센서 불확실성을 고려한 PMBM 필터 기반 다중센서 융합 추적알고리즘

Abstract

With the recent development of advanced driver assistance systems, the need for reliable perception performance has emerged. Perception in autonomous driving is to properly recognize the surrounding environment. An autonomous vehicle must track both static and dynamic objects, such as roads and buildings, as well as pedestrians and cars. Research on dynamic object tracking is being actively conducted, but existing studies show limitations in special situations where multiple objects overlap each other, or objects are temporarily completely covered during tracking. In addition, various types of autonomous driving sensors such as cameras, LiDARs, and radars are used, and sensor fusion that properly fuses them has recently been emphasized as it can compensate for the shortcomings of each sensor. Therefore, in this paper, we propose a centralized dynamic object tracking algorithm using multiple sensors. We introduce a method for fusing in consideration of unique sensor data uncertainty and tracking based on the Poisson Multi-Bernoulli Mixture filter using fused data. A performance evaluation was performed using actual vehicle data and a NuScenes dataset. Results show that the proposed method not only maintains robustness in situations such as sensor faults but also probabilistically manages the birth and death of tracks, resulting in better results than existing tracking algorithms.