Fujitsu develops model for higher accuracy in complex behavior recognition
Fujitsu has developed a deep learning technology for analyzing how adjacent joints in the body move during complex movements or behaviors to improve the accuracy of its behavior recognition. The example the company provides of such a movement is removing objects from a box.
The company claims it has achieved the highest accuracy in the world in benchmarking to the Kinetics-Skeleton dataset. In the evaluation, the accuracy rate relative to conventional object recognition for simple behaviors like standing and sitting was maintained, while the accuracy for 400 types of more complex behavior was improved by an average of nearly 8 percent, according to the announcement. The success rate at recognizing footage of a person throwing an object leaped from 8 percent to 24 percent in one example.
In the future, Fujitsu plans to add the new model to the model for 100 basic behaviors in its Actlyzer product for higher accuracy. The capability could roll out in fiscal 2021.
The details of the technology were presented by Fujitsu at last week’s virtual 5th International Conference on Pattern Recognition (ICPR 2020).