In this paper we present a direct semi-dense stereo Visual-Inertial Odometry (VIO) algorithm enabling autonomous flight for quadrotor systems with Size, Weight, and Power (SWaP) constraints. The proposed approach is validated through experiments on a 250 g, 22 cm diameter quadrotor equipped with a stereo camera and an IMU. Semi-dense methods have superior performance in low texture areas, which are often encountered in robotic tasks such as infrastructure inspection. However, due to the measurement size and iterative nonlinear optimization, these methods are computationally more expensive. As the scale of the platform shrinks down, the available computation of the on-board CPU becomes limited, making autonomous navigation using optimization-based semi-dense tracking a hard problem. We show that our direct semi-dense VIO performs comparably to other state-of-the-art methods, while taking less CPU than other optimization-based approaches, making it suitable for computationally-constrained small platforms. Our method takes less amount of CPU than the state-of-the-art semi-dense method, VI-Stereo-DSO, due to a simpler framework in the algorithm and a multi-threaded code structure allowing us to run real-time state estimation on an ARM board. With a low texture dataset obtained with our quadrotor platform, we show that this method performs significantly better than sparse methods in low texture conditions encountered in indoor navigation. Finally, we demonstrate autonomous flight on a small platform using our direct semi-dense Visual-Inertial Odometry. Supplementary code, low texture datasets and videos can be found on our github repo: https://github.com/KumarRobotics/sdd_vio.

在本文中，我们提出了一种直接半密集立体视觉惯性里程计 (VIO) 算法，使具有尺寸、重量和功率 (SWaP) 约束的四旋翼飞行器系统能够自主飞行。通过在配备立体相机和 IMU 的 250 g、22 cm 直径四旋翼飞行器上的实验验证了所提出的方法。半密集方法在低纹理区域具有优越的性能，这在基础设施检查等机器人任务中经常遇到。然而，由于测量尺寸和迭代非线性优化，这些方法在计算上更加昂贵。随着平台规模的缩小，板载 CPU 的可用计算变得有限，这使得使用基于优化的半密集跟踪的自主导航成为一个难题。我们表明，我们的直接半密集 VIO 的性能与其他最先进的方法相当，同时比其他基于优化的方法占用更少的 CPU，使其适用于计算受限的小型平台。我们的方法比最先进的半密集方法 VI-Stereo-DSO 占用更少的 CPU 量，因为算法中的框架更简单，多线程代码结构允许我们运行实时状态ARM 板上的估计。使用我们的四旋翼平台获得的低纹理数据集，我们表明该方法在室内导航中遇到的低纹理条件下的性能明显优于稀疏方法。最后，我们使用我们的直接半密集视觉惯性里程计在一个小平台上演示了自主飞行。补充代码，