An IMM-Enabled Adaptive 3D Multi-Object Tracker for Autonomous Driving

Abstract

3D multi-object tracking (MOT) is a crucial part in the field of autonomous driving. Thanks to the recent advances in deep-learning-based detector, tracking-by-detection paradigm has become popular in 3D MOT, which consists of a front-end object detector and a back-end tracker. However, most existing methods only focus on the performance on particular data sets, ignoring the adaptiveness of the tracking algorithm to dynamically changing driving environment. Based on this, we design an adaptive 3D MOT algorithm, which can adapt its behavior to the complex changing environments in real driving scenarios. The system first utilizes a pre-trained 3D detector to produce the observations (detections) for the current frame. Then, a state estimator based on interacting multiple model (IMM), which takes the statistics of the data set into account and switches its state dynamically, provides the adaptive state estimation for target tracking. Experiments show that our algorithm can improve the performance of single-model-based methods, and adapt its behavior dynamically on nuScenes data set.