MCOV-SLAM: A Multicamera Omnidirectional Visual SLAM System

Multicamera-based visual simultaneous localization and mapping (SLAM) systems prove to be more effective and robust for complex scenarios than monocular-based ones because of their capability of capturing more environmental information. However, most existing multicamera SLAM methods only extend on the basis of traditional single-camera methods and just use multiple cameras for tracking more feature points, in which the design of the front-ends and sensor layout are less theoretically grounded, such as the heuristic condition of inserting a new keyframe. Moreover, the omnidirectional perception ability of multicamera system has not been fully utilized in most existing methods. When performing place recognition, existing methods still need to get the point in similar position and orientation like what single-camera methods perform, rather than in any direction. To eliminate human heuristics, elevate loop-closing ability and boost system's performance, this article proposes a multicamera visual SLAM method based on observability and omnidirectional perception. The key novelties of this work are the design of an omnidirectional loop-closing method and a new keyframe decision method based on system's observability analysis. First, an observation model for multicamera system is constructed and analyzed, which provides a theoretical basis for system's sensor layout design and the further enhancement of multicamera visual SLAM method. Then, a feature matching result screening method and a novel keyframe decision method based on observability are proposed to ameliorate the precision and reliability of system. Lastly, an omnidirectional loop-closing method that fuses all cameras' information is proposed to realize loop detection and correction without sensor's direction constraint. Extensive experimental results demonstrate that the proposed MCOV-SLAM method has good augmentation in terms of system's accuracy and robustness.