Predictive Collision Avoidance via Probabilistic Multi-Object Tracking in Dynamic Environments

Abstract

Collision avoidance for industrial AGVs operating in dynamic, shared environments remains challenging because reactive local planners treat moving obstacles as static, leading to conservative or unsafe behavior. We address this by enabling the local planner to reason over predicted obstacle motion rather than instantaneous positions. Specifically, an Ensemble Kalman Filter (EnKF)-based multi-object tracker provides filtered position and velocity estimates that a modified Dynamic Window Approach (DWA) uses to evaluate candidate trajectories against projected future obstacle states. A radius-informed ensemble generation scheme adapts filter uncertainty to the observed object size from 2D LiDAR, and a geometric-center representation provides stable bounding estimates under partial occlusion. The system is implemented as three modular, pipelined ROS 2 nodes. Experiments in Gazebo, Stage (up to five concurrent robots), and on a real industrial forklift demonstrate collision-free navigation in same-direction, crossing, and head-on scenarios with sub-10 ms controller response times.