LiDAR Point Cloud-based Multiple Vehicle Tracking with Probabilistic Measurement-Region Association

TL;DR

The paper tackles LiDAR-based multi-target tracking of extended targets by proposing a probabilistic measurement-region association (PMRA) model that partitions each target's extent into five regions to capture realistic measurement distributions. PMRA is integrated into a Poisson multi-Bernoulli mixture (PMBM) filter, yielding a PMBM posterior $f_{k|k}(oldsymbol{X}_k|oldsymbol{Z}^k)$ with a data-driven, region-aware association mechanism that uses continuous integrals and visible-angle geometry. Compared to GGIW-PMBM and DRA-PMBM baselines, PMRA-PMBM shows higher accuracy in both target center and extent estimation in simulation, albeit with higher computational cost due to the particle-based implementation. This work advances robust, LiDAR-centric extended-target tracking for autonomous driving by improving measurement-region modeling and integration within the PMBM framework.

Abstract

Multiple extended target tracking (ETT) has gained increasing attention due to the development of high-precision LiDAR and radar sensors in automotive applications. For LiDAR point cloud-based vehicle tracking, this paper presents a probabilistic measurement-region association (PMRA) ETT model, which can describe the complex measurement distribution by partitioning the target extent into different regions. The PMRA model overcomes the drawbacks of previous data-region association (DRA) models by eliminating the approximation error of constrained estimation and using continuous integrals to more reliably calculate the association probabilities. Furthermore, the PMRA model is integrated with the Poisson multi-Bernoulli mixture (PMBM) filter for tracking multiple vehicles. Simulation results illustrate the superior estimation accuracy of the proposed PMRA-PMBM filter in terms of both positions and extents of the vehicles comparing with PMBM filters using the gamma Gaussian inverse Wishart and DRA implementations.

Figures (5)

• Figure 1: Illustration of the system model. The positions of the LiDAR and the target are denoted by $(x_{\mathrm{s}},y_{\mathrm{s}})$ and $(x_{\mathrm{t}},y_{\mathrm{t}})$. The target extent is divided into five regions $r_1$-$r_5$. The vertices $\mathbf{p}^1$-$\mathbf{p}^4$ are determined by the eigenvectors $\{\mathbf{e}^1,\mathbf{e}^2\}$ and eigenvalues $\{e^1,e^2\}$ of the extent matrix $E$.
• Figure 2: Illustration of the edge visibility and angles $\theta_{r_n}$. The blue color of region $r_2$ denotes that it is visible from the LiDAR, i.e., $b_{r_2}=1$.
• Figure 3: Visualization of the simulation scenario. The target extent at the start or end of a trajectory (solid line) is represented by a black rectangle, with the corresponding time indicated next to it.
• Figure 4: GOSPA for PMRA-PMBM, DRA-PMBM and GGIW-PMBM in the scenario of Fig. \ref{['Fig3']}, averaged over 100 Monte Carlo simulations.
• Figure 5: Extent estimates of the same target in a Monte Carlo simulation. Grey rectangles are the true extents, and cyan circles are the LiDAR measurements. The estimation results of PMRA-PMBM, DRA-PMBM, and GGIW-PMBM are shown by red solid, blue dashed, and black dotted rectangles, respectively.