Project 1: Developing a benchmark for object detection in aerial images
There are lots of object detection benchmarks, including MS-COCO and Pascal VOC to train and test object detection networks. However, the characteristics of these datasets are quite different than aerial datasets due to the perspective and resolution. In this project, you will train and test well-known object detectors with aerial datasets in the literature.
Project 2: Developing a benchmark for object detection in aerial videos
Object detection is easier in the natural image datasets since objects occupy a large number of pixels. However, an object in an aerial image may occupy a few numbers of pixels due to recording altitude and perspective. Therefore, the transition between frames may help object detection in aerial data. In this project, you will use variants of deep neural networks (maybe design your own) for object detection in aerial videos.
Project 3: Design a Deep Learning-based Perception System for autonomous drone racing
Design a perception system for a racing drone to identify and localize gates and important landmarks in drone racing contests on our drone platform. You may need to generate, examine and train datasets with deep-learning, CNN networks for robot and gate localization. Later, you will implement the perception system into a drone platform and do real-time testing.
Project 4: Study Visual Inertial Odometry for state estimation in fast, aggressive flights using visual cameras and/or Event-based cameras.
State estimation is a crucial component of any robotics application. In this project, we explore and improve the state-of-the-art methods in state estimation using visual cameras and/or event-based cameras, in combination with other sensors (IMUs, 2D LiDAR, …) for aggressive flight applications.
Project 5: Design fast and reliable object detection and estimation using Event-based cameras.
Event-based cameras are novel neuromorphic cameras that operate entirely differently in comparison with normal cameras, but offer significant advantages in dealing with motion blurs, distortion and challenging lighting conditions. In this project, we will leverage them to design a fast and reliable perception system for aggressive flights with MAV, when traditional cameras usually struggle.
Project 6: Motion planning with local mapping for a racing quadrotor.
A successful mission in drone racing requires making quick but correct decisions. In this project, we will address the problem of motion planning and mapping in such contexts. You will first build an effective map of the gate entrance quickly, and then use the map to plan for safe passage in real-time.
Please contact with Erdal Kayacan (firstname.lastname@example.org) for more information.