here, agents (such as autonomous cars or a swarm of unmanned aerial vehicles) are organized in networked teams with the goal of performing tasks jointly. Due to the decentralized topology of most agent networks, state-of-the-art centralized MTMS algorithms are impractical. This calls for the development of a new type of distributed MTMS algorithms that are scalable and robust to agent failure. The proposed research project will address this challenge. The main goals are: to investigate the performance limits of MTMS; to develop distributed and scalable MTMS algorithms. The distributed nature and scalability of the MTMS algorithms to be developed will allow their use in large decentralized agent networks. In a first step, the fundamental limits of MTMS performance will be investigated by deriving meaningful performance bounds. This derivation is expected to lead to a deep understanding of the MTMS problem and to provide valuable guidelines for the development of distributed and scalable MTMS algorithms. In a second step, the obtained insights will be leveraged for developing MTMS algorithms that perform statistical inference by means of so-called message passing algorithms. Here, statistical information in the form of messages is exchanged along the edges of a graph representing the inference problem. Message passing algorithms achieve a very attractive performance-complexity compromise. Surprisingly, until very recently, the message passing approach was ignored by the object tracking and sensor fusion community.

Multiobject tracking (MOT) using multiple sensors has found a wide variety of applications over the last few years. For instance, in today's driver assistance systems, measurements provided by multiple cameras and automotive radar sensors are automatically fused in order to detect and track other road users and obstacles. Further applications of MOT include robotics, remote sensing, oceanography, air traffic control, and computer vision. The main scientific contribution of the project is the insight that graphical models are the ideal tool to describe existing MOT methods on the one hand and to develop advanced methods on a top-down theoretical approach. In particular, the principle of "stretching" or "opening" nodes can replace some of the messages with lower-dimensional messages, resulting in reduced computational complexity and improved scalability. For the first time, filtering and data association for randomly appearing and disappearing objects are described by a joint graph. The message-passing algorithm for this graph can outperform existing algorithms in terms of detection and tracking performance, computational complexity, scalability, robustness, and versatility. This project has developed novel methods for MOT that can substantially improve the performance of perception systems and thus lead to tangible advantages across multiple sectors in industry and government. Variants of the proposed methodologies will be useful in applications, including autonomous driving, medical imaging, and wireless communication.