EO-SLAM: Evolutionary Object Slam in Perceptual Constrained Scene

Object-level SLAM systems can capture semantic instances in the environment to help systems better understand intelligence tasks. However, due to limitation of the sensor's viewing angle, classical object-level SLAM systems fail to parameterize instantiated objects under certain conditions. In this work, we propose a evolutionary object SLAM system which is termed EO-SLAM. In the absence of sufficient angular information, it parameterizes the object's observation as the cylinder model and allows the quality improvement based on multi-frame accumulated observations to upgrade the dual quadric model. Additionally, data association schemes that combine 2D tracing with 3D model fitting are explored, enabling the system to operate accurately and robustly in both short and long term. The evaluation results indicate that EO-SLAM achieves the best tradeoff between speed and accuracy.