Scalable Real-Time Multi Object Tracking for Automated Driving Using Intelligent Infrastructure

Automated driving highly relies on object detection and tracking. While vehicle autonomous solutions represent the state of the art, connected systems are increasingly being developed and researched. Many current approaches to Multi Object Tracking focus on the accuracy, whereas such a system working as the backbone of a live and offline digital twin of a test field mainly needs to be scalable to a large amount of perceived objects and still real-time capable. This paper presents a novel approach to Multi Object Tracking that can reach framerates of up to 604 Hz spanning a digital test field of 0.5 km ² size with still reasonable accuracy on simulated data. The approach can also be applied to real world data where our framework reaches a similar accuracy as the common baseline AB3DMOT on the nuScenes dataset, being 588 % faster. Using a new motion model in combination with an Unscented Kalman Filter, we can further increase the framework's accuracy and robustness while still reaching a framerate of 544 Hz on nuScenes and 367 Hz on the ACCorD test field.