Robust SLAM and motion segmentation under long-term dynamic large occlusions

Abstract

Visual sensors are key to robot perception, which can not only help robot localisation but also enable robots to interact with the environment. However, in new environments, robots can fail to distinguish the static and dynamic components in the visual input. Consequently, robots are unable to track objects or localise themselves. Methods often require precise robot proprioception to compensate for camera movement and separate the static background from the visual input. However, robot proprioception, such as \ac{IMU} or wheel odometry, usually faces the problem of drift accumulation. The state-of-the-art methods demonstrate promising performance but either (1) require semantic segmentation, which is inaccessible in unknown environments, or (2) treat dynamic components as outliers -- which is unfeasible when dynamic objects occupy a large proportion of the visual input.

This research work systematically unifies camera and multi-object tracking problems in indoor environments by proposing a multi-motion tracking system; and enables robots to differentiate the static and dynamic components in the visual input with the understanding of their own movements and actions. Detailed evaluation of both simulation environments and robotic platforms suggests that the proposed method outperforms the state-of-the-art dynamic SLAM methods when the majority of the camera view is occluded by multiple unmodeled objects over a long period of time.