Deploying fast object detection for micro aerial vehicles

Abstract

In this thesis, we present an evaluation of four state of the art convolutional neural network (CNN) object detectors, and a method to incorporate temporal information into an object detection pipeline for a micro aerial vehicle (MAV). This work was done as part of the Defense Advanced Research Projects Agency's (DARPA) Fast Lightweight Autonomy (FLA) program with the goal of creating an autonomous MAV that could explore and map an unknown urban environment. We tested four CNN-based object detectors on a range of compact deployable compute devices in order to select the best detector-hardware pair for our flight vehicle. We chose to use the MobileNetSSD object detector running on an Intel NUC Skull Canyon. Additionally, we developed an efficient object detection method that incorporates temporal information found in sequential camera frames by selectively utilizing an object tracker when the object detector fails. Our temporal object detection method shows promising results improving the recall of the base object detector on two of three datasets while maintaining a high framerate.