Transitional Grid Maps: Joint Modeling of Static and Dynamic Occupancy

TL;DR

This work tackles sensor-limited perception in dynamic environments by jointly modeling static and dynamic occupancy. It introduces Transitional Grid Maps (TGMs), a Bayesian-network framework that uses random transitions and convolution-based predictions to yield tractable per-cell updates while distinguishing static from dynamic content. Through real-vehicle lidar experiments, TGMs demonstrate improved static mapping and bolster SLAM performance in highly dynamic scenarios, while maintaining compatibility with existing localization pipelines. The approach offers a practical path toward more robust perception and planning, with potential extensions to velocity estimation and multi-sensor fusion.

Abstract

Autonomous agents rely on sensor data to construct representations of their environments, essential for predicting future events and planning their actions. However, sensor measurements suffer from limited range, occlusions, and sensor noise. These challenges become more evident in highly dynamic environments. This work proposes a probabilistic framework to jointly infer which parts of an environment are statically and which parts are dynamically occupied. We formulate the problem as a Bayesian network and introduce minimal assumptions that significantly reduce the complexity of the problem. Based on those, we derive Transitional Grid Maps (TGMs), an efficient analytical solution. Using real data, we demonstrate how this approach produces better maps by keeping track of both static and dynamic elements and, as a side effect, can help improve existing SLAM algorithms.

Figures (10)

• Figure 1: Transitional Grid Map (TGM). The dynamic environment at the current time, $m^d_t$, is predicted based on the transition probability between cells, represented with arrows, and the current belief about the static environment, $m^s$, which constrains the feasible transitions.
• Figure 2: Bayesian network of mapping a dynamic environment with known poses. Poses $x_{1:t}$ and measurements $z_{1:t}$ are known, and the goal is to estimate the static part of the environment $m^s$ and the current state of the dynamic part of the environment $m^d_t$.
• Figure 3: Volvo XC90 used for the data collection with sensors at the top.
• Figure 4: Color map used to represent the current belief about each cell for the baselines (left) and for TGMs (right).
• Figure 5: Snapshot of the scenario recorded for the first experiment. The image from the camera shows the ego vehicle stopped with a vehicle in front (top). The lidar scan shows the same vehicle at the front, another vehicle at the back and a few buildings (bottom).