Royal Academy of Engineering/ EPSRC Research Fellowship
Random-Set Filtering Techniques for Multi-Sensor Multi-Object Tracking and Data Fusion
Summary:
There is substantial interest in the development of autonomous systems to enable remote surveillance of environments where it is too dangerous or too costly for humans to go. For example, scientific investigation of Mars has been made possible by the development of NASA's Mars Exploration Rover and the EU's Beagle missions. The DARPA Grand Challenge was created to accelerate research and development in autonomous ground vehicles. Unmanned Aerial Vehicles (UAVs) are remotely operated or self-piloted aircraft that can carry cameras, sensors, communications equipment or other payloads and have been used in reconnaissance since the 1950s. Advances in sensing technologies enable these to conduct more challenging roles. Therefore, the principal aim of this project is the development of a framework for the detection, identification, and tracking of targets to enable the autonomous surveillance of complex multi-target multi-sensor environments.
Target tracking is a necessary part of systems that perform functions such as crime prevention, defence and anti-terrorism. Tracking algorithms take their input measurements from sensors which provide the signals such as radar, sonar or video. The measurements are taken at regular intervals and the task is to estimate the state of a target at each point in time, such as its position, velocity or other attribute. Successive estimates provide the tracks which describe the trajectory of a target.
The extension from a single-target scenario to a multiple-target environment is non-trivial since the number of targets may not be known and varies with time, there are missed detections where the target is not observed and observations may be false alarms due to clutter. In addition, the identities of the targets may need to be known to determine their trajectories. The usual method for solving this problem is to assign a single-target stochastic filter, such as a Kalman filter or an extended Kalman filter, to each target and use a data association technique to assign the correct measurement to each filter. This complexity of this approach can be extremely expensive due to the data association problem, and the mathematical foundation for multi-target stochastic filtering is unclear.
An alternative solution to the multiple target tracking problem is to view the set of observations collectively, and try to estimate the set of target states directly, where the correct report-to-track association is not considered observable. The novelty of the random-set approach to multi-target tracking is that it bypasses the data association problem, thereby drastically reducing the computational complexity. The approach is inspired by the recent success of association-free multi-target filtering algorithms such as the particle-PHD filter and Gaussian mixture PHD filter, which are derived from the random finite set formulation of the multi-target problem. These filters have led to robust multiple-target tracking algorithms which are effective in estimating both the correct number of targets and their state vectors in data with high false alarm rates and missed detections. Daniel Clark established the convergence properties of these algorithms, developed techniques to use them for multiple-target tracking, gave the first demonstration of the techniques on real data, and successfully deployed them in commercial trials for BP.
Recent studies show that in some scenarios the PHD filter algorithms have outperformed the industry standards, Multiple Hypothesis Tracking (MHT) and Joint Probabilistic Data Association (JPDA), which are commonly believed to be optimal solutions. This approach has attracted international attention and several practical applications have appeared in the literature. These include tracking vehicles in different terrains, tracking targets in passive radar located on ellipses, tracking mines in forward scan sonar, tracking feature points in images sequences, locating an unknown time-varying number of speakers and visual tracking. An extensive review of the mathematical foundations, algorithmic implementations and different practical applications is given in the new book by Ron Mahler.
This project will establish foundational techniques in sensor fusion to detect, identify and track formations of targets to ensure that higher-level real-time operational decisions can be achieved in complex sensor environments. The usual assumption of independent target motion is undermined in many applications where it is required to identify and track convoys of vehicles or aircraft flying in formation. Generalising single-target methods to group entities presents significant challenges since measurements are generating by independent targets within the group without completely correlated motion. This is further complicated by the need to identify the correct number within a group which is changing over time. If there are many targets within a region, then it may not be possible to identify individual targets, though the overall group or crowd dynamic could be modelled.