AI Solution for Task Motion Recognition
This research project is to develop a solution that analyzes task motions - The goal of this project is to recognize 40 detailed motions and to put 20 sets of subdivided motions in order - We are using 3D CNN models to recognize task motions in real time.
We have changed the class definition to fit the task data. At first, the class was defined by referring to the criteria of the external datasets (ver.1), but after that, the motion classes were classified according to the number and length of data (ver.2 & 3). We split the data by dividing the 3 classes with the most data into subclasses (ver.4).
The data we currently have is not long enough and it’s biased towards some kinds of motions. In addition, it takes a lot of effort and time to create detailed motion data for training in the model. So, we are securing additional data using open datasets and self-produced data.
We plan to achieve our performance objectives by updating the performance of motion classification and motion segmentation. Currently, the research on motion classification is almost completed, so we are now focusing on motion segmentation.