In:
Journal of Robotics and Mechatronics, Fuji Technology Press Ltd., Vol. 32, No. 6 ( 2020-12-20), p. 1200-1210
Abstract:
Object detectors using deep learning are currently used in various situations, including robot demonstration experiments, owing to their high accuracy. However, there are some problems in the creation of training data, such as the fact that a lot of labor is required for human annotations, and the method of providing training data needs to be carefully considered because the recognition accuracy decreases due to environmental changes such as lighting. In the Nakanoshima Challenge, an autonomous mobile robot competition, it is challenging to detect three types of garbage with red labels. In this study, we developed a garbage detector by semi-automating the annotation process through detection of labels using colors and by preparing training data by changing the lighting conditions in three ways depending on the brightness. We evaluated the recognition accuracy on the university campus and addressed the challenge of using the discriminator in the competition. In this paper, we report these results.
Type of Medium:
Online Resource
ISSN:
1883-8049
,
0915-3942
DOI:
10.20965/jrm.2020.p1200
Language:
English
Publisher:
Fuji Technology Press Ltd.
Publication Date:
2020
detail.hit.zdb_id:
2587053-1
SSG:
31
Permalink