本文へジャンプ

成果報告書詳細
管理番号20190000000141
タイトル*2018年度中間年報 次世代人工知能・ロボット中核技術開発 次世代人工知能技術分野 AI・クラウドソーシング・ハイブリッド型広域人命捜索システム
公開日2019/6/19
報告書年度2018 - 2018
委託先名株式会社ロックガレッジ
プロジェクト番号P15009
部署名ロボット・AI部
和文要約
英文要約Title: Hybrid AI crowdsourcing search system for human lives in a wild-area. (FY2018-FY2019) FY2018 Annual Report

Rock Garage Co., Ltd.

1. Investigation/Examination

The following are based on the results of hearings conducted for fire fighting officials, the drone mountain rescue team and the associate professor of drones.

According to the users, there is some unexpected needs for the system, such as water rescue as well as mountain rescue, and search for a human body which is partly concealed under the object.
The user reaction was quite good after testing the system, which fully automatically conduct from the drone flight to the human search.
It will be an advantage that real-time processing, which is highly evaluated by the users, also has the merits when compared with another post-processing system to prevent the risk of data lost from any drone crush accidents.


2. R&D to improve AI accuracy
In order to increase the recognition accuracy of image recognithion by AI, the following items were conducted.

(a)Increase the number of teacher data
(b)Implement the automatic generator of teacher data
(c)Review the structure of deep learning network

(a)Increased up to 17000 data.
As a result, the data recognition taken from directly above the position which was worse before has been improved.

(b)Made program to compose data between human images with transparent backgraund and random background images.
As a result, false detection has been reduced with extraneous object in the background.

(c)Tested more simple, original U-Net system for deep learning.
The performance is not enough yet though it could detect small target at high speed.
Improve the performance; to be discussed.



3. R&D for data buffering due to local server, network issues

The uplink bandwidth changes due to the local radio wave environment or a time period of a day.
The system transmits fixed amount of data at all times through an uplink, therefore the change of bandwidth often made processing errors when if it occuers because of the transmission failure.
The implementation of buffering function enabled stable signal transmission regardless of the bandwidth fluctuations.


4. R&D, determining specification, design, coding oparation for front-end and back-end system

The android application and back-end system were designed and it has been started coding.
This android app is to control multiple drones at the same time, it can be used by the search team members at multiple locations, and cloud/server distributes the result data in the group as well as transmits data to the crowdworkers to complement AI.
ダウンロード成果報告書データベース(ユーザ登録必須)から、ダウンロードしてください。

▲トップに戻る