LSTM and CNN Based IMU Sensor Fusion Approach for Human Pose Identification in Manual Handling Activities

Abstract

In recent years, human pose estimation has become a very important research topic in the context of control engines, and exoskeletons. In this paper, we propose a Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) based Hybrid Deep Neural network, aimed to estimate human pose while handling of loads. The proposed model is capable to identify three such activities, i.e. load lifting from the ground, load shifting, and uplifting of the load. For this purpose, a inertial sensor unit (IMU)-based system was developed to collect the raw data. Next, to obtain more robust and accurate results, Kalman filtering has been used as a fusion technique. Rigorous fine-tuning and simulations show that the model obtained from the Kalman filtering achieves better results as compared to the raw data. Our proposed model can classify the target activities with a test accuracy of 86%.