Fine-grained activity classification in assembly based on multi-visual modalities

Assembly activity recognition and prediction help to improve productivity, quality control, and safety measures in smart factories. This study aims to sense, recognize, and predict a worker's continuous fine-grained assembly activities in a manufacturing platform. We propose a two-stage network for workers' fine-grained activity classification by leveraging scene-level and temporal-level activity features. The first stage is a feature awareness block that extracts scene-level features from multi-visual modalities, including red–green–blue (RGB) and hand skeleton frames. We use the transfer learning method in the first stage and compare three different pre-trained feature extraction models. Then, we transmit the feature information from the first stage to the second stage to learn the temporal-level features of activities. The second stage consists of the Recurrent Neural Network (RNN) layers and a final classifier. We compare the performance of two different RNNs in the second stage, including the Long Short-Term Memory (LSTM) and the Gated Recurrent Unit (GRU). The partial video observation method is used in the prediction of fine-grained activities. In the experiments using the trimmed activity videos, our model achieves an accuracy of > 99% on our dataset and > 98% on the public dataset UCF 101, outperforming the state-of-the-art models. The prediction model achieves an accuracy of > 97% in predicting activity labels using 50% of the onset activity video information. In the experiments using an untrimmed video with continuous assembly activities, we combine our recognition and prediction models and achieve an accuracy of > 91% in real time, surpassing the state-of-the-art models for the recognition of continuous assembly activities.