Safety-Driven and Localization Uncertainty-Driven Perception-Aware Trajectory Planning for Quadrotor Unmanned Aerial Vehicles

Recent advances in trajectory planning have enabled quadrotor unmanned aerial vehicles (UAVs) to navigate autonomously in complex environments. However, most of the existing methods do not consider the perception quality and the safety simultaneously. This article proposes a perception-aware trajectory planning strategy for quadrotors, which can ensure the safety and localization accuracy. In contrast to the existing methods, the main idea of the proposed method lies in that the yaw angle trajectory is planned to actively obtain more information in the environment to improve the localization accuracy and keep the safe flight simultaneously. Following the mainstream two-stage motion planning framework, a coarse-to-fine graph search strategy is proposed to search for a safe and perception-aware yaw angle path in the first stage. Specifically, a Yaw Safety Corridor (YSC) is proposed to guarantee the safety, which can observe the obstacles directly along the tangent direction of the position trajectory. In addition, a dedicated map Fisher Information Field (FIF) is employed to evaluate the perception quality. In the second stage, a path-guided optimization method is proposed to quickly generate a safe and perception-aware trajectory. Finally, comparative simulation and real-world experiments are conducted to verify the superior performance in terms of the perception quality and the safety of the proposed method.

I. Introduction

Unmanned aerial vehicles (UAVs), especially quadrotors, equipped with the visual sensors, are suitable for different tasks such as mapping [1], transportation [2], [3], autonomous exploration [4], [5], [6], [7], inspection [8] and rescue [9]. However, it is essential to design robust localization [10] and planning algorithms to keep UAVs safely navigating in complex environments. Visual-inertial state estimation has received much attention in recent years [11], which assists the vision system with a low-cost inertial measurement (IMU) [12] for the localization. Typical visual-inertial navigation systems such as SVO [13], OKVIS [14], VINS [15] and ORB-SLAM3 [16] all track multiple visual features in the keyframes and adopt nonlinear optimization technique to estimate the state of the robot. The quantity and the quality of the visual features in the view will affect the localization quality, resulting in the localization uncertainty. For the planning algorithm, most of the existing methods focus on the safety of the position trajectory and time-optimal performance [17], [18], [19], [20]. However, the localization uncertainty is not considered as shown in Fig. 1. During the flight, the motion planning algorithm, without considering the localization uncertainty, potentially results in facing featureless areas such as white walls, which may cause a localization failure. Thus, a catastrophic crash may happen. However, the trajectory considering the localization uncertainty can keep more visual landmarks in the view to increase the localization accuracy as shown in Fig. 2. Therefore, it is necessary to consider the perception quality in the motion planning, which is called , to improve the performance on the localization accuracy. Fig. 1.

Without considering the perception awareness, the planned trajectory is oriented along the tangent direction of the position trajectory. The texture information in the environment is not utilized, resulting in relatively considerable localization uncertainty.

Fig. 2.

Considering the perception awareness, the planned trajectory is oriented along the textured direction. The texture information in the environment is fully utilized to reduce the localization uncertainty.