Multi-object Tracking for Autonomous Driving with a Monocular Camera Subject to Missing Data

Abstract

In autonomous driving, visual surveillance may be blocked or damaged by external environment, leading to missing data for multi-object tracking (MOT) algorithms and tracking accuracy degradation. To overcome this problem, this work proposes a vehicular monocular camera MOT framework with missing data based on an expert evaluation error detection mechanism. In addition, our method avoids suffering from expensive hardware compared with those using LiDar and Radar and is more interpretative compared with those using end-to-end tracking algorithms relying on deep neural networks. The proposed method is tested on the KITTI dataset following several benchmark metrics (e.g., MOTA and ID-switch) as an evaluation criterion. Experimental results demonstrate that tracking with missing data based on our approach is still ideal compared with tracking without missing data.