MOT-AS: Real-Time Scheduling Framework for Multi-Object Tracking Capturing Accuracy and Stability

Abstract

Unlike existing accuracy-centric multi-object tracking (MOT), MOT subsystems for autonomous vehicles (AVs) must accurately perceive the surrounding conditions of the vehicle and timely deliver the perception results to the control subsystems before losing stability. In this paper, we proposed MOT-AS (Multi-Object Tracking systems capturing Accuracy and Stability), a novel handover-aware MOT execution and scheduling framework tailored for AVs with multi-cameras, which aims to maximize tracking accuracy without sacrificing system stability. Given the resource limitations inherent to AVs, MOT-AS partitions the handover-aware MOT execution into two distinct sub-executions: tracking handover objects that move across multiple cameras (referred to as global association) and those that move within a single camera (termed local association). It selectively performs the global association only when necessary and carries out local association with multiple execution options to explore the trade-off between accuracy and stability. Building upon MOT-AS, we developed a new scheduling framework encompassing a new MOT task model, offline stability analysis, and online scheduling algorithm to maximize accuracy without compromising stability. We implemented MOT-AS on both high-end and embedded GPU platforms using the Nuscenes dataset, demonstrating enhanced tracking accuracy and stability over conventional MOT systems, irrespective of their handover considerations.