Persistent Identity Tracking of Traffic Cones for Autonomous Robotic Grasping

Abstract

Object tracking is a key component of autonomous robotic manipulation when multiple visually similar objects are present, as the system must maintain consistent object identities over time. In this work, we investigate multi-object tracking of traffic cones for an autonomous cone placement and retrieval system. A task-oriented benchmark was created using real-world recordings acquired with an eye-in-hand camera setup. The evaluation dataset consists of five sequences covering multiple object-count scenarios. Each scenario was assessed under both clean and artificially degraded detection conditions to analyze the impact of degraded detection conditions on tracking performance. Several state-of-the-art tracking-by-detection algorithms were compared using standard multi-object tracking metrics. The results reveal how increasing scene complexity and detection degradation affect identity preservation and tracking robustness. The findings highlight the limitations of current approaches in robotic manipulation environments and provide guidance for selecting and developing tracking methods for reliable cone handling tasks.