Multi-Object Recognition Method Inspired by Multimodal Information Processing in the Human Brain
Ryoga Seki, Daichi Kominami, Hideyuki Shimonishi, Masayuki Murata, Masaya Fujiwaka · 2022 IEEE Globecom Workshops (GC Wkshps) · 2022
In order to realize the Digital Twin, it is necessary to instantly understand various objects existing in the real world through various sensor devices. In recent years, the development of machine learning technologies such as a convolutional neural network has been remarkable, and in the field of video analysis, they achieve very high recognition rates. However, if all sensor data were collected and processed in the cloud, network transmission bandwidth and communication delays would become bottlenecks. The brain is a light weight system that makes decisions based on uncertain observations and We have previously proposed an object recognition method based on a mathematical model of how the brain recognizes information. In this paper, we extend this method. The features of the brain’s recognition of spatial context are modeled by a conditional random field and incorporated into it. We show that our proposal can recognize multiple objects with an accuracy of more than 83.2% even from noisy information, and can be applied to 60 fps video in the evaluation environment.