SLAM Gets a PHD: New Concepts in Map Estimation

Abstract

H aving been referred to as the Holy Grail of autonomous robotics research, simultaneous localization and mapping (SLAM) lies at the core of most the autonomous robotic applications [1]. This article explains the recent advances in the representations of robotic sensor measurements and the map itself as well as their consequences on the robustness of SLAM. Fundamentally, the concept of a set-based measurement and map state representation allows all of the measurement information, spatial and detection, to be incorporated into joint Bayesian SLAM frameworks. Modeling measurements and the map state as random finite sets (RFSs) rather than the traditionally adopted random vectors is not merely a triviality of notation. It will be demonstrated that a set-based framework circumvents the necessity for any fragile data association and map management heuristics, which are necessary in vectorbased solutions.The article focuses on an implementation of the simplest Bayesian set-based estimator-the probability hypothesis density (PHD) filter-and its application to SLAM. Using particles to represent hypothesized vehicle trajectories and a Rao-Blackwellized (RB) PHD filter for each particle's map, estimates of the location and, equally importantly, the number of features that have passed through the field of view (FoV) of a vehicle's sensor(s) are estimated.Vector-based SLAM solutions do not estimate this number of features and instead rely on map management and data association heuristics, which are implemented outside of the Bayesian (e.g., Kalman, unscented Kalman, particle filter) estimator to determine which features to update. In the parallel but closely related field of target tracking, recent research has shown that

Multiple target tracking in video data using labeled random finite set

“…The probability of a target being born is set to 0.005 for each location which is gaussian distributed with small variances in x and y directions measured in pixels. For doorways nearer to the camera a much larger variance is set (e.g., diag(100, 25)) and for doorways at the far end of the picture smaller variances are set (e.g., diag (25,9)). The rationale behind this is that doorways far from the camera appear smaller in the video ( perspective effect ) and take a fewer number of pixels to cover.…”

Poisson Multi-Bernoulli Mapping Using Gibbs Sampling

“…Two aspects that make the problem challenging are the unknown data association between the measurements and the landmarks, and the unknown number of landmarks. In [11] and [12], these two aspects are addressed through representing the map and the measurements as random finite sets (RFS), thus incorporating the uncertainties in the data association and the number of landmarks into the model. They use the probability hypothesis density (PHD) filter [13] to recursively approximate the posterior density of the map.…”

“…They use the probability hypothesis density (PHD) filter [13] to recursively approximate the posterior density of the map. In addition, [11] and [12] model the sensor detections as point objects, i.e., each object can generate at most one measurement at each time step. For some automotive sensors, the distance to landmarks is often such that a landmark is covered by more than one resolution cell of the sensor.…”

Distributed Multi - Robot Rao - Blac k wellized Particle Filt er ing Simultaneous Localization and Mapping with Consensus Calculation of Particle Weight and Posterior Parameters

“…Meanwhile, mapping is a computation problem of integrating environmental information obtained from sensor installed on the robot into map representation. There are several map representations that can be built using SLAM; those are: (1) grid map, a map that represents the environment into occupied grid [1,2,3,4,5], (2) topological map, a map that represents the environment in a simplified graph, which only provides vital information such as robot's position and geometric structure of environment remains [1,6,7], (3) line map, a map that represents the environment using line and curve model [1,8,9,10,11], and (4) feature based map, a map that represents the environment into a set of features called landmark [1,12,13,14].…”