ICCV 2021

1st Workshop on Airborne Object Tracking (AOT)






Despite significant advances in Computer Vision, tracking extremely small and far away objects, such as flying Airborne Objects Tracking (AOT) by autonomous flights with Sense and Avoid (SAA) capability, still remains a substantial challenge. Detection, accurate localization and future motion prediction for these objects, which typically cover 0.01% of the image on average, is crucial for the safe navigation of autonomous drones under a wide range of weather and illumination conditions. The 1st Workshop on Airborne Object Tracking will be held virtually on TBD. This workshop aims at promoting research and closing the gap in performance of state-of-the-art Computer Vision-based models in the domain of airborne objects detection and tracking using monocular visual cameras onboard aerial vehicles.

Autonomous drones engage Sense and Avoid (SAA) technology for situational awareness and collision avoidance maneuver around unforeseen airborne obstacles on their mission path. Computer Vision models for detection and tracking of Airborne Objects (AO) provide an integral solution that enables the use of low-cost, lightweight cameras onboard these drones for safe navigation.

The domain of Airborne Objects Tracking (AOT), as opposed to the one of traditional computer vision datasets for detection and tracking, poses a unique combination of challenges. This is due to the dynamic nature of the AOs as well as the sensing camera. Moreover, owing to the tiny size (in pixels) of these distant objects, and low density of AOs in the airspace, the data is extremely sparse both spatially and temporally. A large degree of variation in weather conditions, scene illumination, view angle and background terrain add to the complexity of the task. Accurate localization and highly reliable tracking of AOs over several video frames is essential for avoidance maneuver by the drone, when required. Furthermore, for safety and efficiency, it is imperative that the solution produces a very low number of false alarms. Timely detection within a desirable temporal window is essential for generating reliable and informative tracks useful for future motion prediction. Together these conditions pose some novel challenges for training computer vision models on monocular video data for online detection and tracking on AOs.

Call for Papers

We invite researchers to submit papers to ICCV presenting their solutions to the various tasks on the Airborne Objects Detection and Tracking dataset. The proposed approaches should produce state-of-the-art results on the AOT Challenge dataset -

1. Airborne detection and tracking - Methods that detect AOs and generate tracks to successfully detect airborne encounters in flight video data while producing a very small number of false alarm rate per hour of flight.

2. Frame-level airborne detection - Methods that detect AOs at frame level while producing a very small number of false positives per image.

3. Semi-supervised or Unsupervised Segmentation

4. Monocular Depth Estimation

5. Dataset Bias Analysis


Please visit this website for more details about Airborne Object Tracking challenge







Please contact Yuri Federigi/Maria Zontak if you have question.