Pipeline & Results

Pipeline

We use the saliency-based detection module BASNet to get the obstacle mask M(t). We then combine it with the radial flow R to get the desired flow F ∗(t). We use the synthesized desired flow F ∗(t) along with a flow-based Visual Servoing pipeline to generate a control command using Cross Entropy Method.

Results on the Simulation Benchmark

Minimum Distance and Trajectory Length

Min Dist signifies the minimum distance of MAV from any building in the image. We show that our method has the highest Min Dist among all the methods, which empirically proves that our method maintains the safest distance from the buildings. Our method takes the safest path to the goal while other methods either collide with the building or graze past it, and hence our trajectory length is highest. ”×” indicates that the maneuver was not completed, and the drone collided.

Success Rate Comparison

Our method generalises across different scenes in the Building-99 and UrbanScene3D dataset to give a consistent controller performance

Qualitative results on the benchmark for selected scenes

Our method successfully avoids the obstacles on all the 10 scenes in the building 99 environment and 5 of 6 in the UrbanScene3D environments.

Here we show images for 5 intermittent poses captured during the obstacle avoidance for selected scenes in the simulation benchmark.

In Building 1, we are able to segment and avoid the building even though there are several buildings behind it. Building 4 covers a large part of the image, but our algorithm is able to move in the correct direction. In Building 10, the MAV can navigate the narrow path between the two buildings successfully. We also present results on certain challenging configurations from the real-world dataset UrbanScene3D.

The lines in the trajectory column indicate the following: (RED) –> Our Method, (BLUE) –> Flow Balancing, (GREEN) –> Flow Balancing with Radial Flow. The goal and start positions are marked with a red star and a yellow circle, respectively