A Tutorial on Multiple Extended Object Tracking

Abstract

<div>This tutorial introduces state-of-the-art methods for tracking multiple spatially extended objects based on unlabeled noisy point clouds, e.g., from radar or lidar sensors. In the first part, the focus lies on tracking a single extended object, i.e., the objective is to simultaneously estimate the shape and position of a moving object based on spatially distributed noisy detections. Model-based approaches for tracking elliptical and star-convex shaped objects are treated. Furthermore, recent learning-based approaches are discussed, which learn the spatial distribution of detections from real data. The second part considers the track management and data association problem, i.e., the initialization and termination of tracks as well as the association of detections to objects. After an in-depth analysis of the data association problem for extended objects, various data association approaches are discussed. Finally, recent frameworks for multiple extended object tracking are introduced such as the random finite set-based Poisson multi-Bernoulli mixture framework.</div>

Trajectory PMB Filters for Extended Object Tracking Using Belief Propagation

“…The explicit modelling of undetected objects is beneficial for automotive applications, e.g., in traffic scenes where pedestrians walking on sidewalks are occluded by parked vehicles along the street. More examples can be found in [4], [35].…”

“…This assumption is, however, unrealistic for modern high-resolution radar and Lidar sensors, for which it is common that an object gives rise to multiple measurements per time scan. The tracking of such an object is called extended object tracking (EOT), and overviews of EOT literature can be found in [3], [4]. The focus of this paper is on multiple EOT.…”

Mapping Extended Landmarks for Radar SLAM

“…By assuming some general parametric shape for the target extent, we can approximate different types of targets with varying sizes and shapes. Typical parametric shapes includes employing a specific geometric shape, such as en ellipse or a rectangle [8], and a more general star-convex model to represent an arbitrary shape [10].…”

“…To model the physical extent of landmark n as an ellipsoidal shape at time k, we use a semi-positive definite (SPD) random matrix X n k and adapt the Bayesian random matrix solution in [14] to estimate the landmark extent. Conversion between X n k and its extent parameter e n k can be found in [17] (Equations (8)(9)). In the following, we provide a summary of the steps for implementation of the landmark extent estimation.…”

“…Landmark extent estimation can better facilitate navigation and path planning for the mobile platform, since it provides a more precise map of the surrounding environment. However, existing methods of object extent estimation have mainly been developed in the target tracking literature (typically the single/multiple extended target(s) tracking [7,8]) and adaption of these approaches to radar SLAM is nontrivial. Firstly, the pose (location and orientation) of the mobile platform is unknown and evolves with time, and hence needs to be estimated simultaneously in SLAM.…”