Visual SLAM With Multi-Object Tracking Using Hybrid Features

Visual simultaneous localization and mapping (SLAM) enables ego localization and mapping, and provides critical technical support in autonomous driving and mobile robots. However, the assumption of rigid bodies usually made in the SLAM literature limits performances in dynamic environments. Moreover, incorporating surrounding objects into the unified estimation framework can enhance environmental understanding. In this paper, we present a unified SLAM system based on stereo cameras, which estimates ego pose and tracks multiple objects in dynamic environments. Our system separately estimates ego pose and tracks multiple objects, and creates individual local maps. To identify dynamic points, we calculate dynamic probabilities for each feature and subsequently refine them using local bundle adjustment (BA). For multi-object tracking, we utilize motion information and depth estimation for initialization, and construct an object local BA that includes hybrid feature re-projection factors and motion smooth factors for optimization. Several experiments are conducted to demonstrate the improvement of our system compared to the existing object SLAM systems.