Safety-Aware Perception for Autonomous Collision Avoidance in Dynamic Environments

TL;DR

A safety-aware approach for online determination of the optimal sensor-pointing direction which maximizes the perception of risk within the FOV within the field-of-view (FOV) restrictions is presented.

Abstract

Autonomous collision avoidance requires accurate environmental perception; however, flight systems often possess limited sensing capabilities with field-of-view (FOV) restrictions. To navigate this challenge, we present a safety-aware approach for online determination of the optimal sensor-pointing direction $ψ_\text{d}$ which utilizes control barrier functions (CBFs). First, we generate a spatial density function $Φ$ which leverages CBF constraints to map the collision risk of all local coordinates. Then, we convolve $Φ$ with an attitude-dependent sensor FOV quality function to produce the objective function $Γ$ which quantifies the total observed risk for a given pointing direction. Finally, by finding the global optimizer for $Γ$, we identify the value of $ψ_\text{d}$ which maximizes the perception of risk within the FOV. We incorporate $ψ_\text{d}$ into a safety-critical flight architecture and conduct a numerical analysis using multiple simulated mission profiles. Our algorithm achieves a success rate of $88-96\%$, constituting a $16-29\%$ improvement compared to the best heuristic methods. We demonstrate the functionality of our approach via a flight demonstration using the Crazyflie 2.1 micro-quadrotor. Without a priori obstacle knowledge, the quadrotor follows a dynamic flight path while simultaneously calculating and tracking $ψ_\text{d}$ to perceive and avoid two static obstacles with an average computation time of 371 $μ$s.

Figures (6)

• Figure 1: Safety-Aware Perception. Our optimal safety-aware perception methodology. Using an onboard sensor with limited FOV, the UAV determines its ideal sensor-pointing direction based on a local spatial density function. The peaks of the density function represent risk levels in the local environment, calculated using CBFs.
• Figure 2: Flight Control Block Diagram. The structure of the safety-critical controller for our UAV. Onboard sensors gather UAV and dynamic obstacle state information (green), and provide it to a feedback control algorithm (blue). The controller uses CBFs in a safety-critical QP to avoid collisions while tracking a user-defined position reference $\boldsymbol{r}_\text{d}$ (orange). With our safety-aware perception algorithm, the yaw reference $\psi_\text{d}$ is used to orient a fixed onboard sensor/camera to enable obstacle detection and tracking.
• Figure 3: Density Function Behavior. The density function $\Phi(r,\theta)$ for a particular obstacle changes with its values of collision risk $\alpha_k$ and confidence $\beta_k$. The density function peak heights modulate as a function of $\alpha_k$ while the peak widths expand and contract based on $\beta_k$.
• Figure 4: Dynamic Obstacle Simulations. The mission profiles for the (a) infinity symbol track and (b) sine wave corridor which were used to evaluate various perception methodologies. The success rates for these two profiles, for 100 randomly-generated simulations, are shown in (c) and (d) respectively. Flights were categorized as collision-free if the boundaries of the UAV never intersected those of the obstacles. Flights were further categorized as safe if the boundaries of the UAV never crossed those of the obstacle safe-sets, including the safety factor.
• Figure 5: Flight Demonstration Setup. The location of the Crazyflie at various times throughout the static collision avoidance demonstration. The corresponding onboard camera image is displayed, highlighting that obstacle information is effectively captured via yaw manipulation.