| Presentation | 2018-10-18 [Poster Presentation] An optimization method of wireless communication bandwidth in drone object recognition using edge computing PengFei Sun, Akihiro Nakao, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Real-time object detection is considered challenging in many UAV applications, such as detection, surveillance, search and rescue, etc. There have been many real-time object recognition models based on deep learning, which however require high computational resources and need to be run with high-end Graphics Processing Units (GPUs). Generally, the cloud and edge servers are adopted to perform computing tasks based on images captured by the drone cameras. The most challenging problem with this approach is that, due to a large amount of data transferred between the drone and cloud server, the need for network bandwidth is extremely high for real-time object detection. Although 5G supports high bandwidth, if we would like to scale the number of UEs, the less bandwidth the service uses, the more scalable the service becomes. For this reason, our proposed method splits a deep neural network model and executes a part of computation to obtain intermediary data, which is supposed to be less than the final output data. Our approach preserves the accuracy of object recognition at a high level while reducing the wireless data transfer bandwidth. Moreover, we use additional computational resources for extra layers of the neural network, yet negligible enough to be executed on a drone, to compress the intermediary data so as to reduce the wireless network bandwidth. The usual data compression is for the sake of human eyes recognition, but our data compression is optimized for machine learning and executed by additional neural network layers. Using Yolo V2[1] , we evaluate our proposed method to find the detection accuracy of the proposed method reduces the data transmission by as much as 84% compared with the general approach, while the accuracy of object detection is reduced only by 3%. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | 5G / edge computing / AI / Object detection / drone |
| Paper # | NS2018-114 |
| Date of Issue | 2018-10-11 (NS) |
| Conference Information | |
| Committee | NS |
|---|---|
| Conference Date | 2018/10/18(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Kyoto Kyoiku Bunka Center |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Network architecture (Overlay, P2P, Ubiquitous network, Scale-free network, Active network, NGN, New Generation Network), Next generation packet transport (High speed Ethernet, IP over WDM, Multi-service package technology, MPLS), Grid, etc. |
| Chair | Yoshikatsu Okazaki(NTT) |
| Vice Chair | Akihiro Nakao(Univ. of Tokyo) |
| Secretary | Akihiro Nakao(NTT) |
| Assistant | Kenichi Kashibuchi(NTT) |
| Paper Information | |
| Registration To | Technical Committee on Network Systems |
|---|---|
| Language | ENG-JTITLE |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | [Poster Presentation] An optimization method of wireless communication bandwidth in drone object recognition using edge computing |
| Sub Title (in English) | |
| Keyword(1) | 5G |
| Keyword(2) | edge computing |
| Keyword(3) | AI |
| Keyword(4) | Object detection |
| Keyword(5) | drone |
| 1st Author's Name | PengFei Sun |
| 1st Author's Affiliation | The University Of Tokyo(The University Of Tokyo) |
| 2nd Author's Name | Akihiro Nakao |
| 2nd Author's Affiliation | The University Of Tokyo(The University Of Tokyo) |
| Date | 2018-10-18 |
| Paper # | NS2018-114 |
| Volume (vol) | vol.118 |
| Number (no) | NS-250 |
| Page | pp.pp.45-48(NS), |
| #Pages | 4 |
| Date of Issue | 2018-10-11 (NS) |