From Particles to Self-Localizing Tracklets: A Multilayer Particle Filter-Based Estimation for Dynamic Grid Maps

One of the indispensable functions of a self-driving vehicle is to estimate its dynamic world, which includes various traffic participants within complex driving scenarios. The estimation mechanism has to be flexible, fast, and robust; however, achieving these requirements is still challenging. Dynamic grid maps are one of the possible ways to combine and estimate multisensory information at an intermediate level. In this article, we present a particle filter (PF)-based grid map estimation that addresses several challenges. First, we propose a multilayer PF-based tracking solution that uses two measurement grid channels as inputs: an occupancy grid and a semantic grid. Second, we introduce the concept of structuring the particle population into batches, where each batch represents an individual tracklet. Rather than using one PF for estimating the entire grid, we employ multiple individual PFs (tracklets) that share the same world. Third, a concept of "self-localizing" tracklets is presented. Similar to simultaneous localization and mapping approaches, in our tracking solution, every particle state is extended with a small set of landmarks. This allows a tracklet to "selflocalize" itself with respect to a tracked-object boundary and leads to a more precise velocity estimation. Finally, we introduce an advanced tracklet-management mechanism that allows executing some specific PF operations at the tracklet level. This optimization provides multiple advantages. Experimental results with ground-truth data show improvement in estimation accuracy when compared to similar techniques.