Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Real-time object detection using an ultra-high-resolution camera on embedded systems

Antonakakis Marios, Tzavaras Aimilios, Tsakos Konstantinos, Spanakis Emmanouil G., Sakkalis, Vangelis, Zervakis Michail, Petrakis Evripidis

Full record


URI: http://purl.tuc.gr/dl/dias/D56A1139-65C2-4E32-B5B4-64BB221CF0B3
Year 2022
Type of Item Conference Full Paper
License
Details
Bibliographic Citation M. Antonakakis, A. Tzavaras, K. Tsakos, E. G. Spanakis, V. Sakkalis, M. Zervakis, and E. G. M. Petrakis, "Real-time object detection using an ultra-high-resolution camera on embedded systems," in Proceedings of the 2022 IEEE International Conference on Imaging Systems and Techniques (IST 2022), Kaohsiung, Taiwan, 2022, doi: 10.1109/IST55454.2022.9827742. https://doi.org/10.1109/IST55454.2022.9827742
Appears in Collections

Summary

Unnamed Aerial Vehicle (UAV) - based remote sensing is a promising technology that is being applied for inspecting live scenes from high altitudes (e.g., for surveillance and recognizing emergencies). The evolution of hardware and software technologies in the last few years has generated additional interest in embedded systems research and its implementation in energy-independent UAVs for remote sensing. Alongside, ultra-high-resolution optical sensors are mandatory for acquiring high-resolution images which are necessary for accurate object detection from a distance (e.g., 1,000 meters). The processing of ultra-high-resolution images (e.g., 4K or 8K) is beyond the typical resolutions which are used for object detection (e.g., < 2K) emerging a necessity for special treatment in order to succeed a fast object detection. We propose a three-step approach deployed on a Docker runtime environment in an Nvidia Jetson AGX Xavier board. To support fast object detection, the captured images are split into K parts processed in parallel in separate containers running the YOLOv5 object detection algorithm. A final detection is constructed based on each one of the K detections. The experimental results are a good support to our claims of efficiency: the method can achieve close to real-time object detection for ultra-high (i.e., 8K) resolution images (i.e., in less than 1 second per frame).

Services

Statistics